Bioelectric Source Localization in Peripheral Nerves...Abstract Bioelectric Source Localization in...
Transcript of Bioelectric Source Localization in Peripheral Nerves...Abstract Bioelectric Source Localization in...
Bioelectric Source Localization in PeripheralNerves
by
Jose Zariffa
A thesis submitted in conformity with the requirements
for the degree of Doctor of Philosophy
Edward S. Rogers Sr. Department of Electrical and Computer Engineering
and
Institute of Biomaterials and Biomedical Engineering
University of Toronto
c©Copyright by Jose Zariffa 2009
Abstract
Bioelectric Source Localization in Peripheral Nerves
Jose Zariffa
Doctor of Philosophy
Graduate Departments of
Edward S. Rogers Sr. Department of Electrical and Computer Engineering
and
Institute of Biomaterials and Biomedical Engineering
University of Toronto
2009
Currently there does not exist a type of peripheral nerve interface that adequately
combines spatial selectivity, spatial coverage and low invasiveness. In order to address this
lack, we investigated the application of bioelectric source localization algorithms, adapted
from electroencephalography/magnetoencephalography, to recordings from a 56-contact “ma-
trix” nerve cuff electrode. If successful, this strategy would enable us to improve current
neuroprostheses and conduct more detailed investigations of neural control systems. Using
forward field similarities, we first developed a method to reduce the number of unnecessary
variables in the inverse problem, and in doing so obtained an upper bound on the spatial
resolution. Next, a simulation study of the peripheral nerve source localization problem
revealed that the method is unlikely to work unless noise is very low and a very accurate
model of the nerve is available. Under more realistic conditions, the method had localization
errors in the 140 µm-180 µm range, high numbers of spurious pathways, and low resolution.
On the other hand, the simulations also showed that imposing physiologically meaningful
constraints on the solution can reduce the number of spurious pathways. Both the influence
of the constraints and the importance of the model accuracy were validated experimentally
using recordings from rat sciatic nerves. Unfortunately, neither idealized models nor models
based on nerve sample cross-sections were sufficiently accurate to allow reliable identification
ii
of the branches stimulated during the experiments. To overcome this problem, an experi-
mental leadfield was constructed using training data, thereby eliminating the dependence on
anatomical models. This new strategy was successful in identifying single-branch cases, but
not multi-branches ones. Lastly, an examination of the information contained in the ma-
trix cuff recordings was performed in comparison to a single-ring configuration of contacts.
The matrix cuff was able to achieve better fascicle discrimination due to its ability to select
among the most informative locations around the nerve. These findings suggest that nerve
cuff-based neuroprosthetic applications would benefit from implanting devices with a large
number of contacts, then performing a contact selection procedure. Conditions that must
be met before source localization approaches can be applied in practice to peripheral nerves
were also discussed.
iii
Acknowledgements
This thesis is the reflection of four years of work, during which I have learned many
valuable lessons about conducting research, lessons that cannot be reduced to any of the
chapters, figures, or references in this document. In this respect, I could not have hoped for
a better supervisor than Dr. Milos Popovic. He gave me tremendous freedom and autonomy
in my research, yet at the same time always came through when I needed help and support.
He taught me to see past my thesis project, to understand the bigger context, and never to
forget the people for whom we carry out biomedical research, and for these things I will be
thanking him for many years to come.
Dr. Popovic may have made the last four years interesting, educational, productive,
and occasionally very funny, but it was all the members of the Rehabilitation Engineering
Laboratory who made it so easy to go to work every day. To have such a large group of
people who are all, without exception, kind, smart, funny, and helpful, is surely some sort
of major statistical anomaly, and I’m immensely grateful to all of them for making my PhD
fun.
This work was not conducted in isolation, but was made possible by the support and
collaboration of a large group of people. First and foremost, Dr. Mary Nagai, whose help
with the experiments was invaluable. I owe many thanks as well to our collaborators at
the University of Freiburg in Germany, Drs. Thomas Stieglitz and Martin Schuettler, who
provided us with the electrodes that made the study possible, and to our collaborators at the
Center for Addiction and Mental Health in Toronto, in particular Dr. Jeff Daskalakis and
Lori Dixon, who provided us with the resources and space necessary for our experiments.
Finally, the members of my thesis committee, Drs. Berj Bardakjian and Adrian Nachman,
gave me a great deal of very sound advice, kept me on the right path, and were always
encouraging.
Academia, of course, is only half the story. My studies, and for that matter everything
iv
else I do, would be impossible without my wife Vanessa, whose love and support have been
unflinching since the moment we set eyes on each other. My parents, likewise, have seemingly
endless confidence in me, and for this selective memory I thank them with all my heart.
And last but, as they say, not least, I wish to thank all of my friends and family who have
supported and encouraged me all this years, even when they had no idea what my research
was about.
v
List of Abbreviations
Action potential AP
Akaike’s Bayesian information criterion ABIC
Boundary element method BEM
Central nervous system CNS
Compound action potential CAP
Cross-validation error CVE
Electrical analysis ELECTRA
Electrical impedance tomography EIT
Electroencephalogram EEG
Finite element FE
Finite element method FEM
Flat interface nerve electrode FINE
Focal underdetermined system solver FOCUSS
Functional magnetic resonance imaging fMRI
Generalized cross-validation GCV
Hematoxylin and eosin H&E
Linear programming LP
Linearly constrained minimum variance LCMV
Local auto-regressive averages LAURA
Longitudinal intrafascicular electrode LIFE
Low resolution brain electromagnetic tomography LORETA
lp norm iterative sparse solution LPISS
Magnetic resonance imaging MRI
Magnetoencephalogram MEG
vi
Micro-electrode array MEA
Minimum current estimate MCE
Multi-contact cuff MCC
Multiple signal classification MUSIC
Noise-to-signal ratio NSR
Positron emission tomography PET
Peripheral nervous system PNS
Restricted maximum likelihood ReML
Slowly penetrating interfascicular nerve electrode SPINE
Standardized LORETA sLORETA
Truncated singular-value decomposition TSVD
vii
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.1 Modeling phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.2 Source localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.3 Experimental validation . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Organization of the Document . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Literature Review 10
2.1 Origin of Bioelectric Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Bioelectromagnetic Source Localization . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 The forward problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.2 Equivalent dipole methods . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.3 Distributed linear methods . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.4 The dynamic source localization problem . . . . . . . . . . . . . . . . 33
2.2.5 Choice of regularization parameter . . . . . . . . . . . . . . . . . . . 36
2.3 Peripheral Nerve Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3.1 Electrode types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.2 Signal processing techniques for extraneural measurements . . . . . . 43
viii
2.3.3 Source localization and pathway discrimination using extraneural mea-
surements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3 Reduction of the Inverse Problem Solution Space 48
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.1 Finite element model and leadfield construction . . . . . . . . . . . . 52
3.3.2 Element grouping algorithm . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Leadfield comparison metrics . . . . . . . . . . . . . . . . . . . . . . 57
3.3.4 Example and Complexity Analysis . . . . . . . . . . . . . . . . . . . 59
3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4.1 Size of the reduced leadfield . . . . . . . . . . . . . . . . . . . . . . . 61
3.4.2 Properties of reduced leadfield . . . . . . . . . . . . . . . . . . . . . . 63
3.4.3 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 70
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.2.1 Approach to solving the source localization problem . . . . . . . . . . 70
4.2.2 Evaluation of the results . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.1 One-pathway case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.2 Three-pathways case . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5 Validation of the Source Localization Approach on Physiological Data 91
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
ix
5.2.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.2.2 Construction of nerve-specific leadfields . . . . . . . . . . . . . . . . . 98
5.2.3 Evaluation of the source localization performance . . . . . . . . . . . 100
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.3.1 Using the idealized leadfield without the spatio-temporal constraint . 103
5.3.2 Using the idealized leadfield with the spatio-temporal constraint . . . 104
5.3.3 Using the nerve-specific leadfield without the spatio-temporal constraint107
5.3.4 Using the nerve-specific leadfield with the spatio-temporal constraint 110
5.3.5 Influence of the constraints on the number of peaks in the estimate . 110
5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6 Source Localization Using an Experimentally-Derived Leadfield 120
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2.1 Construction of the experimental leadfield . . . . . . . . . . . . . . . 121
6.2.2 Identification of fascicle combinations . . . . . . . . . . . . . . . . . . 123
6.2.3 Evaluation of the results . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7 Influence of the Number and Location of Recording Contacts on the Se-
lectivity of a Nerve Cuff Electrode 133
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.2.1 Evaluation of the classification success rate . . . . . . . . . . . . . . . 134
7.2.2 Evaluation of the influence of the stimulation artefact . . . . . . . . . 137
7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.3.1 Comparison of the matrix and single-ring configurations . . . . . . . 139
7.3.2 Influence of the stimulation artefact . . . . . . . . . . . . . . . . . . . 141
x
7.3.3 Layout of the most informative contacts . . . . . . . . . . . . . . . . 144
7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8 Discussion 150
8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.2 Comparison of the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
8.2.1 Resolution achievable . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
8.2.2 Validation of the simulations . . . . . . . . . . . . . . . . . . . . . . . 156
8.2.3 Implication of the contact configuration study on the source localiza-
tion results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.3 Limitations of the study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.3.1 Experimental issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.3.2 Use of CAPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.3.3 Use of simplified FE models . . . . . . . . . . . . . . . . . . . . . . . 159
8.3.4 Use of a low-resolution source localization algorithm . . . . . . . . . . 160
8.3.5 Use of peaks in the estimate as a measure of the number of pathways 160
8.3.6 Focus on spatial over temporal resolution . . . . . . . . . . . . . . . . 161
8.3.7 Focus on a specific electrode and nerve . . . . . . . . . . . . . . . . . 161
8.4 Optimal number of contacts . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
8.5 Number of pathways to be localized . . . . . . . . . . . . . . . . . . . . . . . 164
8.6 Factors related to the use of a nerve cuff electrode . . . . . . . . . . . . . . . 166
8.7 Implications for future cuff-based peripheral nerve interfaces . . . . . . . . . 168
9 Conclusions 172
Bibliography 175
xi
List of Figures
1.1 56-contact nerve cuff electrode. . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Schematic overview of bioelectric source localization in peripheral nerves. . . 6
1.3 Anatomy of a peripheral nerve. . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Two implementations of the tripole configuration. . . . . . . . . . . . . . . . 41
2.2 Types of extraneural electrodes . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3 Types of intraneural electrodes . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 Mesh element grouping example. . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Number of columns of the reduced leadfield. . . . . . . . . . . . . . . . . . . 62
3.3 2D projections of the true source distribution and the estimates obtained with
the original and reduced leadfields. . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 FEM geometry cross-sections. . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Source localization metrics for the one-pathway case. . . . . . . . . . . . . . 82
4.3 Localization trial examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4 Source localization metrics for the three-pathways case. . . . . . . . . . . . . 86
5.1 Placement of the cuff electrodes. . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Connector linking the cuff electrode to the amplifier. . . . . . . . . . . . . . 96
5.3 Example of a trial before and after conversion to a common-average reference. 97
5.4 H&E stained sciatic nerve cross-sections. . . . . . . . . . . . . . . . . . . . . 99
5.5 Cross-sections of FE meshes based on nerve samples. . . . . . . . . . . . . . 101
xii
5.6 Summary of the source localization performance (idealized leadfield without
spatio-temporal constraint). . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.7 Summary of the source localization performance (idealized leadfield with spatio-
temporal constraint). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.8 Summary of the source localization performance (nerve-specific leadfield with-
out spatio-temporal constraint). . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.9 Summary of the source localization performance (nerve-specific leadfield with
spatio-temporal constraint). . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.10 Mean number of peaks in the estimates. . . . . . . . . . . . . . . . . . . . . 113
5.11 Spurious and missed pathways in the estimates. . . . . . . . . . . . . . . . . 119
6.1 Means of activity estimates for each branch. . . . . . . . . . . . . . . . . . . 126
6.2 Success rates for identifying the exact combination of active branches (thresh-
old = 0.2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.3 Success rates for identifying the exact combination of active branches (thresh-
old = 0.6). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.1 Contact configurations for matrix and single-ring cases. . . . . . . . . . . . . 135
7.2 Example of a trial before and after conversion to a tripole reference. . . . . . 138
7.3 Maximum classification success rate for each configuration. . . . . . . . . . . 140
7.4 Classification success rate using the first 8 contacts of each configuration. . . 141
7.5 Classification success rate as a function of the number of contacts. . . . . . . 142
7.6 Order in which contacts were added. . . . . . . . . . . . . . . . . . . . . . . 143
7.7 Comparison of artefact and success rate variations between contact rings in
Rat 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
xiii
List of Tables
2.1 Commonly used weight matrices for weighted minimum norm methods. . . . 26
3.1 Parameters for the finite element model of the rat sciatic nerve. . . . . . . . 53
3.2 Metric Met1 for all test cases. . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Simulation results using the original and reduced leadfields. . . . . . . . . . . 66
3.4 Computation time comparison for simulations using the original and reduced
leadfields. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 Parameters for the idealized finite element model of the rat sciatic nerve. . . 74
5.1 Cuff diameters in nerve-specific rat models. . . . . . . . . . . . . . . . . . . . 100
5.2 Influence of constraints on the number of peaks. . . . . . . . . . . . . . . . . 114
7.1 Correlation of the artefact and classification success rate variations between
contact rings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
xiv
Chapter 1
Introduction
1.1 Motivation
The human nervous system is responsible for controlling a staggering number of physiological
functions, both autonomic and somatic. Before any medical intervention related to neural
mechanisms can take place, it is crucial to have an understanding of the corresponding
control systems, and this has made the nervous system one of the most active topics of
research in the last century. There are two major obstacles that make deciphering the inner
workings of the nervous system so difficult: its sheer scale, and the technology available to
us. With over 100 billion neurons in the human brain alone [69], identifying the specific
networks corresponding to given tasks is enormously difficult. To compound this problem,
our ability to examine the activity of a specific neuron or group of neurons is limited by
technological considerations.
When monitoring neural activity in the brain, existing methods rely either on non-
invasively detecting changes in blood oxygen levels (usually using functional magnetic res-
onance imaging (fMRI)), or on recording the bioelectric or biomagnetic fields generated
by neural activity. The latter can be recorded non-invasively using electroencephalography
(EEG) or magnetoencephalography (MEG). Alternatively, invasive measurements of bio-
electric fields with different types of electrodes can be used to obtain electrocorticographic
1
Introduction 2
(ECoG) recordings from the surface of the cortex, local-field potentials (LFPs) from small
groups of neurons, or even single-neuron recordings. Methods based on field recordings have
much higher temporal resolution than fMRI, but none of them can equal the latter for high
resolution images of the entire brain. More recently, source localization approaches have been
applied to EEG or MEG recordings in order to try to obtain whole-brain information with
high temporal resolution, but spatial resolution of these methods still remains substantially
inferior to that of fMRI.
In peripheral nerves, monitoring techniques are currently restricted to recording bio-
electric fields, either from within the nerve (intraneural potentials) or right outside of it
(extraneural potentials). These two approaches represent a trade-off between resolution and
invasiveness. Micro-electrode arrays (MEAs) provide a large number of measurements from
specific locations within the nerve, but are more likely to damage the nerve and may be too
big to be used on smaller nerves. Longitudinal intrafascicular electrodes (LIFEs) are equally
selective and less damaging, but can only monitor a few nerve fibers. Nerve cuff electrodes,
on the other hand, use only extraneural measurements, making them less invasive and likely
to damage the nerve, but also less selective.
Although the neural control of physiological systems is carried out by networks of
neurons in the central nervous system (CNS), the difficulty of identifying the neurons that
are involved in a specific task creates the need for additional sources of information. Con-
sequently, using peripheral nerve recordings to monitor the information being conveyed to
and from the CNS is a complementary and attractive approach for helping to decipher
these control systems. Indeed, the “inputs” and “outputs” of the CNS can provide valuable
information about the control mechanisms being used, and task-specific signals are more
easily discriminated because each peripheral nerve innervates a known region of the body.
Nonetheless, the selectivity of peripheral nerve recordings is still too limited for this strategy
to reach its full potential. This is particularly true in the context of applications in humans,
where the higher selectivity of MEAs cannot be fully taken advantage of because of the risk
of nerve damage associated with those devices. Cuff electrodes, on the other hand, have
Introduction 3
been chronically implanted in humans for up to 12 years with no adverse effects [192].
1.2 Objective
In light of the situation described in the previous paragraph, the goal of the study presented
in this thesis was to improve the amount of information that can be obtained from extra-
neural recordings of peripheral nerve activity. More specifically, the aim was to develop a
system that improves the spatial resolution of traditional cuff electrode recordings, such that
bioelectric fields measured can be associated with specific pathways or groups of pathways
within the nerve.
1.3 Applications
The information obtained from higher resolution peripheral nerve recordings could be used
as follows:
1. Attributing recorded afferent and efferent activity to more specific neural pathways,
allowing us to better understand how the CNS controls the activity of a limb or organ
during complex tasks (for example reaching [159]).
2. Using the greater selectivity to improve the performance and capabilities of neuro-
prosthetic systems, which are defined as artificial systems that interact directly with a
damaged nervous system in order to replace or enhance its function.
The second point is broad. Some existing neuroprosthetic systems use nerve cuff
recordings of afferent signals to provide feedback to a stimulation device (for example systems
correcting for foot-drop [55, 75, 51, 52] or assisting grasping [64, 65]; other applications
under investigation in humans include quiet standing [198, 102, 12] and bladder control
[78]). The feedback provided can represent either a discrete event (e.g. heel strike) or a
Introduction 4
continuous quantity (e.g. a joint angle). Discriminating between the activity of different
sensory pathways would allow for more sophisticated control of the stimulation.
A more ambitious application, and one requiring real-time pathway discrimination
with good resolution, would be to convert efferent neural control signals into commands for
a prosthetic limb after amputation, thereby creating a direct neural interface ([191, 27, 26]).
With sufficient resolution, a “natural” interface could be created, in which the signals of
various pathways are used for their original purpose, enabling the user to control the artificial
limb as he or she would a real limb. With lower resolution, “natural” control may not be
possible, but the activation of different fascicles could be used as part of multiple-switch type
of control scheme.
1.4 Approach
The approach proposed to attain our objective consisted of two main components. First,
instead of a traditional nerve cuff electrode (which typically contains three contacts and
produces a single signal), a multi-contact cuff (MCC) was used (here with 56 contacts),
providing a much larger number of measurements obtained from all around the nerve. This
electrode is shown in Figure 1.1. Second, in order to best deal with this larger amount of
data, the task of locating the bioelectric sources within the nerve was treated as an inverse
problem of source localization. This is in contrast with the signal processing techniques
traditionally used to interpret nerve cuff recordings. A schematic overview of the approach is
shown in Figure 1.2. The design and fabrication of the MCC was performed by collaborators
at the Laboratory for Biomedical Microtechnology at the University of Freiburg, Germany.
Therefore, this thesis deals mainly with the mathematical treatment of the inverse problem,
and with the experimental validation of the system.
Introduction 5
Figure 1.1: The 56-contact MCC used in this study.
1.4.1 Modeling phase
The source localization process requires us to have a model describing the relationship be-
tween bioelectric sources at different locations in the nerve and the MCC measurements. In
addition, such a model makes it possible to generate simulated measurements and therefore
to investigate source localization approaches before experimental data is collected. The first
step of this project was therefore to construct a model of a nerve. The anatomy of a typical
nerve is shown in Figure 1.3. Individual fibers are grouped into bundles known as fascicles,
and the membrane delimiting each fascicle is the perineurium. The fascicles are in turn en-
cased in a tissue known as the epineurium to form the nerve trunk. The inside of the fascicles
is known as the endoneurium. Finite-element analysis can be used to model the electric fields
produced by electric sources in regions of complex geometry and inhomogeneous, anisotropic
conductivity. It is used here to create a model that reflects the nerve geometry, the different
Introduction 6
Figure 1.2: Schematic overview of our approach to localizing active pathways in peripheral
nerves: an MCC is used to record the electric fields generated by the neural activity at various
locations around the nerve. These recordings are used as the input to a source localization
algorithm, which produces a three-dimensional estimate of the bioelectric activity in the nerve.
conductivities of the endoneurium, perineurium, and epineurium, the shape and properties
of the MCC placed around the nerve, and those of the tissues and fluids around and between
the nerve and the electrode.
1.4.2 Source localization
Using the nerve model, simulation studies can be conducted to investigate different algo-
rithms for solving the inverse problem of source localization. The mathematical treatment
of this problem can be directly based on another bioelectric source localization problem, elec-
troencephalogram (EEG) source localization [104]. The goal in that context is to determine
Introduction 7
Figure 1.3: Anatomy of a peripheral nerve (from [199])
the distribution of brain activity that is responsible for producing a set of measurements
observed on the scalp. The EEG and peripheral nerve source localization problems are
similar in that they both involve the same underlying biophysics governing the relationship
between the electrical activity of neurons and extracellular potentials. In both cases, any
given set of measurements can be produced by an infinite number of source configurations,
such that the problem lacks a unique solution (this was shown by Helmholtz in 1853 [58]; a
translation of the relevant section of his often-cited paper is available in [59]). Small changes
in the measurements (i.e. noise) can also have a large effect on the solution, making the
problem unstable. A problem with these characteristics fails to meet Hadamard’s definition
of a well-posed problem (namely, that a solution exists, is unique, and depends continuously
on the data [47]). The bioelectric source localization problem therefore belongs to the class
of ill-posed inverse problems.
The most obvious difference between the EEG and peripheral nerve source localiza-
tion problems is the geometry of the region containing the sources, which is why the model
described in the previous section is necessary. In addition, solving this type of ill-posed
problem requires the imposition of constraints onto the solution to overcome the underde-
Introduction 8
termined nature of the problem. These constraints should ideally be tailored to the problem
and accurately reflect available information (e.g. biophysical, anatomical or electrophysio-
logical characteristics). For that reason, it is to be expected that not all the constraints that
are applicable to the EEG source localization problem are applicable to the peripheral nerve
problem, and vice versa. Part of the process of finding an appropriate algorithm for the
peripheral nerve source localization problem is therefore to select useful and well-justified
constraints.
1.4.3 Experimental validation
After having selected a source localization algorithm on the basis of simulated performance,
the last step of the project was to test the complete system (MCC and algorithm) using
physiological measurements. For this purpose, MCC recordings were obtained from rat
sciatic nerves during a series of in vivo experiments. By stimulating specific pathways, the
output of the source localization system can be evaluated by comparing it to the known
positions of the stimulated pathways. The experiments can also be used to validate the
influence of different constraints on the performance, as predicted by the simulation results.
1.5 Organization of the Document
The work presented in this thesis is divided into several individual studies, each of which
stands independently but flows logically from the previous ones. Each of these studies
constitutes a separate chapter, with its own introduction, method, results, and discussion
sections. The remainder of this document is therefore organized as follows. Citations in
parentheses refer to publications by the author that relate to each chapter.
Chapter 2 provides an overview of the relevant literature on bioelectric fields, source
localization, and peripheral nerve interfacing.
Chapter 3 describes a method for reducing the number of variables in the inverse
Introduction 9
problem ([203, 206, 205]).
Chapter 4 provides a simulation study of the peripheral nerve source localization
problem ([204, 207]).
Chapter 5 describes the experimental component of this work, as well as the per-
formance of the source localization approach on experimental data, whereas Chapter 6 is
concerned with an alternative approach to the source localization problem ([200, 201]).
Chapter 7 uses the 56-contact cuff used in this work to investigate the influence of the
number and location of the recording contacts on the spatial selectivity of the cuff ([202]).
Chapter 8 provides a discussion of the results, and the main conclusions are summa-
rized in Chapter 9.
Chapter 2
Literature Review
2.1 Origin of Bioelectric Fields
Peripheral nerves consist of axons linking the body’s sensors and effectors with the more
proximal components of the nervous system. Specifically, these axons can be those of auto-
nomic, dorsal root ganglion, or ventral horn neurons [62]. The axons and their associated
Schwann cells are encased in a collagen matrix, the endoneurium. Nerve fibers are grouped
into fascicles according to a topographical organization, and the endoneurium of each fas-
cicle is surrounded by the perineurium, which acts as a blood-PNS barrier. The fascicles
are then held together by the epineurium, a collageneous tissue that provides most of the
overall nerve’s structure and strength. The axons within a peripheral nerve can be either
myelinated or not, and exhibit a range of diameters. The combination of these characteris-
tics produces conduction velocities ranging from approximately 0.4 to 120 m/s. Myelinated
axons are faster and tend to be used in tasks when speed is important, such as the control
of movement or the perception of rapidly changing stimuli. Conversely, unmyelinated axons
are generally involved in the control of smooth muscle and the perception of slower events
such as pain and temperature changes. The nomenclature of axon fibers is based both on
the size and speed of the fibers and their function. Cutaneous fibers are refered to as, in
10
Literature Review 11
order of decreasing size, Aα, Aβ, Aδ, B, and C. Muscle sensory nerves are classified as Ia,
Ib, II, III, and IV, whereas muscle motor nerves are known as α, β, and γ, again in order of
decreasing size.
In all cases, information travels along nerves in the form of action potentials (APs),
which are temporary changes in a neuron’s transmembrane voltage. When an AP occurs at
a given point along a neuron’s axon, it results in charge redistributions that cause another
AP to be initiated further down the axon. This process repeats itself, allowing the AP to
“travel” down the axon. The frequency of the APs is responsible for encoding information, for
example the intensity of a stimulus [130]. The mechanism underlying the generation of an AP
is the opening of ion channels in the cell membrane, allowing ions to temporarily redistribute
themselves and alter the transmembrane voltage. This movement of ions produces small
electric currents across and along the membrane. Because the extracellular medium is a
volume conductor, these currents generate electric fields in the vicinity of the cell [92].
Therefore, it is possible to obtain information about the activity of the cell by recording
potentials in the extracellular medium, a small distance away from the cell. Because electric
fields add linearly, the activity of a population of cells can also be detected from further
away, yielding for example electroencephalographic (EEG) recordings, which are reflections
of brain activity obtained from the scalp.
The currents arising from ion movements during APs are nonconservative currents
resulting from the conversion of chemical energy to electrical energy [92], and are known
as impressed currents. This electrical activity imposes an electric field in the surrounding
volume conductor, and this field in turn creates a conduction current, known as the return
current, which prevents the impressed current from resulting in a charge buildup. Denoting Ji
the impressed current density, E the electric field, and σ the conductivity of the extracellular
medium, the total current density J can be expressed as shown in Equation 2.1 [92].
J = Ji + σE (2.1)
Literature Review 12
Note that in practice the conductivity is not constant throughout the medium and
in some regions is a tensor rather than a scalar (i.e. the medium is anisotropic). Because
of the frequency band of the bioelectric signals under consideration and the properties of
the medium, it has been shown that these bioelectric phenomena can be examined under
quasistatic conditions [131]. In other words, at any given instant, the fields can be considered
stationary and any time-dependent effects can be ignored. As a consequence, the electric
field can be expressed as the negative of the gradient of the potential field Φ, such that
Equation 2.1 becomes Equation 2.2.
J = Ji − σ∇Φ (2.2)
According to Maxwell’s equations under quasistatic conditions, the divergence of the
total current density J must be 0 (Ampere’s law, using the fact that the divergence of a curl
is 0). Taking the divergence of Equation 2.2 therefore leads to Equation 2.3.
∇.(σ∇Φ) = ∇.Ji (2.3)
This is known as Poisson’s equation, which governs the potential distribution in the
extracellular medium around active nerves. While analytical solutions to Poisson’s equation
are available and useful for simple cases (for example, sources in an infinite homogenous
medium), they quickly become unwieldy for more complex situations (a more complete
mathematical treatment of this issue can be found in [92]). In more realistic scenarios, the
equation is usually solved using numerical techniques (see Section 2.2.1 for more details).
2.2 Bioelectromagnetic Source Localization
Bioelectromagnetic source localization problems are a type of inverse problems aiming to
determine the location of biological current sources within a volume, based on electric or
magnetic potential measurements obtained from the surface of that volume. Until now, the
Literature Review 13
main applications of these problems were to determine the location of brain activity using
EEG and/or magnetoencephalographic (MEG) measurements, or to study the distribution
of electrical activity on the heart using electrocardiogram (ECG) measurements. The very
high temporal resolution of the source localization approach makes it an attractive method
in these contexts. Although the problem discussed in this thesis is to localize sources within
a peripheral nerves, the algorithms used for EEG/MEG source localization remain valid.
Before the inverse problem can be solved, it is necessary to be able to solve the
forward problem, that is to say to compute the measurements that would result from a known
source distribution. Once this is achieved, the inverse problems algorithms are divided into
two broad classes: equivalent dipole models and distributed linear models. The underlying
assumption of equivalent dipole models is that the surface measurements can be modeled
using a small number of current sources. The number of sources is chosen in advance,
and an optimization problem is solved to determine which source locations, magnitudes
and orientations will produce measurements most similar to the recorded ones. Distributed
linear models, on the other hand, do not make any assumptions about the number of sources.
Rather, the activity is estimated at a large number of fixed points forming a 3D grid in the
volume. This approach corresponds to estimating the direction and magnitude of a large
number of dipoles, each of which has a fixed location. Because the number of points in
the solution grid is typically much greater than the number of measurements, the problem is
severely underdetermined. In other words, any set of surface measurements can be explained
by infinitely many source configurations. To overcome this problem, various constraints can
be imposed on the solution. These constraints can be purely mathematical in nature, or
can be based on anatomical or physiological information about the problem. Additionally,
the sequential entries in a measurement time series can be viewed as independent, or as
belonging to a system evolving in time. In the latter case, temporal information can be used
to constrain or guide the solution. In all distributed systems, the ill-posed nature of the
problem creates the need for regularization, which can take various forms corresponding to
the different algorithms. In all of those cases, however, the choice of regularization parameter
Literature Review 14
has a considerable impact on the solution, and therefore a number of methods exist to help
guide this choice.
2.2.1 The forward problem
Both equivalent dipole methods and distributed linear methods rely on the availability of a
model that allows us to compute the measurements resulting from a given source configu-
ration. This computation is known as solving the forward problem [48]. It is a well-defined
problem, and can be solved analytically if the geometry and properties of the volume con-
taining the sources are sufficiently simple [109]. On the other hand, accuracy in the forward
problem is important for the quality of the inverse solution, so models with realistic geome-
tries are often preferable to simplified ones [50, 196, 100, 19]. When the geometry becomes
too complex for analytical solutions to be convenient, the forward problem is instead solved
with the help of numerical methods, such as boundary element or finite element methods
(BEM and FEM, respectively). BEM only discretizes the boundaries between regions of dif-
ferent conductivities, rather than the whole three-dimensional volume [104, 23]. In contrast,
FEM must discretize the whole volume into a three-dimensional mesh. The advantage of
the BEM approach is that possible source locations are not constrained by the shape of a
mesh, as they are in FEM. They can be left unconstrained, which is useful for equivalent
dipole methods, or organized into a regular grid of arbitrary coarseness, which is useful for
distributed linear methods. BEM is the method most commonly used in EEG/MEG source
localization studies (e.g. [109, 163, 194]). On the other hand, the method is not as well suited
as FEM for dealing with volumes that contain regions of anisotropic conductivity, and FEM
is therefore the method of choice in those cases [10, 193, 5, 160]. The operator mapping
electrical activity at any location in the region of interest to the measurements is known as
the leadfield, and it is computed by solving the forward problem. Different quantities can
be chosen to reflect this electrical activity, and it follows that different source models can be
used when solving the forward problem.
Literature Review 15
The finite element method
The distribution of bioelectric potentials in a conducting volume is governed by Poisson’s
equation, as given by Equation 2.3 of Section 2.1, and subject to Dirichlet or Neumann
boundary conditions, or both. Although this equation has an analytical solution in the case of
a homogeneous region of isotropic conductivity, the analytical approach becomes intractable
when the volume is made up of several regions of different geometries and conductivities,
creating the need for a numerical method such as FEM. The first step in the finite element
(FE) solution is to define an approximation Φ to the potential field Φ in terms of a set of
basis functions, as shown in Equation 2.4 [105].
Φ =n∑
a=1
φaNa (2.4)
The set of piecewise-smooth functions Na spans an n-dimensional space. The criterion used
to select the best such approximation is the minimization of a weighted residual integrated
over the volume Ω, leading to the problem in Equation 2.5.
∫
Ω
waR dΩ = 0, a = 1, 2, ..., n (2.5)
where R = ∇.(σ∇Φ) + Iv is the residual, Iv is the current source density, and wa are the
weighting functions. Using the divergence theorem and substituting the expression for R
into equation 2.5, we obtain Equation 2.6
∫
Ω
∇wa.σ∇Φ dΩ =
∫
Ω
waIv dΩ +
∮
Γ
wa[n.σ∇Φ] dΓ (2.6)
Here, Γ is the boundary of Ω and n is the normal component of the current density.
The weighting functions wa are usually selected to be equal to the functions Na, and
this choice is known as the Galerkin method. By substituting the definition from Equation
2.4 into Equation 2.6, we finally obtain Equation 2.7
Literature Review 16
n∑
b=1
3∑
i,j=1
[∫
Ω
∂Na
∂xi
σij
∂Nb
∂xj
dΩ
]
φb =
∫
Ω
NaIv dΩ +
∮
Γ
Na[n.σ∇Φ] dΓ (2.7)
x1, x2, and x3 are the three Cartesian coordinates. Note the presence of both Na and Nb.
Equation 2.7 is established for each value of a, ranging from 1 to n. Within each of these
equations, all other weighting functions appear because of Equation 2.4, hence the index b,
which also ranges from 1 to n. Using the boundary conditions and appropriately defined
weighting functions, Equation 2.7 can be written as a linear system where the only unknowns
are the coefficients φ1...n (the vector φ in Equation 2.8, see [105] for details).
Kφ = F (2.8)
The matrix K is know as the stiffness matrix and F is the force vector, these terms
having their origin in structural engineering. K contains the information about the geometry
and conductivities of the volume, whereas F reflects the boundary conditions and sources
within the volume. Once the coefficients have been computed, the final solution can be
recovered using Equation 2.4.
The remaining component of the process is to partition the volume into a mesh
composed of geometrical elements such as tetrahedra or wedges. The key to making the
method efficient is to define each basis function Na so that it is non-zero at one mesh
node and zero at all the other nodes. Between the non-zero node and its neighbours, the
function is generally a polynomial. With a definition of this type, the system in Equation
2.8 become sparse and can be readily solved. By constructing the mesh out of irregularly
shaped geometrical elements, the method is able to deal with complex volume geometries.
Furthermore, the conductivity tensor is defined separately for each mesh element, making it
possible to incorporate regions of different conductivities as well as anisotropic conductivities.
Literature Review 17
The boundary element method
The starting point of BEM is to segment the inhomogeneous region into sub-regions of
homogeneous conductivity. The effect of sources on the potential at the outermost surface is
then in effect computed by replacing sources within each region by equivalent distributions
of sources on the boundary between that region and the one containing it. The potentials
on all the surfaces are governed by Equation 2.9 [50, 99, 109].
σ0Φ∞(r) =(σ−
j + σ+j )
2Φ(r) +
1
4π
m∑
i=1
(σ−i − σ+
i ).
∫
Γi
Φ(r’)ni(r’).d/d3dr’ (2.9)
where Φ∞ is the potential field that the source distribution would produce in an infinite
homogeneous region of unit conductivity σ0 (this can be computed analytically), d = r - r’
(with magnitude d) is the distance between the observation point r and the source point r’,
Γi is the boundary between the ith and i+ 1th regions, m is the total number of boundaries
between regions, r is a location on Γj, and σ−j and σ+
j are the conductivities inside and
outside the surface Γj, respectively. Equation 2.9 can be written more simply by noting that
the right-hand side of the equation is a linear operator on Φ(r), leading to the formulation
L(Φ(r)) = Φ∞(r) [109]. Like in the FEM, the problem can then be cast as the minimization
of weighted residuals integrated over all the boundaries (Equation 2.10).
∫
(L(Φ(r)) − Φ∞(r))w(r)dr = 0 (2.10)
Then, the solution Φ and the weighting function w(r) are expressed as expansions in terms
of (different) basis functions (Equations 2.11 and 2.12).
Φ(r) ∼=n∑
a=1
φaNa(r) (2.11)
w(r) =n∑
a=1
βaψa(r) (2.12)
Literature Review 18
The solution basis functions are defined as in FEM, which is to say that each one is equal
to 1 at one node and to 0 at all the others, usually with a polynomial interpolation between
nodes. Using these definitions and Equations 2.10 to 2.12, we can derive the linear system
of equations shown in Equation 2.13 (see [109] for details).
(ψ1(r),Φ∞(r))
. . .
(ψn(r),Φ∞(r))
=
(ψ1(r), L(N1(r))) . . . (ψ1(r), L(Nn(r)))
. . .
(ψn(r), L(N1(r))) . . . (ψn(r), L(Nn(r)))
.
φ1
. . .
φn
(2.13)
The linearly independent basis functions ψa of the residual weighting functions are
often chosen to be equal to the solution basis functions Na (Galerkin form), or alternatively
can be defined as a set of Dirac delta functions each centering on one node (collocation form)
[109]. By carefully choosing the basis and residual weighting functions in this manner, the
first two matrices in Equation 2.13 can be conveniently computed, leaving a linear system
(Equation 2.14) whose only unknowns are the coefficients φa, which in turn yield the solution
Φ.
G = Hφ (2.14)
Like the matrix K in FEM, the matrix H depends only on the geometry and conductivities
of the region, not on the sources present, and therefore can be precomputed. The resulting
BEM system is smaller than in FEM but, unlike in FEM, is dense, not sparse [23].
Source models
Whether equivalent dipole or distributed linear methods are used, bioelectric source local-
ization problems typically seek to describe the distribution of currents in a region. The goal
of the forward problem is therefore to compute the measurements that would result from a
current dipole at a given location. In other words, the sources are modeled as equivalent
dipoles that reflect the current density in the vicinity of their position.
Literature Review 19
If FEM is being used, the principle of reciprocity can be used to more rapidly compute
the influence of a current dipole in each mesh element on the measurements. The principle
states that the potential measured between two surface points in the presence of a unit cur-
rent dipole is equal to the magnitude of the electric field that would occur at the dipole’s
position and orientation if a unit current was created between the two surface points. As
a consequence, by placing a current source at an active electrode and a current sink at the
reference electrode and computing the resulting electric field throughout the region, the in-
fluence on the active electrode of current dipoles in every mesh element can be obtained. The
number of potential distributions to compute is therefore equal to the number of electrodes,
rather than the number of mesh elements [193]. On the other hand, if BEM is used, the
principle of reciprocity is not as beneficial, but the positions of the dipoles are not restricted
by an FE mesh and can be chosen as the investigator sees fit (e.g. the dipoles can be placed
on a regular grid whose coarseness is appropriate to the problem at hand).
It is also possible to model the sources as something other than current densities.
Grave de Peralta Menendez et al. [42] have proposed a framework, called electrical analysis
(ELECTRA), which establishes relationships between different possible types of sources to
be recovered and the leadfields corresponding to each one. At its core, ELECTRA relies on
the observation that only irrotational sources are capable of producing measured potentials
at the electrodes [39]. Three types of source models are consistent with irrotational current
sources: current density vectors (the current dipoles most often solved for in bioelectric source
localization problems), current source densities (the divergence of the current density vector
field), and electric potentials. By providing a framework relating the different source models,
ELECTRA makes it possible to solve for any one of them. The current source density and
potential are both scalar fields, such that using these source models can reduce the number of
variables in the inverse problem by a factor of three. Nonetheless, the current density vector
source model remains the one used most often in the literature, perhaps because it has the
most direct relationship with the neuronal transmembrane currents. In the remainder of this
thesis, unless otherwise noted, the source model used is the current density vector.
Literature Review 20
2.2.2 Equivalent dipole methods
Equivalent dipole models rely on scanning the solution space and solving a forward problem
for each location under consideration. The location that results in simulated measurements
that best correspond to the real ones is selected as the solution. The drawback of this
approach is that an exhaustive search of all possible locations can quickly become impractical,
particularly when the method is expanded so that the solution consists of more than one
dipole. Instead of an exhaustive search, therefore, optimization algorithms are used to find
the dipole locations that result in the smallest error [70, 63, 186]. In order to further ease
the process, Mosher et al. proposed uncoupling the unknown variables that have a nonlinear
effect on the measurements (the dipole locations and orientations) from those that have
a linear effect (the dipole magnitudes). The search algorithm is used only to determine
the values of the nonlinear parameters. The linear ones can be factored out of the process
and, once the nonlinear values have been solved for, can be estimated using a pseudo-
inverse approach similar to the one used in the distributed linear methods presented in the
next section [110]. This uncoupling strategy is part of the widely used MUltiple SIgnal
Classification (MUSIC) algorithm. Those same authors later proposed a scheme the avoids
having to search for all the dipoles simultaneously. In the recursive MUSIC algorithm (R-
MUSIC), sources can be identified one by one: the source that comes closest to explaining
the data is selected first and retained, the search is repeated to find the source that best
explains the remaining part of the data, and the process is repeated until all the dipoles have
been localized [108]. R-MUSIC also extended the concept of a source from a single dipole to
a set of several synchronized dipoles.
Scherg et al. [149] have also proposed a forward model that, in addition to the location
and orientation of the dipoles, includes the temporal waveform of activation at each dipole.
The recorded data from a block of time rather than a single instant is then used to evaluate
the ability of a given set of model parameters to explain the data. In other words, temporal
information is incorporated into the problem in order to increase the likelihood of obtaining
Literature Review 21
a plausible solution.
Regardless of the search method employed, the biggest difficulty inherent in equivalent
dipole methods is that number of dipoles must be chosen a priori. Even if the investigator
has some knowledge about the phenomenon causing the electrical activity, estimating the
number of sources may be quite difficult. Some authors have suggested adding dipoles one at
a time, until the amount of unexplained activity stops decreasing significantly [148]. Another
possibility is to use additional imaging modalities to guide the decision. For EEG/MEG
source localization, fMRI has been used to estimate the number of different active areas,
and that information can be used either to choose the number of dipoles to be localized
[30, 74, 101] or to evaluate the plausibility of sources localized without using the fMRI
information, and to adjust the number of dipole if necessary [1]. The main problem with
this approach is that the relationship between the neural activity underlying EEG/MEG and
the hemodynamic response underlying fMRI is not well understood, thereby making the use
of fMRI information in EEG/MEG source localization potentially problematic [189, 25, 90].
For example, EEG/MEG and fMRI are not always equally sensitive to different source
configurations, meaning that it is possible for an area to be detected as active by one method
but not the other [104]. Furthermore, although EEG/MEG have a temporal sensitivity on
the order of the millisecond, the hemodynamic changes that fMRI responds to are much
slower. In a situation where brain activity associated with a given task is spread across
several areas, each of which may be active at different times, it is impossible to use fMRI
information without losing the fine temporal resolution that is the most attractive feature
of EEG/MEG.
An alternative to using other imaging modalities is to use the mathematical properties
of the recorded signals themselves to choose the number of dipoles. For example, Mosher
et al. use the eigendecomposition of the data recorded over a time interval to estimate the
dimensions of the signal subspace and noise subspace, then use this information to select
the number of dipoles (the relationship between the leadfield of each location and the signal
subspace is then used to estimate the location of the dipoles)[110]. Other spatio-temporal
Literature Review 22
decompositions that have been proposed include common spatial patterns decomposition,
which relies on comparing the signal of interest with a control signal to identify important
components [72, 73], and independent component analysis [71].
Use of equivalent dipole methods to obtain time series of activity
As mentioned previously, the temporal resolution of EEG and MEG is the most attractive
feature of bioelectromagnetic source localization methods. When using equivalent dipole
methods, the time series of dipole magnitudes can be recovered in a straightforward manner
once the dipoles have been localized. This is made possible by the linear relationship between
those magnitudes and the measurements. In the MUSIC algorithm, the complete relationship
between the measurements and the dipole magnitudes is described by Equation 2.15, using
notation from the original paper [110].
B = H(L,M)S (2.15)
The matrix H represents the influence on the measurements B of a set of dipoles whose
locations and orientations are given by L and M, respectively. S represents the magnitudes
of the dipoles. Computing the entries of H for given values of L and M corresponds to
solving the forward problem. Finding the values of L and M that minimize the difference
between the recorded measurements F and B (the simulated measurements) is a nonlinear
search problem, and is at the heart of equivalent dipole methods, as describe in the previous
section. Once this step has been accomplished, however, the magnitudes S can simply be
obtained as shown in Equation 2.16.
S = H†F (2.16)
H† denotes the pseudo-inverse of H. If the dipole locations are assumed to be constant
over a certain time interval, then Equation 2.16 can be applied at each time instant and
an activation time series obtained. These high-resolution time series are ideally suited to
Literature Review 23
studying the temporal relationships and frequency components of different brain regions
during a given task [1, 111].
2.2.3 Distributed linear methods
Distributed linear methods represent a different approach to the source localization prob-
lem, and accordingly the set of challenges associated with them is distinct from the issues
encountered with equivalent dipole methods. As discussed above, the equivalent dipole prob-
lem involves a complex but well-determined nonlinear optimization problem. In contrast,
distributed linear methods formulate the problem as a linear system of equations. Indeed,
since the location and orientation of the dipoles are fixed, only their amplitudes remain to
be estimated. Because electric fields add linearly, the relationship between the dipole ampli-
tudes and the measurements is linear. The problem can therefore be expressed as shown in
Equation 2.17.
d = Lj + ǫ (2.17)
d is an Mx1 vector containing the recorded data from the M electrodes contacts, j
is an 3Nx1 vector whose entries represent the magnitudes of the current dipoles, and L is
a Mx3N matrix known as the leadfield matrix whose entry (i,j) represents the influence of
a unit current at dipole j on the potential recorded at electrode i. The leadfield matrix has
three columns for each of N dipole locations, corresponding to the three orthogonal dipole
orientations. ǫ is the additive noise.
The difficulty in distributed linear methods comes from the fact that that system of
equations in Equation 2.17 is severely underdetermined. To obtain a satisfactory solution,
additional information must be incorporated into the problem. This information takes the
form of constraints on the solution, which can be either mathematical in nature or anatom-
ically derived.
Literature Review 24
Weighted minimum norm methods
The fundamental idea underlying all minimum norm solutions to the bioelectric source local-
ization problem is that, out of the infinite set of solutions to the underdetermined problem,
the chosen solution should be the simplest one. In other words, the goal is to find the sim-
plest configuration of sources that explains the measurements. The notion of “simplest” is
then formalized as the minimization of a given norm.
The use of the solution minimizing the l2 norm was first proposed in the context
of bioelectromagnetic source localization by Hamalainen and Ilmoniemi [49]. Many of the
source localization methods that have been proposed have been shown to be equivalent to
this classical minimum norm approach if no constraints are added [56]. Although this method
could achieve reasonable performance in two dimensions, it was not adequate for realistic
three-dimensional problems [122]. The main drawback of the minimum norm approach is
that sources near the electrode positions have a much greater influence on the measurements
than deeper sources, and the resulting solutions therefore tend to feature mainly superficial
sources. Deep sources are usually severely mislocalized. This problem led to the development
of weighted minimum norm approaches. Instead of minimizing the l2 norm of the solution,
the l2 norm of a weighted version of the solution is minimized instead. A variety of weight
matrices can be chosen, corresponding to different constraints on the solution. The simplest
option is to use a diagonal weight matrix that normalizes each leadfield column by its norm
[122]. Since sources near the electrodes correspond to leadfield columns with larger norms,
this weighting has the effect of compensating for the distance between sources and the
electrodes, thereby improving the localization of deep sources. An alternative basis for
weighting source locations is to use information from other modalities. Specifically, fMRI
images can be used to obtain an estimate of which locations are more likely to be active,
and MRI and PET images provide anatomical information that can be used to restrict the
solution space [22, 87, 127, 86]. By using a non-diagonal weight matrix, more complex spatial
features can be incorporated into the constraints. The most common example of this is the
Literature Review 25
low resolution brain electromagnetic tomography (LORETA) method, which uses a discrete
Laplacian weight matrix to impose a smoothness constraint [124]. LORETA is better able
to localize sources in three dimensions than the weighted minimum norm solutions described
above. Its disadvantage, however, is that the resulting solutions are blurred, representing a
trade-off between accuracy of peak locations and spatial resolution. The anatomical validity
of the smoothness constraint is also debatable unless the solution grid is very fine [120].
Another method that uses a non-diagonal weight matrix is known as local auto-regressive
averages (LAURA), which couples elements of the solution based on the distance separating
them and the rate of change of electric fields with distance [42].
Having chosen an appropriate norm to evaluate the solution, the problem remains that
the inverse problem is ill-posed and extremely sensitive to noise. Minimum norm solutions
(weighted and non-weighted) fit the data exactly, so when this data is noisy, the attempt to
fit the noise will lead to significant errors in the solution. This problem must be addressed,
through a process known as regularization. In essence, regularization involves accepting
some error in the fit to the data, in exchange for increased stability of the solution [54]. Two
types of regularization methods are commonly used in bioelectromagnetic source localization:
Tikhonov regularization and truncated singular-value decomposition (TSVD). In Tikhonov
regularization, the problem is formulated as shown in Equation 2.18 [179, 54].
j = arg minj‖C−0.5
ǫ (Lj − d)‖2 + λ‖Hj‖2 (2.18)
j is the current source distribution estimate, Cǫ is the noise covariance matrix (the
noise is assumed to be zero-mean Gaussian), and H is the weight matrix. The expressions
for the various weight matrices described in the previous paragraph are given in Table 2.1.
λ is the parameter that balances the minimization of the residual and the minimization of
the weighted solution norm. A small value of λ will give preference to fitting the data. As
λ is increased, more importance is given to the a priori constraints. In other words, the
noisier the data, the larger the value of λ should be, so that more importance is given to the
Literature Review 26
Table 2.1: Commonly used weight matrices for weighted minimum norm methods.
Method Weight Matrix Expression
Minimum Norm I
Weighted Minimum Norm
Hwmn|Hii = LT1..n,iL1..n,i;Hij = 0, i 6= j
LORETA BHwmn
LAURA HwmnA⊗ I3 where Aii = NNi
∑
k⊂Vi
d−3
ki
Definitions B is the Laplacian operator
⊗ is the Kronecker product
Vi is the set of Ni neighbours of location i, of size at most N
dki is the distance between locations k and i
investigator’s prior knowledge of what characteristics the solution should have. Equation
2.18 has a closed-form solution, which is shown in Equation 2.19 [126].
j = (HtH)−1Lt[L(HtH)−1Lt + λCǫ]−1d (2.19)
As for the TSVD method, it can be derived starting from the expression for the
unregularized, unweighted solution, which is simply the Moore-Penrose pseudo-inverse of
the leadfield, multiplied by the data (Equation 2.20).
j = L†d (2.20)
If the weight matrix is square and non-singular, it can be incorporated in Equation
2.20 by first re-writing Equation 2.17 as Equation 2.21, such that the new solution is as
shown in Equation 2.22.
d = (LH)(H−1j) + ǫ (2.21)
j = H(LH)†d (2.22)
Literature Review 27
Note that Equation 2.22 is equivalent to Equation 2.19, with λ set to 0 [37]. The
next step involves a singular value decomposition, which decomposes a matrix into two
orthonormal matrices U and V and a diagonal matrix S. U and V can be interpreted as
containing basis vectors describing the structure of the data, whereas the singular values
in S reflect the relative importance of these different components. Now, using the singular
value decomposition LH = USVT, Equation 2.22 can be written as shown in Equation 2.23.
Here, S−1 is defined as the matrix in which each nonzero entry of the diagonal matrix S has
been replaced by its inverse.
j = H(VS−1UT)d (2.23)
The instability of the solution in an ill-posed inverse problem is related to the small
singular values in the matrix S [54]. Based on this insight, the TSVD regularization method
functions by setting to zero the entries of S that are below a certain threshold. The solution
is then obtained by using the modified singular values matrix S in Equation 2.23. Tikhonov
regularization can be understood within the same framework of singular value decomposition,
where it corresponds to adding λ to each singular value in S, with λ being assigned a value
that is much greater than the smallest singular value but much smaller than the largest
singular value. This will reduce the instability associated with taking the inverse of a very
small number, without having a significant impact on the large singular values. Several
methods exist to help with choice of λ; they are described in Section 2.2.5.
Iterative minimum norm methods
Gorodnitsky et al. have proposed an algorithm by which more focal solutions can be achieved
than the ones resulting from the minimization of a weighted l2 norm. The algorithm, known
as the focal underdetermined system solver (FOCUSS), is essentially an iterative application
of a weighted minimum norm method (Equation 2.19), where the weight matrix at each
step is a diagonal matrix whose entries are based on the source estimates from the previous
Literature Review 28
iteration [37]. In this way, locations that were found to have significant activity in one
iteration are favoured in the next iteration. As the number of iterations increases, the
activity becomes concentrated in a small number of locations, and tends to zero elsewhere.
The process stops when the solution is no longer changing significantly from one iteration to
the next, or if the number of nonzero elements in the solution starts to increase. FOCUSS
can obtain at most as many focal solutions as there are measurements [37, 38].
Before FOCUSS can be applied, an initial estimate must be obtained, which can
be done using any distributed linear method. The original paper proposed a variant of the
depth-weighted minimum norm solution [37]. More recently, good results have been obtained
using sLORETA (see section 2.2.3) as the initial estimate [88]. FOCUSS is very sensitive to
the initial estimate, so having a good estimate is critical to the success of the method. Addi-
tionally, the regularization of the algorithm can be difficult, making it potentially vulnerable
to noise. FOCUSS can be regularized using the Tikhonov or TSVD methods [37], but the
process is made more difficult by the fact that a good choice of regularization parameter
is needed at every iteration. Rao et al. [134] have proposed parameter choice methods to
overcome this problem (see Section 2.2.5).
Standardized minimum norm methods
The classical minimum norm solution given by Equation 2.19 inherently has a magnitude
distortion for each element in the estimated solution. This can be seen using the concept of
the resolution matrix. If the result of the (weighted or unweighted) minimum norm solution
is as given in Equation 2.24, then the relationship between the true source distribution and
the estimated source distribution can be obtained by combining Equations 2.17 and 2.24, as
shown in Equation 2.25.
j = Td (2.24)
Literature Review 29
j = Td = TLj = Rj (2.25)
For perfect reconstruction, the resolution matrix R should be the identity matrix,
but because of the ill-posed nature of the problem there does not exist a transfer matrix T
that results in R being the identity. Pascual-Marqui proposed using the bias information
contained in R to normalize the results of an unweighted minimum-norm solution, as shown
in Equation 2.26.
jT
l (Rll)−1jl (2.26)
The resulting algorithm is known as standardized LORETA (sLORETA) [123]. Al-
though there is no explicit smoothness constraint like in the original LORETA algorithm (see
Section 2.2.3), sLORETA nonetheless results in a smooth solution with well-localized peaks.
In the noiseless, single-source case, sLORETA has been shown to have zero localization error
[123, 162, 43].
The standardization method used in sLORETA has also been incorporated into the
FOCUSS algorithm by Liu et al [88]. Their algorithm, known as standardized shrinking
LORETA-FOCUSS (SSLOFO), recomputes the resolution matrix at every iteration based
on the updated transfer matrix T, then applies the normalization described by 2.26. Note
that in this case, T results from a weighted minimum norm solution, as per the FOCUSS
algorithm, rather than the unweighted minimum norm solution used in sLORETA. The
resulting algorithm outperforms the original FOCUSS [88].
An alternative method using the resolution matrix was proposed by Grave de Peralta
Menendez et al. [41]. Instead of normalizing the solution elements individually using local
sections of the resolution matrix, they proposed creating an approximate inverse of the
entire resolution matrix. The unmodified resolution matrix is usually not invertible, but the
approximation can be obtained by inverting the sum of the resolution matrix and a diagonal
perturbation. Calling this perturbation D, the approximate inverse is shown in Equation
Literature Review 30
2.27, and the transfer matrix is updated as shown in Equation 2.28.
M = (D + R)−1 = (D + TL)−1 = D−1 − D−1T[I + LD−1T]−1LD−1 (2.27)
T := MT (2.28)
This method’s aim was to overcome the limitations of partial inversions such as
sLORETA, namely that they work better for single sources than for linear combinations
of sources.
Methods using lp norms
Although the minimum norm methods described so far usually minimize the l2 norm, this
does not have to be the case: a different lp norm can be chosen. In particular, several studies
have investigated the minimization of the l1 norm, which produces sparser solutions than
l2 [94, 95, 187, 165, 194]. Using these methods, solutions are obtained that have at most
as many nonzero elements as there are sensors. The main disadvantage of this approach is
that, unlike the l2 norm minimization problem, the equivalent l1 problem does not have a
direct solution. An optimization problem must be solved instead or, if the dipole orientation
is fixed, a simpler linear programming (LP) approach can be used [32, 187]. The practical
outcome is that the computational demands of l1-based methods are much higher than those
of l2-based methods.
Notable l1-based methods include the approach of Matsuura and Okabe [95], the
Minimum Current Estimate (MCE) introduced by Uutela et al. [187] and the lp Norm
Iterative Sparse Solution (LPISS) by Xu et al. [194]. Matsuura and Okabe minimized the
l1 norm under inequality constraints corresponding to an acceptable data misfit tolerance,
thereby making the procedure more robust in the presence of noise. MCE determines the
dipole orientations using a traditional weighted minimization of the l2 norm, then estimates
Literature Review 31
the dipole magnitudes by using a LP approach to minimize the l1 norm of the solution. The
minimization is constrained by a regularized equality between the estimated solution and the
data. LPISS is a FOCUSS-like algorithm that replaces the l2 solution norm in Equation 2.18
by an lp (p ≤ 1) norm, as shown in Equation 2.29. By using the l1 norm LPISS reconstructs
sparse distributions more reliably than FOCUSS. Because of the presence of both an l1 and
an l2 norm in Equation 2.29, the problem cannot be solved using LP, and a more complex
optimization problem must be solved. This algorithm is therefore particularly time intensive,
requiring roughly two orders of magnitude more computation time than FOCUSS.
j = arg minj‖Lj − d‖2 + λ‖j‖1 (2.29)
Local linear estimators
Most of the minimum-norm methods discussed so far can be described as global linear
estimators, because they produce an entire source distribution designed to account for the
measurements as well as possible. An alternative approach is to use local linear estimators,
which focus on the recovery of each target location individually, without explicitly minimizing
the overall misfit to the data [43]. The rows of the resolution matrix (Equation 2.25) provide
information about the bias of a given solution method for each individual location, as well
as about the influence of activity at other locations on the target location. Similarly, the
columns of the matrix provide information about the blurring that can be expected when
recovering a single point source. Local linear estimators function by constructing transfer
matrices designed to produce resolution matrices with specific properties, rather than to
minimize the data misfit. Each row of the transfer matrix is treated as an individual problem,
and produces a resolution matrix row (also known as a resolution kernel) with the chosen
properties. Specifically, the problem can be formulated as a minimization of the norm of the
resolution kernel under certain constraints, which vary from one type of linear estimator to
the other.
Literature Review 32
Examples of constraints that have been proposed include ensuring that the solution
has a gain of one at the target location (distortionless estimators [43]) or making the resolu-
tion kernel as close as possible to a delta function centered around the location of interest and
integrating to one (the oft-cited Backus and Gilbert method first introduced in geophysical
inverse theory [4] and improved in [40]). Note that the sLORETA algorithm is a local lin-
ear estimator, although it was presented separately in Section 2.2.3 because of its frequent
mention in the EEG/MEG source localization literature (it is also worth clarifying that
the original LORETA algorithm is a global linear estimator, and therefore fundamentally
different from sLORETA).
All of the local estimators in the previous paragraph are non-adaptive estimators,
meaning that the solution is derived independently of any data. Adaptive local estimators
have also been described, which assume some knowledge of the measurements. The best-
known example is the linearly constrained minimum variance (LCMV) spatial filter [188],
which is the “beamforming” technique common in radar and sonar technology. As in the
non-adaptive estimators, the idea is to obtain an estimate at each location by constructing a
spatial filter that excludes as much as possible signals originating from outside the location
of interest, while having unit gain at that location. Because of the underdetermined nature
of the problem, these filters do not have enough degrees of freedom to adequately exclude
the signals from all undesired locations. The LCMV approach formulates the problem as
the minimization of the variance of the filter output, constrained by the requirement of unit
gain at the target location. The variance of the filter output (i.e. of the sources) can be
expressed in terms of the signal covariance and the leadfield. Consequently, the beamforming
approach relies on knowledge of the signal covariance matrix and uses this information to
allocate the resources of the filters in a way that blocks the activity at the locations most
likely to interfere with the activity at the target location. The signal covariance matrix is
generally estimated directly from the data. If the underlying signal is stationary and a good
estimate of the signal covariance matrix is available, the beamformer can perform very well
[162]. If the assumption of stationarity cannot be justified, on the other hand, the method
Literature Review 33
may produce misleading results.
A commonly used variant of the LCMV beamformer is known as synthetic aperture
magnetometry (SAM) [137, 190]. In addition to estimating the source power at each location,
the contribution of the sensor noise to the source power estimate at that location is also
estimated, and the ratio of source power to noise variance is used as the final tomographic
image. The normalization by the noise variance improves the resolution of the algorithm by
emphasizing regions with statistically significant brain activity.
2.2.4 The dynamic source localization problem
In addition to constraints derived from anatomical and mathematical considerations, the
temporal behaviour of the bioelectromagnetic sources can be taken into account in order to
obtain realistic solutions. This is a natural approach considering that the fine temporal res-
olution of bioelectromagnetic measurements is one of their main advantages when compared
to other imaging modalities (for example MRI and PET), and makes them useful in the
examination of temporal characteristics such as frequency, synchronization, and the order of
activation of different regions. Two approaches have been used to impose temporal relation-
ships between the solutions at various time instants: grouping several time instants together
into a single large linear system [24, 154, 153, 155], or using the results of one time instant to
obtain initial conditions for the solution of the next instant [195, 33]. Additionally, methods
have been proposed in which the localization algorithms are applied to a time-frequency
decomposition of the signal, in order to focus on sources producing a certain frequency band
[84, 91, 174, 161, 35, 106, 118].
Coupled spatio-temporal system approach
Ideally, the whole spatio-temporal system could be expressed and solved as a single linear
system, by concatenating the measurement and source vectors for all time instants [8, 24,
154]. This is shown in Equation 2.30, which is very similar to Equation 2.17, with the
Literature Review 34
exception that the vectors dc, jc, and ǫc include the complete information about the system
at all time instants in the observation time window (i.e. dc = [d(1)Td(2)T...d(n)T]T, with
equivalent definitions for jc, and ǫc). Lc is a block diagonal matrix, where each block is equal
to L.
dc = Lcjc + ǫc (2.30)
The advantage of this approach is that temporal patterns can be included as con-
straints, in the same way that spatial constraints were included in the instantaneous prob-
lem. Since the solutions at different time instants are incorporated into a single vector,
the solution of the inverse problem can include a weight matrix that relates activations at
different times, according to an a priori pattern. One possible constraint is to impose tem-
poral smoothness. The obvious drawback of this approach is that the combined system is
extremely large and quickly becomes intractable, unless the number of variables per time in-
stant is very small. Efficient algorithms have been proposed to try to overcome this problem
[8, 153, 155]. Another notable approach within the framework of coupled spatiotemporal
systems was proposed by Greensite, who described how to explicitly incorporate a non-
informative temporal prior, rather than assume independence or smoothness between time
instants. That is to say, the Greensite prior can be used to eliminate any such assumptions
and reflect the fact that we have no information about the relationship between the time
instants or their independence [44].
State-space representation
An alternative way of incorporating temporal information into the problem is to use the
estimates of past time instants as initial conditions for the current time instant, under the
assumption of temporal smoothness. A weighted difference between the solution and the
initial condition can be incorporated as an additional term in Equation 2.18, as shown in
Equation 2.31, or similarly in Equation 2.32 [195].
Literature Review 35
jt = arg minjt
‖C−0.5ǫ (Ljt − dt)‖2 + λ1‖Hjt‖2 + λ2‖H(jt − jt−1)‖2 (2.31)
jt = arg minjt
‖C−0.5ǫ (Ljt − dt)‖2 + λ‖H(jt − φjt−1)‖2 (2.32)
Alternative formulations can include using derivative information from previous time
instants to obtain more precise initial conditions [117, 178]. Additionally, Yamashita et
al. point out that Equation 2.32 can be interpreted in terms of a state-space representation,
where the observation and state equations are shown in Equations 2.33 and 2.34, respectively
[195].
dt = Ljt + ǫt ǫt ∝ N(0, σ2Cǫ) (2.33)
jt = φjt−1 + ηt ηt ∝ N(0, τ2(HTH)−1) (2.34)
Equation 2.34 can be modified to incorporate more complex relationships between
time instants, as shown in Equation 2.35, where information from the previous p time points
is used.
jt =
p∑
i=1
Aijt−i + ηt (2.35)
Where the A matrices are state transition matrices. The advantage of this formulation
is that complex relationships between time instants can be expressed without increasing the
size of the system to be solved at each instant (unlike the coupled approach presented in the
previous paragraph). On the other hand, its drawback is that a poor estimate at one time
instant will lead to poor initial conditions in future time instants, potentially causing the
error to propagate.
Taking the implications of the state-space approach further, the same group suggested
using a Kalman filtering approach to solve the source localization problem [33]. In order
for the problem to remain tractable, the high-dimensional problem was decomposed into
Literature Review 36
a set of uncoupled low-dimensional problems. This requires a number of assumptions, for
example concerning the connectedness of different brain regions and the dynamical models
underlying the activity. The method has the potential to be powerful, on the condition that
the assumptions can be kept both realistic and informative.
Time-frequency approach
Another way to incorporate temporal information into the problem is to focus on limited
frequency bands. Sources corresponding to each frequency band can be localized separately,
which can be of considerable interest for studying issues such as oscillatory activity in a
specific region of the brain, or the synchronization of activity among different regions. Early
methods of this type used the Fourier transform (both amplitude and phase) of the recorded
signal to focus on specific frequencies and explain them using dipole fitting methods [84, 91,
174]. Because the Fourier transform is not appropriate for the analysis of non-stationary
signals, these methods led to ones based on various types of time-frequency decompositions,
for example wavelets. Different decompositions have been combined with dipole fitting
approaches [118], the MUSIC algorithm [161], beamforming approaches [20], and distributed
linear methods [35, 106]. In this last category, the study by Gonzalez Andino et al. suggested
using the time-frequency decomposition to identify times at which the frequency spectrum is
simplest and therefore likely to correspond to a small number of generators, then performing
source localization at those times.
2.2.5 Choice of regularization parameter
Both Tikhonov and TSVD regularization involve a parameter that has a significant impact on
the solution and must be carefully selected. The parameter is λ in Tikhonov regularization,
and the number of singular values to be truncated in TSVD. There is no simple way to find
the best values for these parameters, but several methods exist to guide the choice. We
restrict ourselves here to methods that have been used in the EEG/MEG source localization
Literature Review 37
literature; a more general review can be found in [54].
L-curve
The most commonly used parameter-choice method in the EEG/MEG source localization
literature is the L-curve method [53, 54]. It creates a plot of the norm of the residuals versus
the norm of the solution, as the regularization parameter is varied. In log-log scale, this plot
is typically shaped like an “L”, whose corner corresponds to the best tradeoff between the
simplicity of the solution and the data misfit.
In some cases, the log-log curve may not be a perfect “L”, in which case choosing the
point of maximum curvature may not yield the best regularization parameter. Rao et al.
observed this phenomenon in relation to the FOCUSS algorithm, and proposed a modified
L-curve method in which the parameter is restricted to values producing a residual norm
within a given interval [134]. This approach ensures that the parameter choice is always
reasonable and results in a more robust algorithm.
Cross-validation
Cross-validation is a common method in model selection problems. One entry (or a subset)
of the data is left out, the model is trained on the remaining data, and the predicted value
of the data that was left out is compared to the correct values. This is repeated for each
data point or subset, and the mean predicted error is used as a metric of the quality of
the model fitting. In a regularization context, the chosen regularization parameter is the
one that minimizes the cross-validation error (CVE). Pascual-Marqui suggested using CVE
minimization as the regularization method of choice for his sLORETA algorithm [121].
A variant of the cross-validation method, generalized cross-validation (GCV), is dis-
cussed in the regularization literature [54], but rarely mentioned in the context of bioelectric
source localization. One inverse electrocardiography study found the performance of GCV
to be similar to that of the L-curve [164].
Literature Review 38
AIBC
Many source locations algorithms have been derived using a Bayesian approach [152, 21, 31,
128, 147, 126, 180, 68]. It can be shown that under certain common assumptions (e.g. Gaus-
sian priors), the resulting algorithms are equivalent to the ones derived using the minimum-
norm least-squares approach described above. If more complex prior information needs to
be incorporated, however, or if we want to incorporate information from several models si-
multaneously [181], the Baysian approach is attractive because of its versatility. From a
regularization point of view, the Baysian formalism allows us to treat the regularization
parameters as model hyperparameters that need to be determined and who are associated
with their own prior probability distributions. One example of this approach is the use of
an approximation of the Akaike’s Bayesian Information Criterion (ABIC) by Yamashita et
al. [195, 33]. The ABIC is a metric for the evaluation of a statistical model, which includes
both a measure of the log-likelihood of the parameters and of the number of parameters. By
minimizing the value of the ABIC, a model is found that fits the data as well as possible
while still keeping the number of parameters as small as possible. Aside from the obvious
application as a regularization tool, Yamashita et al. point out that the ABIC can be used
as a way to compare several models (i.e. different source localization algorithms or differ-
ent priors). The actual minimization of the ABIC is performed by solving an optimization
problem, making this method potentially more time-consuming than the L-curve or cross-
validation. Furthermore, if probability distributions are not assumed to be Gaussian, the
use of the ABIC becomes analytically very complex.
ReML
An alternative statistical approach to the regularization problem has been suggested by
Phillips et al. [128, 126]. They advocate computing the restricted maximum likelihood
(ReML) solution, which in this case is equivalent to the more general method of using an
expectation-maximization (EM) algorithm [31]. This approach is well established in the
Literature Review 39
statistical literature, in situations where both parameters and hyperparameters of a model
must be estimated from the data. Here, the model parameters are the dipole magnitudes
describing the source distribution, and the hyperparameters are the variances of the source
and sensor signals (the ratio of which is the regularization parameter used in the previous
techniques). The algorithm consists of an iterative process in which the dipole magnitudes
and regularization parameters are each held fixed in turn, while the value of the other
component is optimized to best account for the data. This guarantees that the resulting
dipole and regularization parameters will account for the data in an optimal way, in the
absence of any additional information.
2.3 Peripheral Nerve Monitoring
As the previous section demonstrates, inverse problems of source localization are complex and
difficult to solve. They are therefore not a natural choice if simpler alternatives are available.
In the case of monitoring the activity of peripheral nerves, there are a variety of other
methods available [142, 115], but none of them are entirely satisfactory. Any given system
for extracting information from recordings of the electrical activity of a nerve can be divided
into two components: a recording device, and a set of mathematical tools for interpreting
the data obtained. Until recently, the majority of research in peripheral nerve recordings
has been focused on the development of better electrodes, while relying on traditional signal
processing techniques. Indeed, if the electrode signals do not contain the desired information,
no amount of processing will be able to compensate. The fundamental trade-off in the
design of peripheral nerve electrodes is to balance the amount of information obtained with
the increased complexity and invasiveness of the device. A survey of the available types of
electrodes reveals that they typically represent various attempts to deal with this trade-off
in a way that is satisfactory for a specific application.
Literature Review 40
2.3.1 Electrode types
Extraneural Electrodes
Extraneural electrodes do not penetrate into the nerve, and therefore have less chance of
damaging it than intraneural electrodes. On the other hand, because of the greater distance
between the nerve fibers and the recording site, the information that these electrodes provide
is less spatially specific.
The most common type of extraneural electrode is the nerve cuff electrode [166, 89,
112, 113, 167]. A nerve cuff is an insulating sheath that encircles the nerve, with electrical
contacts placed on the inner surface of the tube (Figure 2.2(a)). This configuration isolates
the electrode contacts from signals originating outside the nerve segment contained by the
cuff. In addition, cuff electrodes are easier to implant and less invasive than intraneural
electrodes. Because of these advantages, much attention has been paid to the potential of
nerve cuffs to serve as chronically implanted neural interfaces. Studies have examined the
effects of the cuffs on the underlying nerves after implantation periods of several months
[76, 80, 171, 46, 138, 11, 96, 139], and although minor nerve damage was observed in certain
studies, it was usually not severe enough to cause any functional deterioration. In particular,
a design known as the spiral cuff has been used, which consists of a “self-sizing” cuff that
adjusts to the diameter of the nerve and has been shown to cause less damage in chronic
implantations than cuffs with fixed sizes [112, 46, 138, 139]. In humans, nerve cuff electrodes
have been used for periods of up to 12 years [192].
The traditional contact configuration for a cuff electrode is the tripole configuration,
in which the potential at a ring contact at the centre of the cuff is measured using as reference
the average potentials from two other ring contacts, one at each extremity of the cuff. In
practice this derivation is implemented using either the quasi-tripole or the true tripole
configurations, which are shown in Figure 2.1. Sources outside a cuff produce a signal that
varies linearly along the length of the device, but this is not true of sources located inside the
cuff [170, 132, 133, 2]. The tripole configuration therefore attenuates signal produced outside
Literature Review 41
Figure 2.1: Two common implementations of the tripole configuration: the quasi-tripole
(a) and the true tripole (b). From [133].
the cuff, but not those from inside. A significant drawback of the device is that a cuff with a
single tripole configuration produces a single signal, which reflects all the activity anywhere
in the nerve, and makes the traditional cuff a device with very poor spatial selectivity. To
remedy this problem, alternative cuff designs can be used that include a higher number
of contacts. Depending on the number of contacts and their configuration, a certain level
of selectivity can be achieved (see Section 2.3.3 for more details). Whenever using cuff
electrodes, it is important to realize that the device acts as a filter on the recorded signal,
and that the geometry of the cuff and contact configuration have a significant impact on the
shape of the recorded signal [168, 173, 199].
A modified version of the cuff electrodes is the Flat Interface Nerve Electrode (FINE)
[183, 60, 150], which aims to improve selectivity by slowly reshaping the nerve into a flatter
configuration, such that every fascicle is close to the surface (Figure 2.2(b)). It has been
demonstrated that if the forces applied are small enough, the nerve can be successfully
reshaped without damaging it [184]. Although it has not yet been used in a clinical setting,
the FINE is a promising technology for selective recording and stimulation of peripheral
nerve signals, because it addresses the key issue of deeper sources having less influence on
the measurements than surface sources.
Other types of extraneural electrodes exist, although they are usually used in more
Literature Review 42
specific contexts than cuff electrodes. For example, epineurial electrodes (Figure 2.2(c))
are sutured to the surface of the nerve [115], and their stable positioning gives them good
selectivity for certain fascicles close to their location. On the other hand, they provide no
information about nerve fibers further away. Helicoidal electrodes (Figure 2.2(d)) are helix-
shaped electrodes [115] that are easy to place and remove, but whose selectivity is limited by
their open shape. Epineurial and helicoidal electrodes are usually used for stimulation rather
than recording, most notably for phrenic nerve stimulation [18] and vagus nerve stimulation
[97], respectively.
Intraneural Electrodes
Intraneural electrodes penetrate the nerve, making them much more spatially selective than
extraneural electrodes, but also more delicate to implant and more likely to damage the
nerve. Intrafascicular electrodes provide the best spatial selectivity available, by recording
directly from a small area inside a nerve fascicle. For example, the longitudinally implanted
intrafascicular electrode (LIFE) [83, 81, 82, 27, 103] consists of a thin insulated wire im-
planted directly into the fascicle. The recording site is a small length of wire where the
insulation is removed (Figure 2.3(a)). Although providing excellent functionality if only a
small region is of interest, these electrodes are difficult to implant, and implanting more than
a small number of them quickly becomes impractical. They are therefore not appropriate
for systems in which a large number of nerve fibers need to be monitored.
Other common types of intraneural electrodes are the penetrating microelectrodes and
microelectrode arrays [142, 115]. These exist in a variety of configurations, including 1D, 2D,
and 3D arrays. Single microelectrodes can be inserted percutaneously, while larger arrays
may require more invasive implantation procedures. These multielectrode arrays (MEA)
can contain more than 100 needle or wedge-shaped electrodes recording from within fascicles
(Figure 2.3(b)), and therefore currently represent the best option for monitoring a large
number of locations within a peripheral nerve [143, 144, 7, 191]. The major disadvantages
Literature Review 43
of MEAs are that they are very invasive, modifying the shape of the nerve and possibly its
conduction properties, they are difficult to implant, and they are difficult to manufacture,
requiring sophisticated microfabrication techniques. In addition, chronic experiments with
these devices have revealed poor stability of the recording over time, although stimulation
could be maintained with much higher stability [6].
The slowly penetrating interfascicular nerve electrode (SPINE) [182] is a different
electrode design, which is placed around the nerve like a cuff electrode but features inner
plates that penetrate the nerve (Figure 2.3(c)). The contacts are thus positioned between
the fascicles, rather than inside them. The penetrating elements are inserted into the nerve
slowly, minimizing damage. The electrode is therefore designed to combine the simplicity
and safety of a cuff electrode with intraneural recording capabilities. The trade-off is that
the number of contacts and the specificity of the information obtained is lower than in an
MEA.
Another nerve interfacing technology is the regeneration sieve electrode [29, 114, 79],
which is placed between the two stumps of a transected nerve. The device has numerous small
holes surrounded by electrode contacts. When nerve fibers regenerate, they grow through
the holes in the electrode, such that the contacts will be specific to individual axons or small
groups of axons (Figure 2.3(d)). While an interesting concept and useful for some studies in
animal models, there are few situations in which regenerative electrodes are appropriate for
use in humans, because of the heavy nerve damage involved and the time required for fiber
regeneration. Currently the only target application for these electrodes in amputees, where
the nerve has already been severed.
2.3.2 Signal processing techniques for extraneural measurements
In order to successfully control an assistive device, the signals recorded from a cuff electrode
must be processed and used to generate appropriate commands. The simplest way to ac-
complish this is to filter the signal so as to retain only the frequencies of interest, group
Literature Review 44
(a) (b)
(c) (d)
Figure 2.2: Types of extraneural electrodes: a) traditional cuff electrodes, b) FINE, c)
epineurial electrode and d) helicoidal electrode.
the measured samples into time bins of a few milliseconds, perform bin-by-bin integration of
the rectified signal, and finally compare the value obtained for each bin to a threshold [55].
The rectified and bin-integrated signal can also be used in conjunction with more sophis-
ticated techniques, for example by being used as the input to an artificial neural network
[102, 12, 16]. Alternatively, the variance of the contents of each bin can be computed (instead
of integrating), and Student’s t-tests performed between bins to detect statistically signifi-
cant increases that mark the presence of a signal [78]. Other approaches include subspace
decomposition methods, which use statistical information about a segment of the signal
to determine whether it should be characterized as noise or as a mix of signal and noise.
Specifically, this information can be obtained by examining the singular values of matrices
containing information either about the autocorrelation or the third-order cumulant (under
the rationale that noise can be modeled as a Gaussian process, and therefore has no non-
zero cumulants of order greater than two) [185]. Lastly, if a model of the expected signal is
available, then optimal filtering techniques such a Wiener filter or a matched filter can be
Literature Review 45
(a) (b)
(c) (d)
Figure 2.3: Types of intraneural electrodes: a) LIFE, b) MEA, c) SPINE, and d) sieve
electrode.
used to separate the signal from the noise [67].
2.3.3 Source localization and pathway discrimination using extra-
neural measurements
Simple nerve cuffs with a single channel reveal only one aspect of the nerve’s activity, and
provide little opportunity to go beyond traditional signal processing techniques such as the
ones described in the previous paragraph. As the electrodes become more sophisticated and
start providing several recording channels, more options become available for analyzing the
Literature Review 46
data obtained. Nerve cuffs are then no longer used simply as a binary tool to detect whether
or not a signal is present, but can be used to provide more complex information.
The first pathway discrimination method to be proposed was based on the conduc-
tion velocity of the nerve fibers. Because different types of fibers have different conduction
velocities [69], identifying the velocity provides one way of discriminating the activity of
various pathways. Before sophisticated nerve cuffs were available, this type of method was
explored using extraneural measurements at two different sites (algorithms explored included
correlation-based methods [57] and matched filters [3]). Developments in nerve cuff manu-
facturing technology led to the introduction of a velocity-selective system that uses several
measurements sites within a single cuff [173, 136, 135]. Using measurements at more than
two sites provides increased robustness.
Alternatively, methods have been proposed that attempt to discriminate pathways
not by using velocity information, but by taking advantage of the fact that an AP traveling
along a fiber at a given location in the nerve will produce different measurements at dif-
ferent contacts of a multi-contact cuff. Early attempts examined the recording differences
in traditional cuffs (with eight to twelve contacts) when action potentials traveled in vari-
ous fascicles or pathways [85, 169, 145, 17]. Developing this approach further, later studies
were able to quantify to what extent sources between and within fascicles produced different
measurement patterns, using FINEs with thirteen channels or less [125, 197]. However, iden-
tification of the active fascicle was not discussed in detail, beyond matching measurements
with known patterns. Extending this methodology to finer resolutions would soon become
infeasible, because the number of possible patterns would grow very large, all the more so
if any number of locations were allowed to be active at the same time. Other attempts to
separate the activity of different fascicles in a nerve based on extraneural recordings have
included the use of blind source separation (again with a FINE) [177, 175, 176] and linear
regression (using a four-channel round cuff) [14]. In both of those cases, the resolution was
limited to discriminating the activity of two fascicles. An earlier study attempted to localize
activity at an arbitrary location in the nerve with an eight-contact cuff, but used a very
Literature Review 47
simplified model of the nerve’s electrical properties and therefore obtained only coarse map-
pings [85]. Simple selectivity studies with multi-contact spiral cuffs have also been reported
that focused on discriminating activity related to specific innervated muscles, rather than to
specific fascicles [140, 141].
The method proposed in the rest of this thesis, based on an inverse problem of source
localization, in effect generalizes the ideas in the previous paragraph and aims to obtain a
more flexible framework, capable of differentiating activity not only between fascicles but also
within a single fascicle (albeit with limited resolution). This approach has previously been
suggested for the purposes of electrode targeting in the spinal cord [107]. That study focused
on localizing a single locus of activity in a two-dimensional cross-section of the spinal cord,
and found that both distributed linear and equivalent dipole methods were able to obtain
localization errors of less then 300µm. They determined that the process was reasonably
robust to errors in conductivity values and electrode position, but less so to high levels of
noise and to errors in the partitioning of the region into white matter and gray matter areas.
Preliminary research into the peripheral nerve source localization problem has also recently
started to emerge from other groups [28].
Chapter 3
Reduction of the Inverse Problem
Solution Space
3.1 Introduction
The aim of the first step in our study was to reduce the number of variables to be estimated
in the source localization problem by taking into account the inherent limitations in the
measurement setup. The guiding principle for the reduction was that if the properties of
the nerve, the electrode, and the discretization of the solution space are such that sources
at two adjacent locations will produce indistinguishable measurements, then the activity at
those two locations can be represented by a single variable. The main benefits of eliminating
unnecessary variables would be improvements in computation time and storage requirements,
as well as improved insight into what resolution limitations are intrinsic to a given nerve
anatomy and electrode geometry.
FEM is a more appropriate method than BEM for constructing a leadfield for the
peripheral nerve source localization problem, due to the anisotropic conductivity of the
region. The accuracy of the FE model both in terms of anatomy and tissue conductivities
will influence the quality of the solution when the leadfield is applied to measurements
48
Reduction of the Inverse Problem Solution Space 49
obtained in vivo. When a FE model is used to construct the leadfield, the fineness of the
mesh will have a significant impact on the accuracy of the matrix entries. Unfortunately,
each element in the mesh corresponds to a variable to be solved for in the inverse problem,
creating a conflict between the needs to make the leadfield as accurate as possible and to keep
the number of variables as small as possible. A simple way to reduce the number of variables
without coarsening the original FE mesh would be to decimate the solution space for the
purposes of the inverse problem. Here, we propose an alternative method to reduce the
number of variables, where the decisions are based on quantitative information contained in
the leadfield. Each column of the leadfield matrix describes the set of measurements obtained
when a unit source is placed at a given location (the forward field of that source). Therefore,
if the difference between two columns corresponding to adjacent locations is extremely small,
then the two corresponding sources can be said to be indistinguishable by the measurement
setup. This implies that two variables are being used when only one is useful. We investigated
how much reduction in the number of variables can be achieved by grouping together leadfield
columns that are indistinguishable, and what, if any, detrimental impact this transformation
may have on the process of solving the inverse problem.
The notion of varying the mesh coarseness in accordance with the capabilities of
the measurement setup has been previously proposed in the context of a different inverse
problem, electrical impedance tomography (EIT) [13, 66]. By using the differences between
leadfield columns as a quantitative grouping criterion, we are adapting this idea to a bio-
electric source localization problem, and developing a method by which the accuracy of the
forward problem does not need to be compromised.
3.2 Theory
Determining whether two leadfield columns correspond to indistinguishable elements is equiv-
alent to determining whether all the entries in the two vectors are indistinguishable down to
a given precision. The precision of the sensor and the amount of noise present will determine
Reduction of the Inverse Problem Solution Space 50
the minimum difference that must exist between two values for them to be reliably distin-
guishable. Calling this minimum distance δ, two vectors v1 and v2 are deemed distinguishable
if they meet the condition in Equation 3.1.
‖v1 − v2‖∞ > δ (3.1)
The remaining necessary step before being able to apply this criterion is to determine
δ. We can consider sensor precision and measurement noise within a single framework by
interpreting sensor error as noise. To relate δ to the noise level, two values will be considered
distinguishable if they are separated by at least twice the standard deviation of the noise. By
using the definition of noise-to-signal ratio (NSR) in Equation 3.2 (the standard deviation
of the elements of the noise vector over the standard deviation of the elements of the signal
vector), the condition stated above can be reformulated as shown in Equation 3.3.
NSR =std(noise)
std(signal)(3.2)
‖v1 − v2‖∞ > 2 ∗NSR ∗ std(signal) (3.3)
Using this criterion, groups of indistinguishable adjacent elements can be formed. The
leadfield columns corresponding to all the elements in a group are then replaced by a single
vector equal to the average of all the selected original columns, thereby reducing the total
number of columns in the leadfield.
It is typical for the leadfield in a bioelectric source localization problem to contain
three columns for each location to be solved for, corresponding to the three orthogonal
components of the dipole at that location. In that case, two elements can truly be said to
be indistinguishable only if all three pairs of leadfield columns prove indistinguishable. In
addition, in order to guarantee that the elements are indistinguishable for a unit source of
any orientation, the threshold in Equation 3.3 must be divided by√
3. To see this, consider
the following argument. If the elements are distinguishable for one of the three orthogonal
Reduction of the Inverse Problem Solution Space 51
orientations corresponding to the leadfield columns, then they are not grouped and no further
examination is necessary. Therefore, for the rest of this discussion, the following inequalities
are assumed to be true:
‖v11 − v21‖∞ ≤ δ (3.4)
‖v12 − v22‖∞ ≤ δ (3.5)
‖v13 − v23‖∞ ≤ δ (3.6)
where v11, v12, and v13 are the three leadfield columns corresponding to the first mesh element,
and v21, v22, and v23 are those corresponding to the second mesh element. If the orientation
of the source is described by the triplet (a,b,c), then the measurements produced when that
source is placed in the first of the two mesh elements will be given by a ∗ v11 + b ∗ v12 + c ∗ v13,
with an analogous expression for the second mesh element. The difference between the
measurements produced by that source in each of the two mesh elements will then satisfy
the following inequality, which follows from Equations 3.4 to 3.6 and the properties of the
infinity norm:
‖a(v11 − v21) + b(v12 − v22) + c(v13 − v23)‖∞ ≤ (a+ b+ c)δ (3.7)
Because the source under consideration has unit magnitude, the values (a,b,c) must
satisfy the constraint:
√a2 + b2 + c2 = 1 (3.8)
It is easily verified that the highest value of (a+ b+ c) attainable under constraint
3.8 is√
3. Combining this information with Equation 3.7, we can conclude that in order to
guarantee that the measurement difference will remain under a certain threshold ∆ regardless
of the orientation of the unit source, then the threshold used in Equations 3.4 to 3.6 should
be equal to δ = ∆/√
3.
Reduction of the Inverse Problem Solution Space 52
In peripheral nerves, however, the current sources generating the electric potentials
outside of the axon can be modeled as dipoles oriented along the axis of the nerve [92]. We
will therefore restrict the solution space to dipoles oriented in that direction only, since the
components in the other two directions should be significantly smaller. This choice reduces
the size of the leadfield from Nx3M to NxM. Furthermore, the grouping of elements is based
on the comparison between a single pair of vectors, instead of three pairs, and using the
threshold in Equation 3.3 without adjustment.
It should be noted that the criterion outlined here assumes that there is no significant
difference in the noise level at the various electrode contacts. This is a simplification, because
potentials originating from sources outside the cuff vary linearly with longitudinal position
along the cuff [2], such that noise levels will be different for contacts located at different points
along the cuff. If sufficient information about the noise distribution at the various contacts
was available, it could be incorporated into the distinguishability criterion by making the
threshold different for each entry in the vector difference.
3.3 Methods
3.3.1 Finite element model and leadfield construction
An FE model of a unifascicular section of the rat sciatic nerve was constructed for this portion
of the study. The model consisted of a cylindrical nerve surrounded by a cuff electrode. The
nerve consisted of a single fascicle and was modeled as three concentric cylinders representing
the endoneurium, perineurium and epineurium layers. The nerve and the cuff were separated
by an encapsulation tissue layer and a saline layer. The whole structure was placed in a saline
bath. The dimensions and conductivities of the various part of the nerve model are given
in Table 3.1 and based on similar models and anatomical studies described in the literature
(relevant references for each parameter are given in the table).
From this model, five meshes of varying coarseness were generated. The solution space
Reduction of the Inverse Problem Solution Space 53
Table 3.1: Parameters for the finite element model of the rat sciatic nerve.
Parameter Values References
Nerve length 5 cm
Endoneurium radius 415 µm [151]
Perineurium width 35 µm [36, 119, 125]
Epineurium width 35 µm [36, 119]
Encapsulation tissue layer width 7.5 µm
Saline layer width 7.5 µm
Cuff length 2.3 cm [157]
Cuff width 30 µm [157]
Cuff radius 500 µm [157]
Cuff starting height 1.35 cm
Saline bath length 5 cm
Saline bath radius 0.485 cm [125, 197]
Endoneurium conductivity 8.26 × 10−2 S/m [36, 98, 197]
(radial)
Endoneurium conductivity 0.571 S/m [36, 98, 15, 197]
(longitudinal)
Perineurium conductivity 2.1 × 10−3 S/m [36, 15, 197]
(all directions)
Epineurium conductivity 8.26 × 10−2 S/m [15, 197]
(all directions)
Encapsulation tissue conductivity 6.59 × 10−2 S/m [119]
(all directions)
Saline conductivity (all directions) 2 S/m [36, 119, 15, 197]
Cuff conductivity (all directions) 1 × 10−7 S/m [197]
Reduction of the Inverse Problem Solution Space 54
for the source localization problem was restricted to the endoneurium. The number of mesh
elements for the endoneurium region only was 75600, 38400, 19200, 12000, and 8400 for the
five meshes, respectively. The other variable of interest is the number of contacts in the cuff
electrode. Four values were considered: 104, 56, 24, and 12. The 56 contacts were chosen
from the original 104, the 24 from the 56, and the 12 from the 24, such that each set was a
subset of the previous case (specifically, the contact configurations consisted of 13 rings of 8
contacts, 7 rings of 8 contacts, 3 rings of 8 contacts, and 3 rings of 4 contacts, with the rings
in all cases being positioned symmetrically with respect to the middle of the cuff). The 56-
contact electrode was modeled on the MCC used in the rest of this thesis [157], and the other
cases are merely extensions or subsets of this pattern of contacts. The electrode dimensions
in the model are also taken from this design. The 1 mm diameter of this cuff is very close to
the approximate diameter of the rat sciatic nerve, hence the tight fit of the cuff in the model.
In total, the five coarseness levels and 4 electrode patterns resulted in 20 leadfield matrices.
The finite element analysis was conducted using the SCIRun computing environment (an
open source platform for the modeling, simulation and visualization of scientific problems
[158]), and the rest of the leadfield computations were performed using the Matlab software
(The Mathworks Inc., Natick, MA). The procedure for obtaining the leadfield from the finite
element model is described by Weinstein et al. [193].
3.3.2 Element grouping algorithm
The algorithm described in the Theory section is designed to reduce a leadfield matrix based
solely on the entries of that matrix, rather than using information contained in a specific set
of measurements. If we want to incorporate information about the noise level into the choice
of threshold, however, the standard deviation of the signal must be known (see Equation 3.3).
In order to circumvent the problem, we base the threshold calculations on a collection of
sample signals obtained from the leadfield columns themselves. Each column corresponds to
the measurements obtained when a unit source is placed at a given location, and therefore
Reduction of the Inverse Problem Solution Space 55
can be seen as a simple sample signal. We choose a subset of all these possible sample
signals. For each of the column vectors in this set, the standard deviation of the entries of
the vector is computed. The average of these standard deviation values is then used as the
signal standard deviation in Equation 3.3. The columns included in the subset are those
corresponding to the mesh elements lying in the endoneurium between the heights of 2 and
3 cm along the 5 cm nerve model, under the assumption that the signals originating close
to the midpoint of the cuff are most representative of the signals that we are interested in
localizing.
Having now obtained all the information necessary for determining whether two el-
ements are distinguishable, we can proceed to use this criterion to form groups of indistin-
guishable elements. Although theoretically groups could consist of non-adjacent elements
spread throughout the solution space, the leadfield reduction process could then no longer be
interpreted as a search for the inherent resolution achievable by the measurement setup. In
this study, the additional restriction is therefore imposed that groups should be formed from
connected mesh elements. This has the additional advantage of considerably reducing the
time required to test every eligible pair of elements for distinguishability. To further simplify
the grouping process, we use the fact that the mesh was constructed using an extrusion pro-
cess that results in distinct layers of elements, and only compare elements to others within
the same layer. Within each layer of the mesh, the algorithm used is as follows. First, a
chain of connected elements is formed, going along the outer edge of the layer and spiraling
inwards until the center is reached. If a point is reached where all the elements connected to
the current element are already part of the chain, then the chain continues at the next free
element. Once all the elements have been added, comparisons are performed on successive
elements of the chain. For instance, the first and second elements are compared, and if they
are indistinguishable they are grouped together. If the third element is also indistinguish-
able from both of the first elements, it is added to the group. This process is continued
until an element is found that cannot be added to the group. A new group is then begun,
and so on until all the end of the chain is reached. A sample grouping is shown in Figure
Reduction of the Inverse Problem Solution Space 56
Figure 3.1: Example of the results obtained when the grouping algorithm is applied to
a layer of the FE model at z = 24.94mm, under the assumption of 0.1% NSR. Elements
surrounded by thick lines belong to the same group.
3.1, corresponding to a layer of the FE model midway up the cuff electrode (this particular
grouping is based on the assumption of a 0.1% NSR).
The algorithm just described clearly does not capture all possible groupings. The
chaining process essentially restricts the comparisons to a one-dimensional geometry instead
of a two-dimensional one, and some elements that are indistinguishable will therefore not be
grouped. Additionally, restricting the comparisons to elements within the same layer will also
omit some possible groupings. The spiral shape is dependent on the geometry of the mesh,
which is irregular and so can cause some distortion and influence which elements are grouped
(this effect is visible in Figure 3.1 toward the centre of the cross-section). Unfortunately,
developing an algorithm able to maximize the number of three-dimensional groupings is far
from a trivial task, and is not the main concern of this study. The chain algorithm used
is capable of significantly reducing the size of a leadfield, as will be shown in the Results
section, and therefore is deemed sufficient for the purposes of this study.
Reduction of the Inverse Problem Solution Space 57
The following NSR values were used to choose a grouping criterion: 0.1%, 5%, 10%,
15%, and 20%. Having an error equal to 0% is unrealistic, if only because of numerical
precision issues. For the purpose of this study, a value of 0.1% was therefore used instead of
0% to capture very small errors, such that elements that are extremely similar will be still
be grouped together.
3.3.3 Leadfield comparison metrics
The success of the leadfield reduction process will be evaluated primarily in terms of the
number of columns of the reduced leadfield, compared to the original leadfield. Nonetheless,
the advantages derived from a smaller matrix may not be worthwhile if the new matrix has
a detrimental impact on the quality of the inverse problem solution. For this reason, it is
important to have a set of metrics that reflect the difficulty of solving the ill-posed problem
using a given matrix.
Rank and condition number
The rank of the leadfield matrix should be equal to the number of electrode contacts. Oth-
erwise, some of the measurements would be linearly dependent on others and not providing
any additional information. It is therefore important to check that a reduced leadfield still
has full row rank, in order to confirm that the reduction process has not led to a loss of
information. The other matrix property that will be examined is the condition number,
which is a measure of the sensitivity of the solution of a linear system of equations to errors
in the right hand side [34, 54]. The condition number can be computed as the ratio of the
largest and smallest singular values of the matrix. Matrices with relatively small condition
numbers are said to be well-conditioned, otherwise they are ill-conditioned. The leadfield
matrix in the source localization problem is typically ill-conditioned, and a variety of regu-
larization techniques exist to deal with this problem (usually by finding ways to minimize
the destabilizing effect of very small singular values, for example by adding a constant to all
Reduction of the Inverse Problem Solution Space 58
singular values or by truncating the smallest ones) [54]. Our concern here is therefore not
the absolute value of the condition number, but we need to check that the condition num-
ber of the reduced leadfield is not significantly greater than that of the original leadfield.
Otherwise, the reduction process may have made the problem more difficult to solve. The
metric used is therefore the ratio of the condition number of the reduced matrix to that of
the original matrix.
Geometry noise
The two main sources of error in the source localization problem are measurement noise
and geometry noise. Geometry noise is the error introduced by inaccuracies in the leadfield
matrix, which occur because the models used to solve the forward problem are not com-
pletely accurate representations of the true anatomy. During the leadfield reduction process,
the columns corresponding to elements deemed indistinguishable are averaged together. De-
pending on how rigorous a threshold is used, this averaging process could introduce some
additional geometry noise. For each column in the original leadfield, a geometry noise vector
is defined as the difference between the original leadfield column and the corresponding col-
umn in the reduced leadfield. By analogy with the measurement noise, the geometry noise
level is then defined by using this geometry noise vector and the original leadfield column
in Equation 3.2. The average of this value for all the columns in the original leadfield is
then used as a metric quantifying the amount of additional geometry noise introduced by
the reduction process. Defining Met1 to be this metric, signali to be the ith column of the
original leadfield, and gei to be the geometry error vector corresponding to that column, the
metric is computed as shown in Equation 3.9 (where M is once again the number of leadfield
columns).
Met1 =
(
M∑
i=1
std(gei)
std(signali)
)
/M (3.9)
Reduction of the Inverse Problem Solution Space 59
3.3.4 Example and Complexity Analysis
Although the rank, condition number and geometry noise metrics have the advantage of not
being dependent on a specific source configuration, it is nonetheless beneficial to consider spe-
cific examples. These examples can provide some confirmation that the leadfield reduction
process does not significantly increase the localization error, as well as illustrate the reduc-
tions in computation time that can be achieved. We focus here on two cases: the 56-contact,
75600-element model, with 0.1% noise and 10% noise. The reduced leadfields constructed for
NSR = 0.1% and NSR = 10% are used for the first and second of those cases, respectively,
and the localizations are compared with those obtained using the original leadfield.
For each case, we generate simulated measurements corresponding to an action po-
tential traveling along a single myelinated fiber. A myelinated mammalian nerve fiber action
potential is first simulated using the model described by Sweeney et al. [172]. In order to re-
main consistent with the EEG/MEG source localization literature, equivalent current dipoles
are used to approximate the electrical activity of the nerve fibers. The magnitude waveform
of the current dipole is therefore obtained from the first derivative of the transmembrane
potential during the action potential [129]. The waveform is then propagated from one node
of Ranvier to the next at a speed of approximately 50 m/s [172]. The nodes of Ranvier are
placed 1 mm apart, which is consistent with a 10 µm-diameter fiber. For the purposes of
this example we restrict the localization to a single time instant, when the action potential’s
peak is in the nodes of Ranvier near the middle of the cuff. Once the current dipole locations
and magnitudes have been obtained in this way, the measurements are computed using the
original leadfield, and the appropriate amount of noise is added.
The source localization is performed using the sLORETA algorithm [123]. For each
of the two example cases, 100 trials are conducted, with the position of the active fiber in
the nerve cross-section generated randomly at every trial. To evaluate the results, the three-
dimensional solution is projected onto a two-dimensional cross-section, because our primary
interest is to determine which pathway is active. From the two-dimensional projection,
Reduction of the Inverse Problem Solution Space 60
localization error is obtained as the distance between the true pathway location and the
location of the peak of the estimate. The error is averaged over 100 trials for each of the two
example cases. Note that in order to compare the results, the reduced estimate is mapped
back to the full solution space by assigning to each variable in the original space the value
of the corresponding variable in the reduced space.
The computation time gains that can be achieved using the leadfield reduction de-
pends on the algorithm used to solve the inverse problem. In the case of sLORETA, recall
from Section 2.2.3 that the first step of the algorithm is to solve a minimum-norm least-
squares problem, shown in Equation 3.10, with the closed-form solution in 3.11.
j = arg minj‖(Lj − d)‖2 + λ‖j‖2 (3.10)
j = Td = LT[LLT + λI]−1d (3.11)
Here, the value of λ is chosen using the cross-validation functional [121]. The second
step of sLORETA is to normalize the solution from 3.11 using the resolution matrix, as
shown in Equation 3.12.
jT
l (Rll)−1jl (3.12)
In the case of constrained orientations, where there is a single variable per dipole
location, jl is simply the magnitude of the lth dipole, and Rll is the lth diagonal entry of the
resolution matrix.
Recall that L is of size NxM, with M much bigger than N. As a result, examination
of Equation 3.11 reveals it to be of complexity O(MN2). Each diagonal entry of R can be
computed in O(N), such that the total normalization process can be performed in O(NM).
The choice of regularization parameter relies on an eigenvalue decomposition that is limited
by the rank N of the leadfield [121], and therefore independent of M. We expect to see these
Reduction of the Inverse Problem Solution Space 61
complexities reflected in the computational times of the examples when the original and
reduced values of M are used.
3.4 Results
3.4.1 Size of the reduced leadfield
The number of columns in the original and reduced leadfield are compared in Figure 3.2, for
the five meshes, four electrode contact sets, and five NSR values. In all cases, the original
leadfield could be reduced to half its original size or less. As more measurement uncertainty
was incorporated into the choice of threshold, the size of the reduced leadfield decreased.
This result is in accordance with expectations, since it implies that noise deteriorates the
achievable resolution.
When the grouping thresholds were based on the assumption of noiseless measure-
ments (NSR = 0.1% in Figure 3.2), the number of electrode contacts had very little influence
on the number of columns in the reduced leadfield, suggesting that in ideal conditions the
geometries and conductivities of the nerve and cuff may be the key factors influencing the
reduction achievable, rather than the number of electrode contacts used. On the other hand,
when measurement uncertainty was incorporated into the choice of threshold, the number
of electrode contacts started to be more important. Figure 3.2 shows that in those cases
the reduced leadfields had progressively more columns as the number of electrode contacts
increased, meaning that the increased number of measurements slightly improved the res-
olution in the presence of noise. The sizes of the reduced leadfields corresponding to the
different contact configurations did not, however, vary linearly with the number of contacts.
This suggests that an optimal number of contacts could be found, representing a good trade-
off between achieving better resolution and increasing the number of contacts necessary. In
the present study, the difference between 12 and 24 contacts was very small, implying that
very little is to be gained from using the larger of those two sets. On the other hand, the
Reduction of the Inverse Problem Solution Space 62
0 2 4 6 8
x 104
0
1
2
3
4x 10
4 NSR = 0.1%
Original size
Red
uced
siz
e
104 contacts56 contacts24 contacts12 contacts
0 2 4 6 8
x 104
0
1
2
3
4x 10
4 NSR = 5%
Original size
Red
uced
siz
e
0 2 4 6 8
x 104
0
1
2
3
4x 10
4 NSR = 10%
Original size
Red
uced
siz
e
0 2 4 6 8
x 104
0
1
2
3
4x 10
4 NSR = 15%
Original size
Red
uced
siz
e
0 2 4 6 8
x 104
0
1
2
3
4x 10
4 NSR = 20%
Original size
Red
uced
siz
e
Figure 3.2: Number of columns of the reduced leadfield as a function of the number of
columns of the original leadfield and of the number of electrode contacts, when the grouping
criterion is based on NSR = 0.1%, 5%, 10%, 15%, or 20%. The curves for 104 and 56
electrode contacts nearly overlap, as do the ones for 24 and 12 electrode contacts.
Reduction of the Inverse Problem Solution Space 63
difference between 24 and 56 contacts was appreciable. The difference between 56 and 104
contacts was also noticeable but somewhat smaller, which suggests that out of the four con-
figurations examined, the 56-contact configuration is the best choice for this problem. As
the noise level continued to increase and the size of the all the reduced leadfields decreased,
however, the size differences due to the number of contacts became less significant (compare
for example the NSR = 5% and NSR = 20% cases in Figure 3.2). In other words, there
was a limit to the amount of uncertainty that the number of contacts could compensate for.
Lastly, it is important to note that all of these observations are valid for all five of the meshes
investigated.
3.4.2 Properties of reduced leadfield
The reduced matrices had full row rank for every combination of mesh, number of contacts,
and noise level, meaning that the reduction process never led to any measurements becoming
linearly dependent on others.
The ratio of condition numbers between reduced and original leadfields was in all
cases smaller than 1, indicating that none of the leadfields became more ill-conditioned than
they already were as a result of the reduction process. The mean of these ratios for all the
reduced leadfields constructed was 0.905 ± 0.043, although no significant relationship to the
mesh coarseness, number or contacts, or noise level was observed.
Table 3.2 shows the values of the Met1 metric for all of the reduced leadfields. As
expected, the amount of geometry error increases as the criterion for indistinguishability
is relaxed (a clear example is provided by the row in bold). For all cases, the geometry
error remained relatively small, with Met1 remaining under 2.5% when the least stringent
grouping criterion is used, and under 1% when the most stringent criterion is used. These
values suggest that the increased geometry error is a small enough price to pay for the
significant reduction that was achieved in the number of variables.
Reduction of the Inverse Problem Solution Space 64
Table 3.2: Metric Met1 for all test cases.
NSR = 0.1% NSR = 5% NSR = 10% NSR = 15% NSR = 20%
104 contacts
75600 elements 0.6 +/- 1.4% 1.1 +/- 1.7% 1.3 +/- 1.7% 1.6 +/- 1.8% 1.8 +/- 1.9%
38400 elements 0.6 +/- 1.5% 1.1 +/- 1.9% 1.4 +/- 1.9% 1.6 +/- 1.9% 1.9 +/- 2.0%
19200 elements 0.7 +/- 1.8% 1.3 +/- 2.3% 1.6 +/- 2.3% 2.0 +/- 2.3% 2.1 +/- 2.3%
12000 elements 0.7 +/- 2.3% 1.3 +/- 2.7% 1.7 +/- 2.7% 1.9 +/- 2.7% 2.0 +/- 2.7%
8400 elements 0.4 +/- 1.6% 0.7 +/- 2.1% 0.9 +/- 2.2% 1.2 +/- 2.2% 1.3 +/- 2.3%
56 contacts
75600 elements 0.6 +/- 1.4% 1.1 +/- 1.7% 1.4 +/- 1.7% 1.7 +/- 1.8% 2.0 +/- 2.0%
38400 elements 0.6 +/- 1.5% 1.1 +/- 1.9% 1.4 +/- 1.9% 1.7 +/- 2.0% 2.0 +/- 2.1%
19200 elements 0.7 +/- 1.8% 1.3 +/- 2.3% 1.7 +/- 2.3% 2.0 +/- 2.3% 2.2 +/- 2.4%
12000 elements 0.7 +/- 2.3% 1.4 +/- 2.7% 1.7 +/- 2.7% 1.9 +/- 2.7% 2.1 +/- 2.7%
8400 elements 0.4 +/- 1.6% 0.7 +/- 2.1% 1.0 +/- 2.2% 1.2 +/- 2.2% 1.4 +/- 2.3%
24 contacts
75600 elements 0.6 +/- 1.4% 1.3 +/- 1.7% 1.6 +/- 1.8% 1.9 +/- 2.1% 2.1 +/- 2.5%
38400 elements 0.6 +/- 1.6% 1.3 +/- 1.8% 1.6 +/- 1.9% 1.9 +/- 2.2% 2.1 +/- 2.5%
19200 elements 0.7 +/- 1.8% 1.5 +/- 2.2% 2.0 +/- 2.3% 2.2 +/- 2.5% 2.4 +/- 2.7%
12000 elements 0.8 +/- 2.4% 1.6 +/- 2.7% 1.9 +/- 2.7% 2.1 +/- 2.7% 2.2 +/- 2.8%
8400 elements 0.4 +/- 1.6% 0.9 +/- 2.1% 1.1 +/- 2.2% 1.3 +/- 2.4% 1.4 +/- 2.6%
12 contacts
75600 elements 0.6 +/- 1.4% 1.3 +/- 1.7% 1.7 +/- 1.9% 2.0 +/- 2.2% 2.2 +/- 2.6%
38400 elements 0.6 +/- 1.6% 1.4 +/- 1.9% 1.7 +/- 2.0% 2.0 +/- 2.3% 2.2 +/- 2.7%
19200 elements 0.7 +/- 1.8% 1.6 +/- 2.2% 2.1 +/- 2.4% 2.3 +/- 2.6% 2.4 +/- 2.7%
12000 elements 0.8 +/- 2.4% 1.7 +/- 2.7% 2.0 +/- 2.7% 2.1 +/- 2.8% 2.3 +/- 3.0%
8400 elements 0.4 +/- 1.6% 0.9 +/- 2.1% 1.2 +/- 2.2% 1.4 +/- 2.4% 1.5 +/- 2.6%
Reduction of the Inverse Problem Solution Space 65
Figure 3.3: 2D projections of the true source distribution, the estimate obtained with the
original leadfield, and the estimate obtained with the reduced leadfield, for one trial in the
NSR = 0.1% example.
3.4.3 Simulation results
Figure 3.3 shows the 2D-projections of the results of one trial in the 0.1% noise case, using
both the original and reduced leadfields. It is clear from the figure that the localization
performance for this trial was virtually identical regardless of whether the original or the
reduced leadfield was used. Table 3.3 displays the average over 100 trials of the localization
error for both examples, with the original and reduced leadfields in each case. The results
show that the mesh reduction process resulted in a negligible localization error increase in the
first example and a small decrease in the second example. These results support our claim
that the reduction process does not significantly reduce the quality of the inverse problem
solution.
Table 3.4 displays the total computation times for the two examples, as well as the
breakdown into the different components of the algorithm. The table also shows the ratios
of these different values for the simulations conducted with the original and reduced lead-
fields. The computation times were obtained using Matlab running on a desktop PC with a
3.0 GHz Pentium IV processor. As expected, the reduction had no effect on the speed of the
regularization process, and the ratio of computation times for the normalization step was
roughly equal to the ratio of the number variables in the two leadfields. The ratio for the
Reduction of the Inverse Problem Solution Space 66
Table 3.3: Simulation results using the original and reduced leadfields. All simulations were
conducted with 56 contacts, and all means correspond to 100 trials.
NSR = 0.1% NSR = 10%
Size of the reduced leadfield 30252 21472
(original leadfield size is 75600)
Mean localization error using 0.0801 +/- 0.0505 0.1359 +/- 0.1108
the original leadfield (mm)
Mean localization error using 0.0807 +/- 0.0505 0.1333 +/- 0.0940
the reduced leadfield (mm)
solution of the minimum-norm least-squares problem was slightly larger than expected, but
a substantial reduction was nonetheless achieved. The discrepancy between the expected
and observed reduction may be due to the fact that our analysis was based on a “naıve”
approach to matrix multiplications and did not take into account any of the optimizations
that may be present in the Matlab software.
3.5 Discussion
By using the similarity of leadfield columns as a criteria for grouping several mesh ele-
ments into a single variable, we were able to achieve substantial reductions in the number
of variables that the inverse problem aims to estimate. This solution space reduction was
investigated in the context of the peripheral nerve source localization problem, although it
could also be applied to other bioelectric source localization tasks. The reduction is valu-
able for several reasons. First, the inverse problem is made less ambiguous by the clearer
distinctions between the leadfield columns. The smaller discrepancy between the number of
measurements and the number of variables also means that the problem will be somewhat
better conditioned. Most importantly, the smaller leadfield matrix will translate into faster
computations. This can be particularly advantageous when iterative algorithms are used
Reduction of the Inverse Problem Solution Space 67
Table 3.4: Computation time comparison for simulations using the original and reduced
leadfields (all values are means over 100 trials.)
NSR = 0.1% (Mreduced/Moriginal = 0.4)
Computation step Computation time Computation time Ratio
(original) (reduced) (reduced/original)
Regularization 0.3136 s 0.3046 s 0.9918
Minimum Norm Least-Squares 0.3926 s 0.2467 s 0.4865
Normalization 0.3532 s 0.1397 s 0.3959
Total 1.0594 s 0.6910 s 0.6123
NSR = 10% (Mreduced/Moriginal = 0.284)
Computation step Computation time Computation time Ratio
(original) (reduced) (reduced/original)
Regularization 0.2611 s 0.2615 s 1.0022
Minimum Norm Least-Squares 0.3290 s 0.1176 s 0.3576
Normalization 0.3427 s 0.0953 s 0.2781
Total 0.9328 s 0.4743 s 0.5087
(e.g. FOCUSS [37]), and would become crucial if real-time implementation were attempted
as part of a control system for a neural prosthesis. The reduction in required storage space
would also be valuable for implanted systems. The reduction process did not have a negative
impact on any of the metrics used to assess the difficulty of the inverse problem.
Despite these advantages, however, it should be kept in mind that the smaller number
of variables is a mixed blessing, since it implies lower resolution. The purpose of the technique
presented here is not to make the number of variables arbitrarily small, but rather to try to
approach the number that best represents the inherent resolution that is achievable for the
given nerve properties and measurement setup, without having superfluous variables that
simply make the problem more difficult. In that sense, the proposed technique provides
Reduction of the Inverse Problem Solution Space 68
us with information that could not be obtained using a simple decimation approach, and
is therefore preferable even if both methods provide the same amount of solution space
reduction.
The incorporation of information about the noise level in the criterion to determine the
distinguishability of two mesh elements was also explored. Not surprisingly, increasing the
uncertainty increased the number of element groupings, because pairs that were only barely
distinguishable using perfect measurements could no longer be reliably separated. This is
simply another way of saying that uncertainty in the measurements will negatively impact the
achievable resolution. Additionally, it was found that a large number of electrode contacts
had more impact on grouping decisions when uncertainty was present. Indeed, having more
contacts would provide better spatial sampling of the electric fields, and therefore provide
more opportunities to detect differences between the fields generated by sources in adjacent
locations when very small differences are obscured by noise. This finding argues in favor of
using a large number of contacts in practice, where noise cannot be completely avoided. On
the other hand, it was found that increasing the number of contacts past a certain number
started yielding diminishing returns, suggesting that an optimal number could be found. The
impact of the number of contacts on the selectivity of cuff electrodes has been previously
studied by Yoo and Durand [197], who also noted the existence of a plateau. The optimal
number in that study proved to be 7, however there was a single ring of contacts. The
higher number found here suggests that it is possible to take better advantage of having
more contacts by distributing them in several rings along the length of the cuff, rather than
a single ring. It is important to clarify that our conclusion about the existence of an optimal
number of contacts applies only to improvements in resolution, and says nothing about
the impact that a larger number of measurements will have on the accuracy of the inverse
problem solution.
Note also that in practice, precise information on the noise level may in many cases
be unavailable, in which case grouping decisions will have to be based on the NSR = 0.1%
criterion, or on a conservative noise estimate. Otherwise, some resolution might be needlessly
Reduction of the Inverse Problem Solution Space 69
lost.
It should be kept in mind when analyzing the results of this study that the specific
size reductions achieved are dependent on the grouping algorithm, which is not optimized.
Small changes in the algorithm will affect both the groupings that are formed among indis-
tinguishable elements and the overall number of groupings. It would therefore be worthwhile
to continue improving the grouping algorithm in order to maximize the achievable reduction.
Nonetheless, the present study demonstrated that even with a simple algorithm a substan-
tial decrease in the number of variables could be obtained, highlighting the usefulness of this
solution space reduction technique.
Another factor that was fixed in this study was the geometry of the nerve. Although
the geometry was simplified here to a single cylindrical fascicle, the solution space reduction
technique is in no way dependent on this geometry. In can be applied to any mesh geom-
etry for which a leadfield has been computed. Certain situations, such as a mesh that is
not comprised of well-defined layers or whose cross-section is not roughly compatible with a
spiral shape, may require modifications of the grouping algorithm, but the basic reduction
strategy would remain valid. Changing the geometry would of course alter the exact amount
of reduction achieved, but since the technique proposed here is simply exploiting the limited
resolution inherent in cuff electrode measurements, substantial reductions should be achiev-
able regardless of the details of the geometry. Likewise, the exact impact of the geometry on
the relationship between the number of contacts and the amount of reduction has not yet
been established in the general case, but the underlying insight that having more contacts
can compensate up to a point for the loss of information due to noise is independent of
geometry.
Chapter 4
Simulation of Bioelectric Source
Localization in the Rat Sciatic Nerve
4.1 Introduction
Before attempting to apply bioelectric source localization to experimental peripheral nerve
recordings, we conducted a simulation study of the process. The goal of the simulations was
to obtain an estimate of the expected performance under different conditions. Specifically,
we are interested in how effectively the performance can be improved by constraining the
problem in physiologically meaningful ways, and what amount of performance degradation
can be expected as the noise level increases.
4.2 Methods
4.2.1 Approach to solving the source localization problem
In order to obtain an estimate of the source distribution from the cuff electrode measure-
ments, there are two problems that need to be addressed: the forward and the inverse
problem. As described earlier, the goal of the forward problem is to compute the measured70
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 71
potentials that would result at each contact from a source at a given location. This informa-
tion can be represented in matrix form, and is known as the leadfield matrix. The leadfield
is needed to solve the inverse problem, whose goal is to estimate the source distribution from
the measurements.
Due to the ill-posed and underdetermined nature of the inverse problem of bioelectric
source localization, in order to arrive at a unique solution it is necessary to impose con-
straints on the solution [104]. These constraints should be determined by the anatomical
and physiological features of the problem at hand. In our simulations, we investigated the
performance of the source localization approach when varying levels of information about the
nerve were incorporated into the problem. An overview of the process is given in this section,
and details about each aspect of the simulations are presented in the following sections.
1. Create two numerical models of the nerve: a) one with an idealized geometry and b)
the other with a more realistic geometry. A leadfield was generated from each model,
and the more realistic model was used to generate simulated measurements, which were
used in all simulations.
2. Develop a spatio-temporal constraint based on the electrophysiological behaviour of
myelinated axons.
3. Evaluate the localization performance on four cases, which were: (1) using the idealized
leadfield and no spatio-temporal constraint (case IL), (2) using the idealized leadfield
and the spatio-temporal constraint (IL-C), (3) using the correct leadfield (i.e. the one
obtained from the more realistic geometry and used to generate the measurements) and
no spatio-temporal constraint (CL), and (4) using the correct leadfield and the spatio-
temporal constraint (CL-C). The first of theses cases incorporates the least information
about the problem into the inverse problem solution, whereas the last case incorporates
the most information.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 72
The forward problem
The forward problem is well-defined and can be solved analytically for simple geometries
and conductivity distributions. However, the anisotropic conductivity of nerves and the
potentially irregular shape of the fascicles mean that we must resort to numerical techniques.
Specifically, the forward problem is solved using FEM. In EEG/MEG source localization, the
method of choice is often BEM (e.g. [109, 163, 194]), but FEM is better suited to anisotropic
conductivities [193, 160] and for that reason is used here.
The first FEM model was built based on the idealized geometry of a unifascicular
section of the rat sciatic nerve. This extended unifascicular geometry is a simplification
compared to the real anatomy, which would branch progressively into several fascicles. The
main components of the model were a cylindrical nerve surrounded by a cuff electrode and
placed in a saline bath. The nerve was modeled as three concentric cylinders representing
the endoneurium, perineurium and epineurium layers. The nerve and the cuff were separated
by an encapsulation tissue layer and a saline layer (left panel of Figure 4.1). The dimensions
and conductivities of the various part of the nerve model are the same as those given in
Table 3.1, with the exception of the parameters that are shown in Table 4.1 along with their
new values. 56 electrode contacts were placed on the inside of the nerve cuff, organized in 7
rings of 8 electrodes each. The dimensions of the cuff electrode and the layout of its contacts
are based on the “matrix” MCC used throughout this project [157]. The ratio of the bath
and nerve diameters is large enough to avoid boundary effects [197], and the nerve segment
is long enough that dipoles placed at its ends have a negligible impact on the measurements.
The number of mesh elements in the model’s endoneurium was 56,400.
The second FEM model was based on a trace of a cross-section of a rat sciatic nerve,
at the level where the nerve begins to divide into its tibial and peroneal branches. The image
that formed the basis of the trace was obtained from the literature (Figure 1C in [151]). The
dimensions of the cuff and the bath were the same as in the previous model, as were the
conductivities. In order to fit the nerve into the cuff, it was scaled to 90% of its original size.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 73
Figure 4.1: Cross-sections of the idealized (left) and more realistic (right) FEM geometries
used to construct the leadfields.
This is a considerably simpler process than morphing the mesh to conform to the shape of
the cuff, and was deemed acceptable for the purposes of this study. It is important to clarify
that this geometry is still simplified, in that it is uniform in the longitudinal direction, and
as such does not represent the progressive branching of the nerve. The cross-section that
was selected as the basis for the trace was chosen because it corresponds approximately to
where the half-point of the cuff would be located on the sciatic nerve, assuming the cuff was
implanted just proximal to the division of the nerve into its main branches. It has been shown
that the shape of the nerve will conform to that of the cuff during chronic implantation [139].
Therefore, the geometry used here corresponds to a situation in which the cuff has only been
implanted for a short time, which is interesting for the purposes of this study, because the
difference between the “realistic” geometry and the idealized one will be greatest at that
time. The cross-sections of both FEM models are shown in Figure 4.1. The number of mesh
elements in the model’s endoneurium was 218,400 (the more complex geometry required a
finer mesh).
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 74
Table 4.1: Parameters for the idealized finite element model of the rat sciatic nerve.
Parameter Values References
Endoneurium radius 360 µm [151]
Epineurium width 35 µm [36, 119]
Encapsulation tissue layer width 40 µm [125]
Saline layer width 40 µm [125, 119]
Saline bath radius 0.48 cm [125, 197]
The finite element analysis was conducted using the SCIRun computing environment
[158], and the rest of the leadfield computations were performed using the Matlab software.
The procedure for obtaining the leadfield from the finite element model is described by
Weinstein et al. [193].
Simulated measurements
To generate simulated measurements, a myelinated mammalian nerve fiber action potential
was first simulated using the model described by Sweeney et al. [172]. In order to remain
consistent with the EEG/MEG source localization literature, equivalent current dipoles were
used to model the electrical activity of the nerve fibers. The magnitude waveform of the
current dipole was therefore obtained from the first derivative of the transmembrane potential
during the action potential [129]. The waveform was then propagated from one node of
Ranvier to the next at a speed of approximately 50 m/s [172]. The nodes of Ranvier were
placed 1 mm apart, which is consistent with a 10 µm-diameter fiber. The length of the
simulation was 2 ms.
Once the locations and time courses of the current sources had been determined,
the simulated measurements at the electrode contacts were obtained using the second FEM
model described in the previous section. The reference for the measurements was the average
of the two reference contacts present in the “matrix” cuff design (see [157]). Noise was then
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 75
added to each set of measurements, with the noise standard deviation set to 0%, 10%,
20%, 30%, or 40% of the signal standard deviation. The signal standard deviation for the
purposes of generating the noise was estimated by computing the standard deviations of the
measurements at each of the 8 contacts in the middle ring of the cuff, then averaging those
values. The resulting noise standard deviation was used to generate Gaussian white noise
time series for each of the 56 contacts. This noise represents the remaining contamination
after appropriate filtering and noise reduction measures have been used.
Measurements corresponding to two situations were simulated: a single active fiber,
and three active fibers. Each of the two cases was repeated 100 times, with the positions
of the active fibers within the endoneurium generated randomly every time. For the case of
three active fibers, the waveforms for all three fibers were identical but a random time shift
was applied. The maximum allowable time shift was a quarter of the length of the simulation.
Given the distance between the nodes of Ranvier and the length of the nerve model, each
fiber was composed of 50 dipoles, each with its own time course. The source localization
task described in the following sections was therefore dealing with regions featuring either
50 or 150 dipoles with varying magnitudes.
The inverse problem
There are two broad categories of EEG/MEG source localization methods: equivalent dipole
methods and distributed linear methods. Equivalent dipole methods assume that the po-
tential measurements can be explained using a small number of equivalent dipoles, whose
number is set a priori or estimated using the data and whose location, orientation, and
magnitude must be determined using search algorithms to fit the measurements (see Section
2.2.2). In the peripheral nerve problem, the combination of multiple active fibers and the
distribution of activity along the length of the fibers make the assumption of a small number
of dipoles very hard to justify. The distributed linear methods, on the other hand, formulate
the problem as the estimation of the magnitudes of a large number of dipoles whose loca-
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 76
tions and orientations are fixed. Electric fields add linearly, so the relationship between the
measurements and the dipole magnitudes can be expressed as the linear system in Equation
4.1.
d(n) = Lj(n) + ǫ(n) (4.1)
Here, d(n) is the Mx1 vector of measurements at instant n, j(n) is the 3Nx1 vector
of current dipole magnitudes at instant n, and ǫ(n) is additive noise (assumed Gaussian).
The Mx3N matrix L is the leadfield matrix, which relates the potential measurement at each
electrode contact to the existence of a unit source at each possible location. It is constructed
as described in the section on the forward problem above. There are three leadfield columns
for each of N mesh elements, corresponding to the three orthogonal dipole orientations.
Note that the system at each time instant can be considered as independent of the other
time instants because quasistatic conditions hold [92].
The difficulty of recovering j(n) from d(n) and L comes from the fact that the problem
is underdetermined and ill-posed. In order to obtain a stable solution, additional constraints
must be added to the problem. An overview of the most commonly used types of constraints
was given in the Literature Review section, and can also be obtained in several reviews
[122, 104]. Here, the method chosen is the standardized low resolution brain electromagnetic
tomography (sLORETA) method [123]. In the absence of additional information about the
solution, sLORETA usually outperforms other instantaneous distributed linear methods, but
does this at the cost of decreased spatial resolution. In other words, it produces images that
are blurred but have well-localized peaks; this smoothness is imposed on the solution as a
constraint to deal with the ill-posed nature of the problem. The algorithm has been shown
to localize single sources with zero error in the noiseless case [123]. sLORETA consists of two
steps: first, a regularized minimum-norm least-squares solution of Equation 4.1 is obtained;
second, that solution is normalized using the diagonal of the resolution matrix, which pro-
vides information about the bias of the solution for each entry in j(n). The regularization
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 77
parameter is chosen here by means of the cross-validation error function, as suggested by
Pascual-Marqui [121, 123].
Our choice of algorithm is based on three criteria. First, it should not make the
assumption that there are only a small number of active sources, as was already discussed
above. Second, it should be reasonably fast, in view of possible application in a neuro-
prosthetic system. Lastly, it should not assume that the source distribution generating the
measurements is stationary over a certain time interval, because outside of a controlled ex-
periment there would be no way to ensure that the set of active pathways is not changing
over the observation period. sLORETA satisfies all three of these requirements.
In order to reduce the number of variables that need to be solved for in the inverse
problem, we can further restrict the location and orientation of the dipoles. Active fibers can
only be located in the endoneurium, so the perineurium and epineurium regions are removed
from the solution space. The current dipoles that are used to model the extracellular field of
an action potential are oriented axially along the fiber [92], so that dipoles in the other two
orthogonal directions can be eliminated from consideration, thereby reducing the number of
variables by two thirds. In a real nerve, the fibers are not completely straight, but rather
exhibit a slight wave pattern, meaning that our decision to use a single dipole direction is
an approximation. Nonetheless, dipoles along the axis of the nerve will capture the largest
component of the activity. When we proceed to using this methodology on experimental
data in the next chapter, the inaccuracy introduced by neglecting the radial components of
the current dipoles should lead to a considerably smaller degradation in performance than
that which would result from tripling the number of variables.
Spatio-temporal constraint
Keeping in mind that the basic approach to solving underdetermined inverse problems is
to constrain the solution based on our knowledge of the problem, we investigate a spatio-
temporal constraint based on the electrophysiology of nerve fibers. Assuming that we are
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 78
applying what follows to myelinated fibres, a spatio-temporal constraint can be implemented
based on the formula in Equation 4.2.
V (x, t) = V (x− ds, t− ds/v) (4.2)
V(x,t) is the transmembrane voltage at position x along the fibre and at time t, ds
is the spatial distance separating two consecutive nodes of Ranvier, and v is the conduction
velocity. ds and v can be estimated from the diameter and type of the fibre. The formula is
simply saying that the activity at a node is ideally identical to the activity at the previous
node ds/v seconds in the past.
Putting this equation to use assumes that the type (myelinated vs. unmyelinated)
and diameter of the fibres at a given location in the nerve are known. Alternatively, even if
no information about the location of different fibre types is available, the constraint can be
applied if a single type of fiber is active at a given time, by assuming that the whole nerve
is composed of fibers of that type. In that situation, before performing the localization, one
can identify the nerve conduction velocity (and thus fiber type) from the MCC electrode
using a method such as the one proposed by Rieger et al. [136].
In order to incorporate this information as a constraint, temporal coupling must be in-
troduced into the problem. Ideally, the whole spatio-temporal system could be expressed and
solved as a single linear system, by concatenating the measurement and source vectors for all
time instants [24, 154]. This is shown in Equation 4.3, which is very similar to Equation 4.1,
with the exception that the vectors dc, jc, and ǫc include the complete information about the
system at all time instants in the observation time window (i.e. dc = [d(1)Td(2)T...d(n)T]T,
with equivalent definitions for jc and ǫc, and n being the total number of time samples). Lc
is a block diagonal matrix, where each block is equal to L.
dc = Lcjc + ǫc (4.3)
The problem with this approach is that the number of variables is multiplied by
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 79
the number of time instants, making the system intractable for any realistic application.
Therefore, for computational reasons, we restrict ourselves to coupling two time instants at
a time. Specifically, in accordance with Equation 4.2, we solve the systems corresponding
to the pairs of time instants (t(1), t(1)+ds/v), (t(2), t(2)+ds/v), etc. ds/v is rounded to
the nearest integer. To couple the solutions of the two time instants together in the desired
manner, a coupled leadfield and a non-diagonal weight matrix are constructed as shown in
Equation 4.4. In general, the weight matrix in a distributed linear method for this type of
problem is applied to the norm of the solution in the minimum-norm least-squares problem
and is used to apply a priori constraints to the solution. The incorporation of a weight
matrix into the sLORETA algorithm is described in the original paper by Pascual-Marqui
[123].
Lc =
L 0
0 L
Hc =
I −A0 I
(4.4)
I is the identity matrix corresponding to the number of variables in a single time
instant, and A contains the spatial element of the constraint. If elements i and j of the
solution vector correspond to locations along the same pathway and separated by a distance
ds, then Ai,j is set to 1. In this way, each row of A corresponding to a location in a
constrained pathway contains a single entry of value 1. Hc therefore constrains those values
in the solution at time t+ds/v to have as close a value as possible to the element situated
ds lower in the solution at time t by penalizing differences in these values, as desired.
In the case of our simulations, the constraint assumed that the whole nerve was
composed of myelinated fibres of diameter 10 µm with nodes of Ranvier spaced 1 mm apart
and a conduction velocity of 50 m/s. These values are the same as the ones used to generate
the simulated measurements.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 80
4.2.2 Evaluation of the results
The characteristic of the solution that is of most interest to us is the position of the active
pathways. For that reason, we are primarily interested in the position of the sources in a
cross-section of the nerve, rather than in their longitudinal position. The three-dimensional
solution can therefore be projected onto a two-dimensional cross-section for the purposes
of computing an error metric. This is done by summing the absolute values of the activity
of all the mesh elements with the same cross-sectional position. The value associated with
each location in the resulting two-dimensional projection is therefore a representation of how
much activity was spread along the length of the nerve at that position. This process implies
that a nerve fibre has a straight path with very little radial deviation along the length of
the nerve segment; because there may in fact be small deviations in vivo, we stress that the
actual source localization in no way depends on this assumption. It is simply used to obtain
useful metrics to evaluate the performance of the method in the context of the simulations,
and is justified here because the simulated nerve fibres were in fact straight.
To further simplify the evaluation of the localization process and obtain concise met-
rics, we sum the estimates from all the time instants of a given trial before performing the
two-dimensional projection. The three-dimensional activity over the time interval is there-
fore summarized as a single two-dimensional source distribution. It is important to clarify
that summing the estimates of all the time instants does not presuppose stationarity of the
underlying sources; it simply provides an indication of what sources were active at some
point in the time interval.
Lastly, the estimated source locations are obtained by finding the local maxima in the
final two-dimensional projection. These local maxima are used to estimate the localization
error, as well as the number of missed and spurious pathways. The metrics are computed as
follows:
1. The mesh is interpolated onto a regular grid. The local maxima are detected on this
grid by comparing the activity at each location with the activity at all locations within
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 81
50 µm. The peak coordinates are obtained, and the results mapped back onto the
FEM mesh cross-section.
2. Each peak in the estimate is associated with the true pathway closest to it. A local-
ization error is obtained for each of the true pathways by computing the peak-to-peak
distance between that pathway and the closest of the estimated peaks assigned to it.
The other peaks associated with that true pathway, if any, are counted as spurious
peaks.
3. True pathways that do not have any peaks associated with them in Step 2 are counted
as missed pathways.
4.3 Results
4.3.1 One-pathway case
The mean values over 100 trials of the localization error, number of spurious pathways, and
number of missed pathways when a single true pathway is present are shown in Figure 4.2.
When the idealized leadfield was used with no constraints, the localization error was relatively
independent of the noise level, with non-monotonic variations between a minimum of 0.137
mm and a maximum of 0.166 mm. When the spatio-temporal constraint was added, similar
non-monotonic variations were observed between a minimum of 0.134 mm and a maximum
of 0.182 mm. When the correct leadfield was used, a clear relationship with the noise was
observed: the error increase monotonically from 0.078 mm to 0.166 mm without the spatio-
temporal constraint, and from 0.081 mm to 0.175 mm with the constraint. The statistical
significance of this finding was confirmed by the fact that, when testing the relationship
between error and noise against a null hypothesis of no correlation, p was less than 0.05
only in the cases where the correct leadfield was used. The difference in results between
the simulations with the idealized leadfield and those with the correct one is due to the
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 82
−5 0 5 10 15 20 25 30 35 40 45 500
0.1
0.2
0.3
Noise Level (%)
Loca
lizat
ion
Err
or (
mm
) Average Localization Error, 1−Pathway Case
ILIL−CCLCL−C
−5 0 5 10 15 20 25 30 35 40 45 50
0
2
4
Noise Level (%)
Spu
rious
Pat
hway
s
Average Spurious Pathways, 1−Pathway Case
ILIL−CCLCL−C
−5 0 5 10 15 20 25 30 35 40 45 50−1
0
1
2
3
Noise Level (%)
Mis
sed
Pat
hway
s
Average Missed Pathways, 1−Pathway Case
ILIL−CCLCL−C
Figure 4.2: Localization errors, spurious pathways, and missed pathways for the one-
pathway case (n = 100). IL: idealized leadfield; IL-C: idealized leadfield with spatio-temporal
constraint; CL: correct leadfield; CL-C: correct leadfield with spatio-temporal constraint.
geometry error. In other words, even when the amount of measurement noise is low, the
localization algorithm must still deal with a large amount of error if the leadfield is based
on an inaccurate model of the region. The relatively constant error in the idealized leadfield
cases additionally seems to suggest that there is a plateau in the amount of localization error
as the noise increases, at least for the range of values examined.
Spurious pathways were seen to constitute the biggest obstacle to the applicability of
the source localization approach. The number of spurious pathways increased monotonically
with the noise in all cases (this relationship was statistically significant in all cases except
IL, possibly because the number of spurious pathways seems to reach a plateau early in that
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 83
situation). When the idealized leadfield was used, the metric varied from 1.05 to 3.24. When
the spatio-temporal constraint was added, the range of values was from 1.14 to 2.51. When
the correct leadfield was used without the constraint, the range was 0.02 to 2.62, and with
the constraint it was 0.06 to 1.74. As clearly visible in Figure 4.2, this data reveals that
the amount of information incorporated into the source localization problem has a direct
impact on the number of spurious pathways: the spatio-temporal constraint led to a marked
decrease regardless of the leadfield used, and the combination of the correct mesh with the
constraint noticeably outperformed all of the other cases. One-way ANOVA followed by
a multiple comparison test between the four cases, performed at every noise level, showed
statistically significant (p < 0.05) differences between all pairs except the pair (IL-C, CL).
This confirmed the differences qualitatively visible in Figure 4.2. The exception was the
0% noise case, in which the choice of leadfield led to significant differences but the spatio-
temporal constraint did not. Overall, the presence of the spatio-temporal constraint led
to minor increases in localization error at high noise levels, but marked reductions in the
number of spurious pathways, which seems to be a worthwhile trade-off. The error increase
was not statistically significant (p > 0.05), and may be due to the larger number of variables
in the coupled problem. No missed pathways were observed, which is not surprising given
that a single pathway was present in the region.
Figure 4.3 (panels a), c) and e)) shows an example of a localization trial. The smooth-
ing effect of sLORETA is clearly visible, as is the presence of spurious pathways. Note that
the trials shown in this figure (for both the one- and three-pathways cases) are selected to
help the reader visualize the concepts being discussed; they are not necessarily the most
representative of the method’s performance. For that information, the reader should rely
rather on the metrics discussed in the text and shown in Figures 4.2 and 4.4.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 84
Contact column1 2 3 4 5 6 7 8
1
2
3
4
5
6
7
Contact ring
1−Pathway Measurements
(a)
Contact column
Contact ring
1 2 3 4 5 6 7 8
1
2
3
4
5
6
7
3−Pathways Measurements
(b)
(c) (d)
(e) (f)
Figure 4.3: Left side: a) Normalized simulated measurements obtained at all 56 contacts
for one of the one-pathway trials with a noise level of 20%. c) Estimated pathways obtained
by applying the source localization procedures to the measurements in a), using the idealized
leadfield combined with the spatio-temporal constraint. e) True location of the pathway for
this trial. Right side: Equivalent data for one of the three-pathways trials. The sharp peaks
seen in the estimates are the output of the peak detection algorithm, superimposed on the
source localization solution. Likewise, the nerve outlines on the floors of figures c)-f) have
been added to help the reader visualize the location of the pathways within the nerve.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 85
4.3.2 Three-pathways case
Panels b), d), and f) of Figure 4.3 show an example trial for the three-pathways case. The
mean values over 100 trials of the localization error, number of missed pathways, and number
of spurious pathways when three true pathways are present are shown in Figure 4.4. The lo-
calization error when the idealized leadfield was used with no constraints was again relatively
independent of the noise level (p > 0.05 under a null hypothesis of no correlation), varying
non-monotonically between 0.152 mm to 0.181 mm. With the spatio-temporal constraint,
the range was 0.155 mm to 0.179 mm. The errors when the correct leadfield was used were
once again an increasing function of the noise (p < 0.05), ranging from 0.083 mm to 0.182
mm without the constraint, and from 0.087 mm to 0.180 mm with the constraint. These
results are very similar to those seen in the one-pathway case.
The general trends for spurious pathways were also similar to those in the one-pathway
case, including the effects of the constraints. Although the increase with noise was not
strictly monotonic in all cases, that was nonetheless the trend (p < 0.05 in all four cases).
The number of spurious pathways when the idealized leadfield was used varied from 0.47 to
1.72 without the constraint, and from 0.64 to 1.15 with the constraint. When the correct
leadfield was used, the range was from 0.02 to 1.24 without the constraint, and from 0.02 to
0.88 with the constraint.
When multiple true pathways were present, the number of missed pathways increased
dramatically. The general trend was an overall decrease as the amount of noise increased,
although this proved statistically significant only in the cases with the correct leadfield.
When the idealized leadfield was used, the number of missed pathways ranged from 0.84 to
0.31 without the constraint, and from 0.84 to 0.58 with the constraint. When the correct
leadfield was used, the range was from 1.44 to 0.38 without the constraint, and from 1.57 to
0.62 with the constraint.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 86
−5 0 5 10 15 20 25 30 35 40 45 500
0.1
0.2
0.3
Noise Level (%)
Loca
lizat
ion
Err
or (
mm
) Average Localization Error, 3−Pathways Case
ILIL−CCLCL−C
−5 0 5 10 15 20 25 30 35 40 45 50
0
2
4
Noise Level (%)
Spu
rious
Pat
hway
s
Average Spurious Pathways, 3−Pathways Case
ILIL−CCLCL−C
−5 0 5 10 15 20 25 30 35 40 45 50−1
0
1
2
3
Noise Level (%)
Mis
sed
Pat
hway
s
Average Missed Pathways, 3−Pathways Case
ILIL−CCLCL−C
Figure 4.4: Localization errors, spurious pathways, and missed pathways for the three-
pathways case (n = 100). IL: idealized leadfield; IL-C: idealized leadfield with spatio-temporal
constraint; CL: correct leadfield; CL-C: correct leadfield with spatio-temporal constraint.
4.4 Discussion
The localization of active pathways in a peripheral nerve was approached as an inverse
problem of bioelectric source localization, using simulated measurements from a 56-contact
nerve cuff electrode. This is an ill-posed inverse problem, and problems of this class can
only be solved satisfactorily if sufficient appropriate constraints are imposed on the solution.
We therefore compared the performance of the approach using idealized and correct models
of the nerve geometry, and in the presence of a spatio-temporal constraint based on the
electrophysiology of myelinated nerve fibres. As expected, the overall performance improved
as more information was incorporated. Nonetheless, most of the simulated cases had mean
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 87
localization errors in the 140 µm to 180 µm range (in a 720 µm-diameter endoneurium, with
10 µm-diameter individual fibres), and unacceptably high numbers of spurious pathways.
Only at very low noise levels and with accurate constraints did the performance reach levels
that would make the approach reliably usable in practice. Even in those situations, however,
the number of missed pathways was high, due partly to the low resolution of the sLORETA
method. Note that the decrease in the number of missed pathways as noise increased in
the three-pathways case is related to the way that the metrics were computed. Recall that
each true pathway is associated with the closest estimated pathway. Therefore, if there are
several spurious estimated pathways distributed across the region, then there is a greater
chance that some of them will be associated with a true pathway and therefore reduce the
number of missed pathways. This explains why the number of missed pathways decreases
as the number of spurious pathways increases, and furthermore why cases that have the
best performance in the other metrics (e.g. the combination of the correct leadfield and
the spatio-temporal constraint) have more missed pathways. In light of this limitation of
our metrics, the most revealing values for the number of missed pathways are the ones
corresponding to cases with very few spurious pathways. For instance, the simulations using
the correct leadfield with the constraint at 0% and 10% noise have close to 0 spurious
pathways. In those cases, the number of missed pathways is approximately 1.5, or half the
number of true pathways. This is a very high proportion, but it is not entirely unexpected.
Recall that sLORETA imposes smoothness on the solution, and therefore is not capable of
reliably distinguishing closely spaced sources. Therefore, if two pathways are close to one
another, they will be lumped together in the estimate, thereby producing missed pathways.
In addition, our determination of the location of the estimated pathways is based on local
maxima. Upon visual inspection, however, it becomes clear that pathways do not always
produce local maxima, but may significantly affect the shape of the solution in other ways
(for example by producing elongated ridges). There is therefore a need for a better method
of identifying the estimated pathways, but given the variability of shapes and magnitudes
that may occur this is not a trivial problem.
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 88
Some remarks are in order regarding the way that information was incorporated into
the problem in this study. The simulations that use the correct leadfield and no noise should
be considered a mostly theoretical situation useful for examining the effects of adding in-
formation, because such an exact correspondence will never be found in practice, even if
detailed anatomical information is available. On the other hand, when noise is incorporated
into the simulations, it can be interpreted as a combination of measurement noise and geom-
etry noise (i.e. mismatch between the leadfield and reality), such that the simulations with
high noise may be more indicative of the expected performance in practical situations where
anatomical information is available. As for the spatio-temporal constraint, it has already
been mentioned that, in the absence of anatomical information about the layout of fibre
types in the nerve, it is useful only if a single type of fibre dominates the nerve’s activity at
a given time. Although it is feasible to ascertain in practice whether or not this assumption
is reasonable [136], it does nonetheless restrict the range of practical situations in which the
constraint would be usable. It is therefore also used in this study as much as a theoretical
tool to examine the effects of constraints as a suggestion for a practical technique. Overall,
there is no doubt that more work is required to develop constraints that are both useful from
the point of view of the inverse problem and whose practical implementation is realistic.
More generally, these simulations indicate that in order for the peripheral nerve source
localization approach to ever be usable as part of a neuroprosthetic system, several advance-
ments are needed. First, noise reduction is essential, and could take the form of improved
instrumentation, better isolation of external signals at the cuff level, or neuromodulation
techniques to boost the amplitude of the recorded nerve signal [146]. Second, methods
should be investigated to obtain precise images of a nerve’s anatomy in vivo. For exam-
ple, an adaptation of electrical impedance tomography techniques [9] to peripheral nerves
could be considered, and studies have already shown that fascicles within a nerve can be
imaged using ultrasound [77]. It is interesting to note, however, that in our simulations
the performance using only the spatio-temporal constraint with the idealized leadfield (IL-
C) was not very far from that using only the correct leadfield without the spatio-temporal
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 89
constraint (CL), at least for cases with non-zero noise. This raises the possibility that if
sufficient physiological and contextual information could be provided as constraints, and the
noise brought down to a manageable level, then an idealized geometry could perhaps still
be used. This would be analogous to using a three-sphere head model in the EEG/MEG
source localization problem, as opposed to a patient-specific MRI-based anatomical model.
Another cause for cautious optimism is that the performance for the very idealized case
of noiseless measurements and full constraints approached acceptable levels, which implies
that the very large number of dipoles in the region (50 per pathway in this study) may
not be an insurmountable obstacle in itself to pathway localization. Importantly, although
this study was conducted with an eye to applying the method to neuroprosthetic systems,
the potential for neural system identification (in this case identifying the peripheral control
signals involved in specific tasks) is also very interesting, and has a different set of restric-
tions. In particular, in a context where repeated trials and off-line processing options are
available, the noise could be significantly reduced by trial averaging and a wider range of
algorithms explored, since computation time would no longer be as much of an issue. It may
therefore be worthwhile to keep exploring the source localization approach in this context.
For neuroprosthetic applications, sub-fascicular resolutions do not appear realistic on the
short term, and the usefulness of the method as a framework for determining the activation
of combinations of several fascicles will be contingent on future developments like the ones
discussed above (particularly if numerous small fascicles are involved, as is more likely to be
the case in humans).
The present study aimed to estimate the overall viability of the source localization
approach. There are various topics that were not discussed in detail here but would become
very relevant if the performance was improved enough to make the approach usable. First,
it is important to keep in mind the very high temporal resolution of the method, which is
not reflected in the metrics used in this study. For each dipole location, the algorithm pro-
duces a complete activation time series, making it possible to study transmembrane current
waveforms and firing frequencies for precise locations in the nerve. Second, the influence of
Simulation of Bioelectric Source Localization in the Rat Sciatic Nerve 90
the number of contacts on the performance would need to be explored. We limited ourselves
here to using a model of an existing electrode, but as manufacturing techniques improve so
will the number of contacts that can be placed on a cuff, and this will undoubtedly have an
influence on the source localization performance. There may also be a plateau to the benefits
of increasing the number of contacts, such that the optimal number should be sought. As-
pects of this question will be explored further in Chapter 7. Third, the number of pathways
that the method can accurately localize in a given situation should be determined. The
use of three pathways in this study was motivated by the need to study a simple multi-
pathway case, rather than by physiological considerations. That said, the term “pathway”
should not necessarily be equated with a single nerve fibre, particularly given the limited
resolution and the coarseness of the mesh. Multiple closely spaced nerve fibres that have
a functional relationship and a roughly synchronous firing pattern (e.g. compound action
potential (CAP)) may therefore be considered a single pathway, as could a very small fas-
cicle. Similarly, the currents generated by a few isolated nerve fibres that fire at the same
time as a larger coordinated group of fibres elsewhere in the nerve may be drowned out
and confused with noise; in that case, they may be missed by the algorithm, but may not
significantly hamper the localization of the “main” pathway. The last issue that deserves
to be discussed is that of computation time. Our choice of algorithm was partially based
on its speed. On a 2.33 GHz dual-processor workstation, localization for a single time in-
stant using the idealized leadfield and no constraints took approximately 1 second. When
using the correct leadfield or the spatio-temporal coupling, the computation time increased
roughly proportionally to the number of variables to be solved for. While these computation
times are not yet suitable for real-time implementation, it is nonetheless a realistic target as
computational speed continues to increase and more efficient implementations of the algo-
rithm are explored. As mentioned earlier, in situations where speed is not an issue, slower
algorithms could be explored, for example ones based on lp norms (e.g. [194]).
Chapter 5
Validation of the Source Localization
Approach on Physiological Data
5.1 Introduction
The next step in the study is to collect experimental data under controlled conditions and
use it to test the source localization approach in practice. Despite the fact that the sim-
ulation results in the previous chapter suggest that the performance may be poor due to
high sensitivity to noise and model inaccuracies, experimental data is needed to validate the
qualitative observations regarding the influence of the spatio-temporal constraint and more
accurate leadfield. In addition, even if the bioelectric source localization approach presented
thus far proves to be of limited utility, we will still seek to determine what useful information
can be extracted from the MCC recordings obtained under realistic conditions, which can
only be done accurately using experimental data. For convenience, from this point on, we
will refer to both the spatio-temporal constraint and the use of a nerve-specific leadfield (in-
stead of one based on an idealized model) as “constraints”. Although the two strategies are
of a different mathematical nature, they both represent attempts to improve performance
by incorporating additional information into the problem.
91
Validation of the Source Localization Approach on Physiological Data 92
5.2 Methods
5.2.1 Data collection
Animals
Seven old male Long-Evans breeders (640 g to 850 g) (Charles River Laboratories Inc.,
Wilmington, MA, USA) were used. All rats were acclimatized for one week prior to use in
the experiment. Food and water were provided ad libitum. A 12 hour lights on/off cycle
was used. All animal care and use procedures conformed to those outlined by the Canadian
Council on Animal Care (CCAC). Experiments were performed on seven rats but, due to
technical difficulties with the first animal, that data had to be discarded. Hence, data from
only six animals was analyzed in this study.
Anaesthesia
All animals were anesthetized with a single bolus injection of pentobarbital (60 mg/kg,
intraperitoneal), and their lower backs and legs were shaved and treated with povidone-
iodine. When an adequate depth of anesthesia was attained (loss of corneal reflex and loss
of sharp pain sensation), the animals were positioned prone on the operating table.
Surgical exposure
An oblique incision was centered over the posterior (dorsal) aspect of the hip. The incision
was extended proximally to the midline and distally parallel with the fibers of the gluteus
maximus to the posterior margin of the greater trochanter. The incision was then directed
distally, parallel with the femoral shaft to the posterior fossa of the knee.
The deep fascia was exposed and divided in line with the skin incision. By blunt
dissection, the gluteus maximus was split in line with its fibers and retracted to expose the
sciatic nerve and short external rotator muscles. Care was taken not to disturb the superior
gluteal vessels in the proximal part of the exposure.
Validation of the Source Localization Approach on Physiological Data 93
The sciatic nerve was exposed as far proximally as possible to allow adequate exposure
for application of the recording cuff. The recording cuff was applied to the sciatic nerve
following application of the three stimulating cuffs (see details in the next section).
The sciatic nerve was then followed distally and three branches were identified: the
sural nerve, peroneal nerve, and tibial nerve. The soft tissue surrounding each of these nerves
was carefully blunt dissected to allow a stimulating cuff to be applied to each nerve.
Electrode placement and recording parameters
A matrix design polyimide spiral nerve cuff electrode [157] (Figure 1.1) was placed on the
sciatic nerve, just proximal to its division into its peroneal and tibial branches. This cuff
was 23 mm long, 1 mm in diameter and contained 56 contacts, arranged in 7 rings of 8
contacts. This electrode was used to record the nerve activity during the experiments. In
addition, three tripolar stimulating polyimide spiral nerve cuffs (8 mm long and 1 mm in
diameter) were placed around the tibial, sural, and common peroneal nerves. The center
ring of the stimulating electrodes contained 8 contacts that were shorted together, resulting
in traditional tripole cuffs. The stimulating cuffs were placed first (Figure 5.1(a)), followed
by the recording cuff (Figure 5.1(b)).
The measurements from the cuff on the sciatic nerve were acquired using a SynAmps2
64-channel amplifier (Neuroscan Inc., Herndon, VA, USA), with a sampling rate of 20 kHz
and a gain of x2010. The signals were band-pass filtered between 300 Hz and 3 kHz. The
reference for the recordings was a contact included in the matrix cuff design and located just
outside the cuff. A needle electrode in the calf was used as the ground.
Direct fascicular stimulation using nerve cuff electrodes
The tibial, peroneal, and sural nerves were stimulated using the 8 mm cuff electrodes, first in-
dividually, then in every possible combination. The stimulation pulses were generated using
Compex Motion stimulators (Compex SA, Switzerland). Although the intended stimulation
Validation of the Source Localization Approach on Physiological Data 94
(a)
(b)
Figure 5.1: a) The tibial, peroneal, and sural nerves are exposed. Each has a stimulating
cuff wrapped around it. The sciatic nerve has been exposed but the recording cuff has not yet
been placed. b) The exposed sciatic nerve with the recording cuff wrapped around it.
parameters consisted of 10 µs 2 mA pulses (2 mA being comfortably higher than the thresh-
olds reported in the literature for pulses of this duration [145, 116, 45, 183]), in practice the
pulses were shorter and had more variable amplitudes, due to technical difficulties noticed
only after the fact and explained in more detail below. 100 trials were conducted for each
fascicle, at a frequency of 2 Hz.
Nerve samples
A section of the sciatic nerve of each rat was removed and preserved in a formaldehyde
solution. These samples were later used to obtain images of the nerve cross-sections (see
Validation of the Source Localization Approach on Physiological Data 95
Section 5.2.2), thereby providing anatomical information that can be incorporated into the
source localization process.
Issues encountered
Three difficulties were encountered during the experiments. Their impact on our study is
described below.
1. Recording cuff connector: The current design of the 56-channel recording cuff was
created with in vitro experimentation in mind. Using it in vivo presented a challenge
because the connector linking the electrode to the amplifier was very bulky (Figure
5.2). As a result, it had to be held in place manually, which implies that there was
some variation in the cuff position over the course of an experiment. This will have to
be taken into account when interpreting the results of the source localization using the
nerve cuff data. During one of the experiments, we were able to clamp the connector in
place, and the data from that rat will therefore provide an indication of how much of a
detrimental impact positioning the connector by hand might have had. Clamping was
unfortunately not possible in the other cases, because of the placement angle and the
very limited tension that the cuff could bear without coming off the nerve. The only
satisfactory solution to this problem will be a redesign of the cuff and its connector by
our collaborators at the University of Freiburg, however this was not possible within
the time frame of the current study. To our knowledge, there currently exists no other
nerve cuff design with a sufficient number of contacts. The limitations imposed by the
connector were therefore in our opinion unavoidable in the context of this project.
2. Stimulation artefact: Although we stimulated using very short pulses in an attempt
to limit the stimulation artefacts, capacitive effects nonetheless resulted in substantial
artefacts that overlapped with the action potentials. This issue arose because the
amplifiers were not blanked during the stimulation (our recordings were performed
using AC coupling in order to achieve the necessary gain, and the amplifier’s blanking
Validation of the Source Localization Approach on Physiological Data 96
Figure 5.2: Connector linking the 56-contact cuff electrode to the amplifier and data-
acquisition system. During the recordings, the connector was held up either manually or
using a clamp.
feature was not available in this mode). The amplifiers did not saturate, but they were
susceptible to an impulse artefact with a time constant of approximately 0.5ms and thus
overlapping with the signal of interest. The simplest way to avoid this problem would be
to use equipment with which amplifier blanking is possible, but such resources were not
available in the context of this project. On a more positive note, we can take advantage
of the large number of contacts in the cuff to implement artefact reduction techniques,
such as using a common average reference. Nonetheless, some contamination of the
signal is unavoidable. An example trial is shown in Figure 5.3.
3. Stimulation pulse characteristics: It was discovered after the experiments that
the stimulators used were not able to adequately control the pulse characteristics,
due to the fact that our desired parameters were at the very limit of the stimulators’
specifications. Therefore, instead of 10 µs 2 mA pulses, the stimulators produced pulses
2-4 µs long and with amplitudes in the 0.7 to 3.8 mA range approximately. Fortunately,
these pulses were still able to reliably produce action potentials in the nerve. This was
established both from the muscle twitches observed during the stimulation and from
the spatio-temporal distribution of activity in the cuff, which is consistent with action
Validation of the Source Localization Approach on Physiological Data 97
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
Rat 1, Tibial Nerve, Trial 1
Figure 5.3: Example of the recordings of one tibial branch trial. The upper left-hand plot
shows the raw recordings for all 56 channels. The upper right-hand plot shows those same
recordings after conversion to a common-average reference. The lower plots show the same
data for one contact only, taken from the middle ring of the cuff.
potentials rather than with signals originating outside the cuff. For the purposes of this
study, the presence of action potentials in a given pathway is more important than the
exact amplitude of those action potentials, such that this problem with our stimulation
does not constitute a significant limitation of the study. In addition, another technical
issue with the stimulator resulted in four successive action potentials being generated
at every trial. For simplicity, all of the analysis that follows was conducted using only
the first action potential from each trial.
Validation of the Source Localization Approach on Physiological Data 98
5.2.2 Construction of nerve-specific leadfields
Samples of the sciatic nerves of the rats used in each experiment were collected. A cross-
section was cut 1 cm proximal to the point where the nerve branched into its tibial and
peroneal components, which corresponds approximately to the mid-point of the recording
cuff. The cross-sections were stained with hematoxylin and eosin (H&E staining) in order
for us to obtain an image detailing the position and shape of the fascicles within the nerve.
These images were then traced and converted into finite-element models. Note that the three-
dimensional models were created by extruding the cross-section in both direction, resulting in
models with no longitudinal variation in the anatomy. Because of the progressive branching
of the fascicles along the length of the nerve, these models are still simplifications compared to
the real anatomy, but they are more detailed than the idealized model presented in Chapter
4. The cross sections obtained from the samples are shown in Figure 5.4.
Recall that the MCC is a spiral cuff designed to be implanted on nerves with a
diameter of 1 mm, which corresponds approximately to the diameter of a rat sciatic nerve.
As visible in Figure 5.4, however, the nerves used in practice were slightly larger. This is
likely a consequence of the fact that we deliberately used the largest rats available to us in
order to have space to implant all the electrodes needed for the experiments (the recording
cuff as well as the three stimulating cuffs). Because of the spiral cuff design, which is intended
to allow the cuff to adjust to the size of the nerve, it was still possible to wrap it around the
nerves without excessive compression or deformation. The spread of contacts around the
cuff wall, however, is designed for a 1 mm diameter, such that having a larger diameter will
result in an certain section of the cuff’s inner wall being devoid of contacts. Furthermore,
we do not know where this section is located. In order to deal with this situation in the
FE models, the inter-contact distance was held constant and the number of contacts was
determined by the diameter of the cuff. For instance, if the cuff had a diameter of 1.5 mm
instead of 1 mm, then the number of contacts was 12 contacts per ring instead of 8, for a
total of 84 contacts instead of 56. The leadfield constructed from this model would therefore
Validation of the Source Localization Approach on Physiological Data 99
(a) Rat 1. (b) Rat 2. (c) Rat 3.
(d) Rat 4. (e) Rat 5. (f) Rat 6.
Figure 5.4: Sciatic nerve cross-section of each rat after hematoxylin and eosin (H&E)
staining. The sections were cut 1cm proximal to the point where the nerve branches into the
tibial and peroneal nerves.
Validation of the Source Localization Approach on Physiological Data 100
Table 5.1: Cuff diameters in nerve-specific rat models.
Rat Cuff diameter Number of
(mm) contacts
1 1.5 12
2 1.5 12
3 1.5 12
4 1.4 11
5 1.4 11
6 1.4 11
have 84 rows. Of course, only 56 channels were actually recorded during the experiments,
and so for the purposes of source localization only 56 of these rows were used at a time,
corresponding to 8 adjacent columns of contacts. More detail on the process of determining
which 8 columns to use is provided in Section 5.2.3. Table 5.1 gives the cuff diameter and
corresponding number of columns of contacts for each of the rats. The diameters were chosen
based on the images in Figure 5.4. Cross-sections of the resulting meshes are shown in Figure
5.5.
5.2.3 Evaluation of the source localization performance
The source localization was applied to the experimental data using the same approach as in
the simulation study. Specifically, the sLORETA algorithm was applied in four cases, namely
using the idealized and nerve-specific leadfields, each with or without the spatio-temporal
constraint. We chose not to use the leadfield reduction method described in Chapter 3
because it is not compatible with the spatio-temporal constraint and using it only on the
unconstrained cases would complicate the comparisons. However, a comparison of the results
using the idealized leadfield and no constraint, with and without the leadfield reduction
technique applied, found that the estimates in both cases were nearly identical (results not
Validation of the Source Localization Approach on Physiological Data 101
Figure 5.5: Cross-sections of the finite-element meshes produced for each of the rats using
the cross-sections in Figure 5.4. The colour-coding of the different types of tissue or material
is the same as in Figure 4.1
shown). This reinforces the validity of the technique, and the fact that it is not used here is
simply due to the need to limit the number of factors that vary between the different cases
studied, not to any limitations of the technique.
Before computing the source localization estimate, the data was converted to a
common-average reference, and the corresponding modification was made to the leadfield
as well. In each trial, channels with excessive variance or very small amplitude compared to
the other channels were marked as bad channels, and were not used in the localization. In
other words, the rows of the leadfield corresponding to those channels were removed. The
removal was performed before conversion to the common average reference, in both the data
Validation of the Source Localization Approach on Physiological Data 102
and the leadfield.
For each trial, the source localization was applied to a small time interval delimited
by the peaks of the action potential recordings at the first and last contacts (excluding bad
channels), plus 0.1 ms before and after this interval. The result of the source localization
consists of a three-dimensional estimate for each time instant in the time interval, for each
trial. As a consequence, the amount of data generated by this process is considerable,
and must be summarized in a convenient form in order to gauge the success of the source
localization. The following method was therefore used. First, an estimate of the activity
of each “pathway” (longitudinal column in the FE model) was obtained by summing the
absolute values of the estimated activities in all the elements of that column over the time
interval. Pathways with levels of activity that were equal to at least 50% of the maximum
level of activity in that trial were considered active pathways. Lastly, for each case (i.e.,
combination of branches in a given rat under given constraints), a two-dimensional mesh
cross-section was generated, where the intensity of each mesh element was determined by
the number of trials in which that pathway was judged to be active. This representation
gives a concise view of where the bioelectric activity was estimated to have originated in the
nerve.
Evaluation of the results is somewhat complicated by the fact that the position of the
contacts with respect to the nerve varied from one animal to the next. Attempts to record
the positions of the contacts were hampered by the bulky connector. The implications of
this are different depending on which leadfield is being used.
In the case of the idealized leadfield, the geometry is symmetric, and therefore rotating
the contacts will rotate the estimate but not otherwise alter it. As a result, the images should
be interpreted not by using the absolute location of the activity, but rather by examining
whether stimulating the different nerve branches generated activation at distinct locations
and whether the number and combination of active branches could be identified from the
results.
In the case of the nerve-specific leadfield, the geometry is not symmetric, such that
Validation of the Source Localization Approach on Physiological Data 103
rotating the contacts will result in an altogether different estimate, not just a rotated one.
We therefore rotate the contacts and perform the source localization for every possible ori-
entation. If, as described earlier, the diameter of the nerve was such that a section of the
cuff was devoid of contacts, then this rotation process will also take into account all possible
positions of the blank section. Because estimates are obtained for each possible rotation of
the contacts, the question remains of which set of estimates should be chosen. In the absence
of better information, we assume that the rotation that is closest to the actual position of
the contacts during the experiments will yield the best performance, and therefore simply
choose the best set of estimates when evaluating the performance. For each rat, the rotation
is chosen based on the overall performance for all fascicle combinations, and is fixed for that
animal (i.e., all estimates presented in the Results section use the same rotation for a given
rat). More details regarding the performance metrics that are used to choose the rotation
are given in Section 5.3.3.
5.3 Results
5.3.1 Using the idealized leadfield without the spatio-temporal
constraint
The first case tested was that of the idealized leadfield, without using the spatio-temporal
constraint. The summarized results of the source localization process are shown in Figure
5.6, for each rat and combination of branches. The figure shows that although some dif-
ferences can be observed between the three single-branch cases, these differences are minor.
Furthermore, the results for branch combinations are not recognizable as combinations of
the single-branch cases, nor can the correct number of pathways be readily obtained from
the source localization outcome. Indeed, in all animals, several of the possible combinations
yielded patterns of activity that were effectively indistinguishable from one another. It is
therefore clear that the source localization approach did not perform adequately, and in the
Validation of the Source Localization Approach on Physiological Data 104
form used in this section is not a viable method for identifying the active pathways. The
most likely causes for this failure are a too-large difference between the actual anatomy and
the model used to construct the leadfield, as well as the noise and stimulation artefact in-
terference. These combined factors led to a level of error that was too high for the source
localization to be successful.
5.3.2 Using the idealized leadfield with the spatio-temporal con-
straint
The second localization attempt consisted of once again using the idealized leadfield, this
time in combination with the spatio-temporal constraint. The application of the constraint
requires an estimate of the conduction velocity in the active pathway. This information was
estimated for each trial by computing the time delay between the signal’s arrival at the first
and last rings, given that the distance between these rings is known. The time at which
the signal arrives at a given ring was taken to be the mean of the times at which peaks
were detected for all the contacts in that ring. For the purposes of computing the spatial
distance between successive nodes of Ranvier, the fiber diameter was estimated from the
conduction velocity using a proportionality factor of 6 [69]. Note that we are treating the
CAP as having a single conduction velocity, which is a simplification. In other words, we are
using the dominant velocity in the CAP for the purposes of the spatio-temporal constraint
and ignoring the others. The summarized results of the source localization are shown in
Figure 5.7.
The figure shows results that are very close to the unconstrained case (at least in terms
of the spread of the activity, even if the intensities in the figures are sometimes different), and
therefore did not lead to a noticeable improvement in performance. This is not surprising,
because if the leadfield diverges too much from the real anatomy (as is suspected here), then
the relationships between variables on which the constraint relies will also not be sufficiently
close to reality to improve performance. It was also the case in the simulations that the
Validation of the Source Localization Approach on Physiological Data 105
Figure 5.6: Summary of the source localization performance for each combination of
branches in each rat, when the idealized leadfield is used without the spatio-temporal con-
straint. Refer to the text for the meaning of the intensities in the figures.
Validation of the Source Localization Approach on Physiological Data 106
Figure 5.7: Summary of the source localization performance for each combination of
branches in each rat, when the idealized leadfield is used and the spatio-temporal constraint
is applied.
Validation of the Source Localization Approach on Physiological Data 107
constraint reduced spurious pathways mostly by eliminating small erroneous fluctuations in
the estimate; it generally did not lead to very large qualitative changes in the estimate, and
therefore it is not surprising that it did not have such an effect on the experimental data
either. An analysis of the number of peaks in the estimate is conducted in Section 5.3.5.
5.3.3 Using the nerve-specific leadfield without the spatio-temporal
constraint
The next step was to use the leadfields constructed from the nerve sample cross-sections,
without the spatio-temporal constraint. The summarized results are shown in Figure 5.8.
As discussed above, the performance shown for each rat corresponds to only one possible
rotation of the position of the contacts with respect to the nerve anatomy. The rotation
shown was chosen by first performing the source localization for every possible rotation,
and selecting visually the one that yielded the best performance. The first criterion for
performance was to be closest to the ideal of having clearly different patterns for each single-
branch case, with the activity being strongest in a different fascicle each time in the cases
where the separation of the fascicles is visible in the anatomy. The second criterion was to
be closest to the ideal of having multi-branch cases that are recognizable combinations of
the single-branch cases.
When comparing Figure 5.6 to Figure 5.8, we can make the qualitative observation
that the nerve-specific leadfield led to somewhat better separation of the single-fascicle cases.
Nonetheless, it is in most cases difficult to clearly identify a single active fascicle and, worse,
even when there is a clear difference between the cases, the fascicle that appears most active
in the estimate is not always the correct one. For example, in Figure 5.8, the Rat 1 estimate
when the peroneal branch was being stimulated seems to suggest that the sural branch is
dominant, and vice versa (the fascicles in the figure corresponding, in order of decreasing size,
to the tibial, peroneal, and sural branches). In addition, in all cases where the division of
the fascicles was visible in the cross-section, the single-fascicle estimates should show a clear
Validation of the Source Localization Approach on Physiological Data 108
Figure 5.8: Summary of the source localization performance for each combination of
branches in each rat, when the nerve-specific leadfield is used without the spatio-temporal
constraint
Validation of the Source Localization Approach on Physiological Data 109
focus of activity in one fascicle, and this was usually not the case. In the case of multiple
fascicles, the use of the nerve-specific leadfield did not make an appreciable difference. In
both Figures 5.6 and 5.8, multiple-fascicle estimates generally either produced a pattern
that was reflective of only one of the active fascicles, or produced a pattern that was not a
recognizable combination. The former case is likely due to higher amplitude of the CAPs for
that fascicle (e.g. due to irregularities in the stimulation parameters, as explained above, and
the sizes of the branches), although a viable source localization approach should of course
be able to take into account sources of different intensities.
In the numerous cases where unexpected patterns were produced, for both single-
fascicle and multi-fascicle cases, there was nonetheless a significant amount of reliability
between trials (recall that red mesh elements in the figures represent pathways that were
active in almost all trials). This suggests that performance degradations were not due to
random types of noise, but to more consistent sources of error. The chief candidates here
are the stimulation artefact and the modeling errors. Because substantial reduction of the
artefact was achieved (Figure 5.3), it seems unlikely that it alone could account for errors
of this magnitude. The most likely conclusion is therefore that the nerve-specific leadfields
were still too coarse an approximation to achieve successful localization. The main culprits
are likely to be the fact that the model does not take into account longitudinal variations in
the anatomy, the fact that the nerve may have been pushed into a slightly different shape
when the cuff was placed on it, and the uncertainty in the positions of the contacts around
the nerve (within each one of the rotations that we took into account, there is room for some
shift, because the contacts were modeled as punctual locations but in fact extend over a
small rectangular area).
The conclusion to be drawn from these results is therefore that although using the
nerve-specific leadfields led to a slight improvement over the idealized leadfield, in the form
of increased distinguishability between single-fascicle cases, this improvement was still too
small to make the source localization useful in practice. Indeed, taking any of the images in
Figure 5.8 in isolation (i.e., without having the other estimates as a point of comparison),
Validation of the Source Localization Approach on Physiological Data 110
we cannot correctly infer the number and identity of the active fascicles, and this is the only
test that is relevant in a real-life neuroprosthetic application.
5.3.4 Using the nerve-specific leadfield with the spatio-temporal
constraint
The last case examined was that of the nerve-specific leadfields in combination with the
spatio-temporal constraint. These results are shown in Figure 5.9.
The application of the spatio-temporal constraint had similar effect when the nerve-
specific leadfield was used as when the idealized constraint was used. That is to say, it had
very little effect on the outline of the active regions within the cross-section. Once again
this may be due to the coarseness of the approximation, as well as to fact that the expected
effects of the spatio-temporal constraint (as per the simulation study) may be too subtle to
be easily visible in the type of display used in Figures 5.6 to 5.9. In the next section, we
attempt to compare the performances of the four cases in a more quantitative manner.
5.3.5 Influence of the constraints on the number of peaks in the
estimate
Recall that the simulation study in Chapter 4 evaluated the performance by examining the
localization error, the number of spurious pathways, and the number of missed pathways.
Computing the localization error requires knowledge of the source’s true location, which
is not available here. A qualitative evaluation of the localization accuracy can be made
in Figures 5.8 and 5.9 for rats in which the fascicles were separated, because the correct
fascicle is then known. As discussed above, this criterion indicates that the localization
errors in most cases were large. This observation is, however, difficult to translate into a
rigorous quantitative assessment, because of the spread of activity among several fascicles
in most estimates and the fact that we can only use this approach for four of the rats when
Validation of the Source Localization Approach on Physiological Data 111
Figure 5.9: Summary of the source localization performance for each combination of
branches in each rat, when the nerve-specific leadfield is used and the spatio-temporal con-
straint is applied.
Validation of the Source Localization Approach on Physiological Data 112
using the nerve-specific leadfields, and none of them when using the idealized leadfield. For
these reasons, we turn rather to the number of peaks in the estimates in order to perform a
quantitative validation of the conclusions of the simulation study.
For each combination of rat, branch combination and constraints, the number of peaks
in the estimate for each trial was computed in the same way as in the simulation study. The
means were computed and are reported in Figure 5.10. Before discussing the data in this
figure, it is necessary to point out that the recordings from Rat 3 were less reliable than
the others. Anticipating the results in Chapter 7, the data suggests that these recordings
were obtained with an improperly closed cuff, and therefore are largely contaminated by
the stimulation artefact (this will be more rigorously demonstrated in Section 7.3.2). As an
additional consequence of this problem, delimiting the time intervals on which to perform the
localization in each trial proved less reliable in the recordings from this rat. Overall, then, the
results from Rat 3 cannot be considered reliable, but for completeness they were nonetheless
reported above and in what follows. In the case of metrics that aggregate information from
several animals, we include the results both with and without Rat 3 included.
The data in Figure 5.10 allows us to determine the effect of each combination of
constraints on the number of peaks in the estimate. There are 5 rats with 7 branch com-
binations, for a total of 35 cases to examine (42 if we include Rat 3). To compare two
combinations of constraint, we perform a one-way ANOVA test comparing the two sets of
trials for each of the 35 cases. Out of these 35 results, we count how many are significant (p
< 0.05) and, of those, the number of cases in which each combination of constraints had a
higher mean number of peaks. The results are summarized in Table 5.2, in which the terms
“idealized” and “nerve-specific” refer to the leadfield that was used and “constrained” and
“unconstrained” refer to whether or not the spatio-temporal constraint was applied. As per
the discussion above, the entries in the table include a second number in parentheses that
corresponds to the results when Rat 3 is included.
The results in lines 1 and 2 of the table allow us to gauge the effects of adding the
spatio-temporal constraint, and show that, among the cases that were statistically significant,
Validation of the Source Localization Approach on Physiological Data 113
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 1
Pea
ks
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 2
Pea
ks
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 3
Pea
ks
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 4
Pea
ks
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 5
Pea
ks
T P S TP TS PS TPS0
1
2
3
4
5
6
7Rat 6
Pea
ks
Idealized/Unconstrained Idealized/Constrained Nerve−specific/Unconstrained Nerve−specific/Constrained
Figure 5.10: Mean number of peaks detected in the estimates, for all combinations of
branches and constraints, in each rat. The abbreviations are as follows: Tibial (T), Peroneal
(P), Sural (S), Tibial and Peroneal (TP), Tibial and Sural (TS), Peroneal and Sural (PS),
and Tibial, Peroneal, and Sural (TPS).
Validation of the Source Localization Approach on Physiological Data 114
Table 5.2: Influence of constraints on the number of peaks. The column “A > B” provides
the number of cases in which the combination of constraints in column A had a higher mean
number of peaks than the combination in column B, and analogously for “A < B”.
Line Combination Combination A > B A < B No significant
A B difference
1 Idealized/ Idealized/ 21(26) 0 14(16)
Unconstrained Constrained
2 Nerve-specific/ Nerve-specific/ 12 1(8) 22
Unconstrained Constrained
3 Idealized/ Nerve-specific/ 16 14(18) 5(8)
Unconstrained Unconstrained
4 Idealized/ Nerve-specific/ 9 13(20) 13
Constrained Constrained
5 Idealized/ Nerve-specific/ 18 13(20) 4
Unconstrained Constrained
6 Idealized/ Nerve-specific/ 7 16(22) 12(13)
Constrained Unconstrained
the constraint had a clear tendency to reduce the number of peaks. This is consistent with the
results of the simulation study. In contrast, the effect of using the nerve-specific rather than
idealized leadfield (lines 3 and 4 of the table) was much smaller, and no clear tendency could
be identified one way or the other. This result differs from the finding in the simulation
study, but is not surprising given that the nerve-specific leadfield is likely to still be no
more than a coarse approximation of the true anatomy. In the simulation study, we had
access to an exact model. When comparing cases “Idealized/Unconstrained” and “Nerve-
specific/Constrained” (line 5), between which the simulations had predicted the difference to
be greatest, we find only a slight tendency towards a reduction of the number of peaks by the
Validation of the Source Localization Approach on Physiological Data 115
constraints (and only if Rat 3 is excluded). This discrepancy is not surprising given the results
for the individual effects of the constraints: the brunt of the reduction is produced by the
spatio-temporal constraint, whereas the use of the nerve-specific leadfield is just as likely to
produce an increase as a decrease. Line 6 shows that there were fewer peaks when using only
the spatio-temporal constraint than when using only the nerve-specific leadfield. Although
the simulations did not make any strong predictions in this respect, these observations are
consistent with the other results in Table 5.2.
In summary, then, an examination of the total number of peaks confirmed the pre-
dictions of the simulations regarding the influence of the spatio-temporal constraint, namely
that it can reduce the number of spurious pathways. The prediction that the use of the
nerve-specific leadfield would have a similar effect was not confirmed, but this is easily ex-
plained by the different accuracies of the models used in the simulations and experimental
localizations.
The remaining quantitative metrics that are available to use are the number of spuri-
ous and missed pathways. These cannot be computed in the same way as in the simulations,
where we had access to all the true pathway locations, so we must use a simplified definition
instead. The only information available to us in the experimental recordings is the num-
ber of branches that were stimulated in a given trial. Therefore, we define the number of
spurious pathways as the number of detected peaks in excess of the number of stimulated
branches. The number of missed pathways is defined analogously in cases where there are
fewer detected pathways than stimulated branches. The results are shown in Figure 5.11,
but keep in mind that this is simply a different way of visualizing the data already presented
in Figure 5.10 (i.e., Figure 5.11 is Figure 5.10 with the appropriate constant subtracted from
each column in the bar plot).
Figure 5.11 suggests that there were generally fewer spurious pathways in cases where
several branches were stimulated. Viewed another way, the total number of pathways was not
very different regardless of the number of stimulated branches, which is a further indication
that the source localization process was not a reliable reflection of the underlying activity.
Validation of the Source Localization Approach on Physiological Data 116
It may also be related to the fact, observed in Figure 5.8, that the estimates for combined
cases were sometimes very similar to the estimates for one of the underlying cases. The other
observation to emerge from Figure 5.11 is that there were many more spurious pathways than
missed pathways. This result is also consistent with our previous observations, because the
algorithm’s difficulty in identifying a clear dominant source of activity is reflective of the
several sources of error present in the recordings.
5.4 Discussion
The results of this chapter show that our attempts to apply a source localization algorithm to
MCC recordings did not succeed in producing a reliable and selective neural interface, which
is to say that the estimates obtained could not be used to correctly ascertain the number and
identity of the branches being stimulated. Difficulties encountered during the experiments
included slight movement of the cuff and a stimulation artifact. While these factors mean
that we are not testing the source localization under ideal conditions, they are representative
of realistic conditions (i.e. a moving nerve, and interference from bioelectric sources outside
the cuff), such that our results are indicative of the expected performance in practice. The
performance observed here was not unexpected given the results of the simulations in the
previous chapter. In light of those conclusions, the source localization experiments that were
described in this chapter were mainly intended, first, to validate the results of the simulations
in the previous chapter and, second, to provide insight into other practical factors that may
affect source localization in real-life situations but were not revealed by the simulations.
The experimental data validated the finding that the application of a constraint can
have a beneficial impact on performance; in the case of the spatio-temporal constraint this
change takes the form of a reduction in the number of spurious pathways. The use of a
nerve-specific leadfield was not observed to have a similar benefit. Although this is contrary
to the results of the simulations, it is easily explained by the fact that the anatomical model
constructed from the nerve sample cross-sections is not a perfect representation of the true
Validation of the Source Localization Approach on Physiological Data 117
nerve anatomy. In situations where the more complex model increased the number of peaks,
we may also attribute this effect in part to the more irregular anatomy and, in several cases,
clear separation between fascicles, which may have resulted in a more fractured estimate
and therefore more peaks. The “realistic” model used in the simulations still had a fairly
regular shape, and only a small separation between the fascicles, and may therefore have
avoided this problem. It is also possible that using an incorrect nerve-specific model does
more harm than good when compared to a more idealized model: a higher level of detail is
only beneficial if the detail is reasonably accurate. This hypothesis may explain why Rat
6, which has the simplest anatomy, produced the results that were the most in accordance
with the simulations (Figure 5.10). Even though the nerve-specific leadfield was as likely to
increase as to decrease the number of peaks in the estimate, it did have some minor benefit in
the form of decreased overlap between the estimates for single-fascicle cases. Unfortunately,
there was no indication that this decrease in overlap was combined with an increase in
localization accuracy, such that this cannot be considered a major improvement due to the
nerve-specific leadfield. In addition, this result is partly biased by the fact that the rotation
of contacts that was chosen for each rat was based in large part on its ability to minimize
overlap between single-fascicle estimates.
The process of applying the source localization approach to real data recorded from
a peripheral nerve did not identify any new factors, but it did demonstrate the large im-
pact of several issues on the performance. Chief among these was, as expected, the high
sensitivity of the performance on the accuracy of the measurements and the model. The
combination of electrode motion, stimulation artefact and model inaccuracy prevented the
proposed approach from providing an accurate representation of the neural activity. Another
issue emphasized by the experiments is that of estimating the position of the cuff contacts
with respect to the nerve. If their position is misjudged by more than a very small amount,
then this will constitute another source of error that will significantly degrade performance.
For the purposes of the experiments in this chapter, we chose to deal with this problem by
rotating the position of the contacts in an effort to find the best performance. This approach
Validation of the Source Localization Approach on Physiological Data 118
is sufficient for our current purposes, namely to evaluate the impact of using a nerve-specific
leadfield, as long as we accept the hypothesis that the most accurate rotation will yield the
best performance. Nonetheless, the problem remains that the position of the contacts can
be varied in a continuous manner, and so trying a finite set of rotations does not capture all
possibilities and may not produce the optimal performance. In addition, the method is too
time-consuming to be used in practice. Before source localization techniques can be applied
in the context of a neuroprosthetic system, a method of correctly estimating the relative
position of the cuff and the nerve will be needed. This could, for example, take the form
of techniques to image the area inside the cuff (e.g. using a variant of electrical impedance
tomography, as mentioned in Chapter 4). Using cuffs with non-symmetrical shapes and fixed
sizes may also be beneficial in this respect, unlike the round spiral cuff used here.
An additional consideration when applying the source localization to experimental
data is to what extent the representation of bioelectric activity as a group of dipolar sources
is appropriate. The simulations did not shed light on this issue because the distributed
dipolar model was used to generate the measurements in addition to being used in the
inverse problem. With experimental data, however, if the source model does not accurately
reflect the complexity of the bioelectric activity distribution in the nerve, performance may
well suffer.
Validation of the Source Localization Approach on Physiological Data 119
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 1
Mis
sed
S
purio
us
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 2
Mis
sed
S
purio
us
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 3
Mis
sed
S
purio
us
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 4
Mis
sed
S
purio
us
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 5
Mis
sed
S
purio
us
T P S TP TS PS TPS2
1
0
1
2
3
4
5
6Rat 6
Mis
sed
S
purio
us
Idealized/Unconstrained Idealized/Constrained Nerve−specific/Unconstrained Nerve−specific/Constrained
Figure 5.11: Spurious or missed pathways as measured by the number of peaks in the
estimates compared to the number of stimulated branches, for all combinations of branches
and constraints, in each rat. The abbreviations are the same as in Figure 5.10.
Chapter 6
Source Localization Using an
Experimentally-Derived Leadfield
6.1 Introduction
The results of the previous section show, first, that the source localization approach as applied
so far results in poor performance and, second, that one of the major factors accounting for
this limitation is the discrepancy between the actual anatomy of the nerve and the model used
to construct the leadfield. In this chapter, we investigate an alternative formulation of the
source localization process that does not rely on an a priori anatomical model. Instead, we
use recordings of a few trials of each single-branch case as a training set to build a collection
of known vectors, each vector consisting of the 56 values recorded at all the contacts at a
given time instant. A new “leadfield” is constructed from these vectors. The problem is then
to identify which combinations of these vectors can best account for new recordings, when
either single or multiple branches are active.
Constructing a leadfield out of experimentally observed patterns has both advantages
and disadvantages. On one hand, it does not depend on having an anatomical model of the
nerve, and therefore eliminates a potential source of errors. On the other hand, because the
120
Source Localization Using an Experimentally-Derived Leadfield 121
observed patterns are matched only to a fascicle rather than a more specific location within
the nerve, the solution to this new problem can identify no more than the combination of
active fascicles; sub-fascicle resolutions are not possible. Of course, reliable identification of
fascicle combinations would already be a significant step forward from the performance in
Chapter 5, and so is very much worth investigating.
6.2 Methods
6.2.1 Construction of the experimental leadfield
For the remainder of this chapter, we will use the term “experimental leadfield” to refer to
a collection of observed measurement vectors that is built using a training set and will be
used to classify future observations. This leadfield is a matrix in which each column is a
56-element vector corresponding to the measurements recorded from the MCC at a given
instant. The goal is to construct a set of such vectors that are sufficiently representative of
the activity of each fascicle to be able to correctly identify future recordings. The leadfield
constructed in this way is thus analogous to the leadfield in previous chapters with the
exceptions that:
• Each column corresponds to an instantaneous spatial pattern of activity produced by
an entire fascicle, instead of by a dipolar bioelectric source in a very specific position.
Note that each fascicle will be associated with several vectors, because different patterns
of activity will be produced as a CAP travels in that fascicle along the length of the
cuff.
• The leadfield is constructed using a training set of experimental observations, rather
than using an anatomical model of the nerve. It is therefore not sensitive to modeling
errors.
Source Localization Using an Experimentally-Derived Leadfield 122
The experimental leadfield is constructed using a training set that includes only ob-
servations of single-fascicle activity. Although multi-fascicle cases are of interest and are
included in the testing set, they are not included in the training set because, in order to
be useful, the system should be able to identify combinations of fascicles based only on its
knowledge of the single-fascicle base cases. This principle is analogous to including in a
traditional leadfield only vectors corresponding to single-dipole cases, rather than explicitly
including in the leadfield all possible combinations of dipoles.
The process of constructing the experimental leadfield is as follows:
1. In each trial, identify bad channels in the same way as in Chapter 5 and set the data
for those channels to 0. Afterwards, convert the data to a common-average reference.
2. Divide the observed trials into a training set and a testing set, for each of the seven
combinations of fascicles (see Chapter 5 for details on the data collection procedure).
In the case of the multi-fascicle combinations, all trials belong to the testing set, as
explained above. In each single-fascicle case, the trials are divided into 5 groups, and
the performance will be measured using 5-fold cross-validation. Accordingly, the multi-
fascicle performance will be evaluated 5 times, using a different training set each time
but always the same testing set (i.e., all multi-fascicle trials).
3. Examine the training set of each single-fascicle case and record the measurement vectors
that occur within it. The goal is to build a set of distinct vectors. Therefore, each new
vector must be examined against all the previously seen vectors in order to determine
whether it is a new pattern or one that has already been recorded. Ideally, each
trial should produce exactly the same activity, so that a small number of patterns
(corresponding to the different positions of the CAP along the cuff) should occur many
times. In practice, this may not be the case, due to noise and the slight movement of
the cuff.
For each trial, the time instants examined are those in the same interval used for the
Source Localization Using an Experimentally-Derived Leadfield 123
source localization in the previous chapter. A collection of vectors is built using these
time instants from all the trials corresponding to a given fascicle. From this collection
of vectors we identify recurring patterns as follows.
(a) Normalize each vector with respect to its entry with the largest absolute value.
(b) Compare each new vector to all previously observed ones. Two vectors are deemed
to be the same if the l2 norm of their difference is less than 30% of the norm of
the first vector.
(c) If a vector has not been observed before, record it and move on to the next one.
Otherwise, the current vector must be incorporated into our collection of known
vectors. For the purpose of future comparisons, each vector is represented by the
mean of all its occurrences. For example, the second time a vector is seen, both
occurrences are replaced by a single entry that is the mean of the two. When a
third vector is found that resembles that mean, the entry is replaced by the mean
of all three vectors, etc... The number of times that a vector has been observed
is also recorded.
4. Once all vectors have been examined, any vector that occurred only once is deemed
not to be useful for identifying future observations and therefore is removed.
5. The remaining vectors are gathered to form the columns of the experimental leadfield.
The leadfield contains the vectors identified for all of the single-fascicle cases, but we
keep track of which columns correspond to which fascicle.
6.2.2 Identification of fascicle combinations
Once the leadfield has been constructed according to the procedure in the previous section,
two situations can arise: the problem can be either underdetermined or overdetermined. The
far more likely of the two, and the one dealt with here, is that of an underdetermined problem:
Source Localization Using an Experimentally-Derived Leadfield 124
the number of measurements is much smaller than the number of different patterns observed,
so that the leadfield has more columns than rows. This is the same type of situation that
we faced when using the model-based leadfield. Many of the principles mentioned in that
context for solving an ill-posed inverse problem can therefore be applied here as well. The
sLORETA algorithm, however, is less meaningful here because the variables now represent
more abstract entities than FE mesh elements, and lack clear spatial relationships between
them. The concept of “smoothness” that is central to sLORETA is therefore no longer
relevant. We turn instead to the simpler weighted minimum-norm approach (Section 2.2.3).
This method does not have a geometrical assumption underlying it and simply solves the
inverse problem in the simplest way (finding the minimum-norm least-square solution), with
only the added complexity of compensating for the different norms of the leadfield columns.
Examining the problem more closely, it becomes apparent that a typical solution should
consist of only a small number of non-zero variables. Indeed, a CAP in a given fascicle
should ideally correspond to a single column of the experimental leadfield. Even considering
a complex case of several active fascicles, some of which may contain two or three CAPs at
different positions, the total number of variable will still be fairly small. Accordingly, it is
reasonable to choose an inverse problem method that produces sparse solutions. We choose
for this purpose to use FOCUSS (Section 2.2.3), using as the initial estimate the weighted
minimum-norm estimate. Both methods are regularized using Tikhonov regularization, with
the regularization parameter chosen using the L-curve method (Section 2.2.5).
6.2.3 Evaluation of the results
Each trial was evaluated as follows. First, the estimate for each time instant was normalized
with respect to its entry with the largest absolute value. Second, for each variable in the
experimental leadfield, the mean estimated activity over the time interval was computed.
Next, for each of the three branches, the mean of these results over all the variables corre-
sponding to that branch were computed. This resulted in a set of three numbers for each
Source Localization Using an Experimentally-Derived Leadfield 125
trial. The absolute values were taken and the three numbers normalized with respect to the
highest one. The final product was a set of three values between 0 and 1 describing the
relative estimated activations of the three branches during this trial.
6.3 Results
Excluding Rat 3, the number of columns in the 25 experimental leadfields (5-fold cross-
validation in each of 5 rats) ranged from 176 to 384. This shows that there was a high
level of variability in the measurements: there are three fascicles being used to construct the
leadfield, and given the sampling rate and approximate conduction velocity, it would take on
the order of 10 time samples for a CAP to propagate through the cuff. In the ideal situation,
therefore, there would be only about 30 columns in the experimental leadfield (creating
a overdetermined problem and calling for different methods to solve it). Furthermore, as
described above, these columns are those that remain after patterns that were observed only
once are eliminated, such that the total number of observed vectors was even higher. On the
other hand, the highest number of repetitions for a given pattern in each rat varied from 31
to 127, showing that there was still a measure of regularity to be found in the recordings.
Figure 6.1 shows the mean of the three activity estimates for each branch combination
and each rat. These means are taken on the agglomeration of the results in all 5 testing
sets. Examination of the figure reveals that in the single-branch cases the algorithm was
successful in identifying the stimulated branch as by far the most active. Activity estimates
of the other branches were in most cases small, but not insignificant. As for the multi-branch
cases, the algorithm was less successful in identifying the active branches. Although a few
cases were close to being accurate (e.g. Rat 1, tibial + peroneal and tibial + sural), inactive
branch activities estimates were still high, and on the whole the method was not reliable.
Figure 6.1 gives a useful overview of the algorithm’s ability to assess the activity of
the different branches. From the point of view of a neuroprosthetic system, however, we are
interested in knowing how often the algorithm can correctly identify the exact combination of
Source Localization Using an Experimentally-Derived Leadfield 126
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 1
Bra
nch
Act
ivity
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 2
Bra
nch
Act
ivity
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 3
Bra
nch
Act
ivity
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 4
Bra
nch
Act
ivity
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 5
Bra
nch
Act
ivity
Tibial Peroneal Sural
T P S TP TS PS TPS−0.2
0
0.2
0.4
0.6
0.8
1
1.2Rat 6
Bra
nch
Act
ivity
Figure 6.1: Means of the activity estimates for the three branches, for each rat and branch
combination. The abbreviations are the same as in Figure 5.10.
active branches. In order to measure this aspect of the performance, the activity estimates
computed above were thresholded at 0.2. A branch is deemed active if it is above this
threshold, and inactive otherwise. We then computed the percentage of trials in which the
combination of active branches is exactly accurate, for each branch combination and rat. The
success rates were averaged across the 5 testing sets, and the results are shown in Figure 6.2.
Figure 6.2 is in accordance with Figure 6.1 in showing that the algorithm was more
successful at correctly identifying single-branch cases than multiple-branch ones. Excluding
Rat 3, the mean success rate over the 15 single-branch cases was 68.5%, with a minimum
Source Localization Using an Experimentally-Derived Leadfield 127
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 1
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 2
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 3
T P S TP TS PS TPS
0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 4
T P S TP TS PS TPS
0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 5
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 6
Figure 6.2: Success rate for identifying the exact combination of active branches, for each
rat and branch combination. The standard deviation is based on the 5 repetitions of the
cross-validation process. The abbreviations are the same as in Figure 5.10. A branch is
considered active if its activity estimate is greater than 0.2.
Source Localization Using an Experimentally-Derived Leadfield 128
of 19.4% and a maximum of 95%. On the other hand, the mean success rate over the 20
multiple-branch cases was 25.3%, with a minimum of 1.3% and a maximum of 54.6%. Note
that by consulting Figure 6.1, we can see that single-fascicle cases with low performance
in Figure 6.2 were mostly due to inactive branches erroneously being identified as active,
rather than the correct active branch being missed. The exact proportions of false positives
and false negatives of course depends on the threshold that we use. The use of 0.2 was an
attempt to balance the need of the single- and multi-fascicle cases. If the performance of
the multi-fascicles cases was of no interest, it is clear from Figure 6.1 that the number of
false positives in the single-fascicle cases could be reduced by raising the threshold. This is
illustrated in Figure 6.3, which is obtained in the same way as Figure 6.2 except with the
threshold set to 0.6 instead of 0.2. This change of threshold raises the mean success rate
of the single-branch cases to 89.8%, and lowers that of the multiple-branch cases to 11.0%.
Rat 5, which had the poorest performance, was also the rat in which the least repetition
was found among the observed vectors in the training set, illustrating the importance of a
reliable training set when interpreting cases in the testing set.
6.4 Discussion
In this chapter we investigated a potential solution to the issue of anatomical model depen-
dency in the bioelectric source localization process. By constructing a collection of measure-
ment patterns across the 56 contacts, observed during a few single-fascicle training trials,
we were able to create a leadfield derived entirely from experimental data, rather than an
a priori anatomical model. In doing so, we eliminated one source of error (model inaccu-
racies), at the expense of abandoning the possibility of sub-fascicular resolution. Given the
performance described in the previous chapter, this seemed like a very worthwhile trade-off.
The results using the experimental leadfield were significantly improved in the single-
fascicle cases. On the other hand, success rates for correctly identifying multi-fascicle cases
were still very low. Discriminating the activity of three fascicles when only one is active can-
Source Localization Using an Experimentally-Derived Leadfield 129
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 1
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 2
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 3
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 4
T P S TP TS PS TPS
0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 5
T P S TP TS PS TPS0
20
40
60
80
100
Suc
cess
Rat
e (%
)
Rat 6
Figure 6.3: Success rate for identifying the exact combination of active branches, for each
rat and branch combination, when the threshold for considering a branch active is raised to
0.6.
Source Localization Using an Experimentally-Derived Leadfield 130
not be considered a novel accomplishment considering the results reported in the literature
(as reviewed in Section 2.3.3), but the method proposed here is a novel way of addressing
this task. This line of enquiry is therefore worth pursuing further: casting the issue as an
inverse problem of source localization provides a rich mathematical framework and opens
the door to a large number of techniques. This large amount of existing theory was one of
the reasons behind our initial decision to apply source localization techniques to peripheral
nerves, and it remains just as valid for the experimental leadfield as for the model-based
leadfield.
Of course, one of the benefits of using sophisticated source localization algorithms is
the possibility of dealing with multiple simultaneous sources. This begs the question of why
the experimental leadfield approach was unable to satisfactorily deal with the multi-fascicle
cases. Despite having eliminated the issue of model inaccuracies, there remain other sources
of error, such as interference from the artefact. The movement of the cuff may have also
had led to a situation where performance degraded more with longer periods of time: in
the single-fascicle cases, the training and testing cases were obtained in short succession in
a single series of trials. Multi-fascicle cases were usually obtained a few minutes later, such
that more cuff movement may have occurred, making it more difficult for the algorithm to
“recognize” the base cases that it had been trained on.
In addition to practical issues of this sort, we must remember that the experimental
leadfield does not change the fact that we are dealing with an ill-posed inverse problem with
an infinite number of solutions and poor stability in the face of measurement errors. From
that point of view, the results of this chapter are indicative of the issues that may arise when
several sources are present, even if simple localization is possible. The previous chapter had
poor performance even in the single-fascicle cases, and therefore was less informative in terms
of the differences between single- and multi-fascicle situations. Interference from the artefact
may account for much of the performance gap between these two situations when using the
experimental leadfield, because measurement error will make successful reconstruction of
complex source configurations more difficult than reconstruction of simpler ones. Nonethe-
Source Localization Using an Experimentally-Derived Leadfield 131
less, since the performance of single-fascicle cases was fairly high despite the artefact, we
can conjecture that the poor performance of the experimental-leadfield approach for multi-
fascicle cases is partly the result of fundamental issues in this type of problem, such as the
fact that different combinations of sources can produce the same measurements. It may be
that the constraints and regularization techniques that can best improve the performance in
this problem are different from the ones that are most effective with model-based leadfields.
It is also possible that interactions within the nerve, in particular electric field in-
teractions between fibers, may invalidate the assumptions that the recordings from branch
combinations will be strictly linear summations of single-branch recordings. In that case, it
could be helpful to incorporate system identification methods producing input-output maps
that account for nonlinear behaviour [93].
A disadvantage of the experimental leadfield compare to the model-based leadfield
is the necessity of obtaining training data. We sought to limit this disadvantage by ob-
taining training data only from the single-fascicle cases. It is likely that the necessity of
obtaining this data would not be an insurmountable obstacle to applying this technique in
practice: the training data could be obtained fairly rapidly intra-operatively when the cuff
is implanted, assuming that it is possible to individually stimulate the fascicles of interest,
either directly or through indirect methods such as cutaneous stimulation or passive limb
movements. Additionally, the model-based leadfield’s advantage in this respect is only signif-
icant if no calibration of the anatomical model is required. If it becomes necessary to collect
various pieces of data to refine the model (e.g. ultrasound imaging of the nerve), then the
practicality of both methods becomes comparable. A related issue is that it may be possible
to obtain training data for only a few of the fascicles in a nerve, a situation that would
make the remaining fascicles unidentifiable and in effect impose a different type of resolution
limit on this approach to source localization. The elements of the experimental leadfield
also do not need to correspond to specific nerve branches, but may instead be associated
with more “high-level” functional concepts (e.g. a particular movement in a particular limb,
which may produce activity in several pathways). This alternative approach would modify
Source Localization Using an Experimentally-Derived Leadfield 132
the stimulation protocol needed to construct the experimental leadfield.
Chapter 7
Influence of the Number and Location
of Recording Contacts on the
Selectivity of a Nerve Cuff Electrode
7.1 Introduction
In the face of the poor performance of the source localization approach, using both model-
based and, to a lesser extent, experimental leadfields, we take a step back and try to deter-
mine whether the MCC genuinely contains more information than simpler cuff configurations
used in the literature previously. To this end, in this chapter we use the simpler problem
of discriminating the three possible types of single-fascicle activity, when a training set is
available. Our goal is to evaluate the improvement in fascicle discrimination that can be
achieved with the matrix cuff compared to a set of contacts laid out in a single ring, as in
previously employed multi-contact tripole configurations [197]. To investigate this issue, we
use the matrix cuff and compare the performance of the full grid-like contact configuration
(the 56-contact “matrix” configuration) to the performance when using only the 8 contacts
in the middle ring of the cuff (the “single-ring” configuration). If the matrix performance is
133
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 134
found to be superior, we will seek to determine in addition whether the improvement is due
to the number of contacts used or to their position (or both). The large number of contacts
and grid layout of the matrix cuff makes it ideal to study these issues. In addition to the
source localization considerations, information about the optimal placement of contacts in
a nerve cuff to maximize selectivity would have direct applications in the design and use of
this type of electrode in neuroprosthetic systems.
7.2 Methods
The analysis in this chapter was performed on the data collected as described in Chapter 5,
with the exception that only the single-fascicle cases were used. The methodology described
in the remainder of this section therefore pertains to the processing of the signals.
7.2.1 Evaluation of the classification success rate
We sought to determine whether or not the recordings from the 56-contact matrix cuff gen-
uinely contained more useful information than measurements from a simpler configuration.
We considered the case of a simple feature-based classifier for differentiating the activity of
the three different fascicles, when only one of them is active at a time. In the context of our
experiments this means that our goal is to determine which fascicle was being stimulated in a
given randomly chosen trial, using the measurements from the recording cuff. We compared
the performance of this classifier when using data from all 56 contacts to the performance
when using only data from the 8 contacts in the middle ring of the cuff (ring 4 of 7). The
two configurations are illustrated in Figure 7.1.
The classification process was conducted as follows for each animal:
1. For each trial, the data was converted to a “tripole” reference, which is to say that
the average of all the contacts in the first and last rings was used as the reference (the
term tripole is used loosely here, since there are more than three rings in the cuff).
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 135
Figure 7.1: Contact configurations for the matrix (a) and single-ring (b) cases. The con-
tacts in dark gray are the ones that are available for use in the classification process. The
contacts of the first and last ring are averaged to produce the reference, for both configuration.
Once this was done, the data was normalized using the largest absolute value in this
trial over all contacts. Because of this normalization, the classification is based on
the distribution of activity among the contacts, and therefore on spatial information,
rather than on the magnitude of the activity.
2. A set S of contacts to be included in the feature vector was defined.
3. For each trial, the peak of the action potential recorded at each of the contacts in S
was found (the peaks may not all occur at the same time, since the contacts can be at
different longitudinal positions along the cuff). The feature vector was then defined as
the potential of each contact at its peak, resulting in a vector with one entry for each
element in S.
4. The trials from each nerve were partitioned into a training set and a testing set. The
feature vectors from the training set were averaged for each fascicle, resulting in one
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 136
mean feature vector for each of the tibial, peroneal, and sural branches. Each of these
mean feature vectors was normalized. The three vectors were then collected into a
matrix L.
5. For each trial in the combined testing sets, the normalized feature vector F was classi-
fied by finding the least-squares solution to the overdetermined system LX = F . The
fascicle corresponding to the largest value in X was chosen as the one responsible for
the observed activity in this trial.
6. The classification success rate is the percentage of trials in the testing set that are
assigned to the correct fascicle.
In order to ensure that the results were not biased by the choice of trials included in
the training set, the evaluation of the classifier was performed using 10-fold cross-validation.
In each trial, channels with excessive variance or very small amplitude compared to
the other channels were marked as bad channels and set to 0 before computing the feature
vector. Trials were discarded when more than a quarter of the channels in S were bad or
when the temporal spread of the peaks across all contacts was greater than 1 ms.
Our main concern is whether or not the matrix cuff allows for more accurate classi-
fication than the single-ring configuration. In addition, we would like to establish whether
or not all 56 contacts are needed for an improvement (if any is found). In other words, we
would like to know if the benefit of the matrix cuff stems from having more channels of
information, or if a small subset of contacts could also lead to better performance simply by
virtue of having 56 possible contacts to choose from instead of 8. To answer these questions,
both configurations were investigated by adding one contact at a time and tracking the per-
formance as more contacts were added. The set of available contacts during this process
was in one case all 56 contacts, and in the other case the 8 contacts in the middle ring
(refer once again to Figure 7.1). At each step, the contact added was the one that improved
the performance the most. In other words, we first computed the performance using each
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 137
contact individually (i.e. S had a single element, and the full cross-validation procedure
was performed) and retained the best one. Next, we investigated each remaining contact
in combination with the first contact selected, and again retained the best one. The third
contact was then combined with the first two, and so on, until all the contacts from the set
of interest had been added.
7.2.2 Evaluation of the influence of the stimulation artefact
The interpretation of the results will be complicated by the presence of the large stimulation
artefact in the recordings, as described in Section 5.2.1. Figure 7.2 illustrates this with an
example of one trial, showing both the raw data and the data after conversion to the tripole
reference. This figure is very similar to Figure 5.3, with the exception that the latter used a
common-average reference rather than a tripole reference. The tripole reference was used in
this part of the study for consistency with the single-ring configuration. We must consider the
possibility that stimulation at different sites produces slightly different stimulation artefacts,
and that the classifier is partly taking advantage of this information. If this were the case,
we would expect that the classification success rate would be superior when large artefacts
are present. In order to investigate this possibility, we use the fact that the magnitude of
the artefact is expected to vary between rings of contacts. Indeed, theoretically, the electric
field produced by sources outside the cuff should vary linearly along the length of the cuff
[170, 132, 133, 2]. By examining the magnitude of the signals recorded at each contact before
converting the data to the tripole reference, we can estimate how the size of the artefact varies
between rings. This information can then be converted to an estimate of how the artefact
will vary between rings after the tripole reference is applied. Lastly, to determine whether
the classifier is relying heavily on information in the artefacts, we compute the classification
success rate using each ring in turn as the set S described above. If the classification uses
the artefact, we expect that the performance using the different rings will be correlated with
the estimated size of the artefact at those rings.
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 138
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V m
V)
0 0.5 1 1.5−2
−1
0
1
2
t (ms)
V (
mV
)
Rat 1, Tibial Nerve, Trial 1
Figure 7.2: Example of the recordings of one tibial branch trial in Rat 1. The upper left-
hand plot shows the raw recordings for all 56 channels. The upper right-hand plot shows
those same recordings after conversion to a tripole reference. The lower plots show the same
data for one contact only, taken from the middle ring of the cuff.
To estimate the artefact variations, the recordings of each contact are first averaged
over all trials of all three fascicles combined, then rectified and integrated. The signals used
in this step are the raw measurements, recorded with respect to the outside contact rather
than using the tripole reference. The size of the artefact at each ring is estimated using the
average of the obtained values of each contact in the ring. The resulting set of seven values
(one per ring) is normalized using the largest value. By subtracting the mean of the first
and last values and taking the absolute values of the results (to take into account the tripole
reference in the classification), an estimate is obtained of how the classification performance
would be expected to vary from ring to ring if the size of the artefact was the determining
factor. Lastly, the correlation between this series and the performance actually obtained is
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 139
computed.
7.3 Results
7.3.1 Comparison of the matrix and single-ring configurations
Figure 7.3 shows the maximum success rate achieved for each animal using each method.
These results demonstrate that in all cases a better classification success rate was achieved
using the matrix configuration. The question now is whether the improvement is due sim-
ply to the sheer number of contacts. To resolve this issue, the success rate was computed
using the first eight selected contacts of the matrix configuration, versus the eight contacts
of the single-ring configuration. The results are shown in Figure 7.4, and once again the
matrix configuration results in clear improvements. For each comparison, an ANOVA test
was conducted using the 10 results of the cross-validation procedure for each contact config-
uration. The differences in Figures 7.3 and 7.4 were shown to all be significant (p < 0.05),
with the exception of the rat 4 comparison in Figure 7.4, although the matrix configuration’s
performance was still higher in that case.
Figure 7.5 plots the classification success rate of the matrix configuration as a function
of the number of contacts for each rat. Markers on each plot indicate the point at which
maximum performance is achieved, and the point at which the success rate exceeds the
maximum success rate achieved with the single-ring configuration. As an example of the
contact selection process, Figure 7.6 shows the order in which the contacts were selected in
the case of Rat 1 for each of the two configurations, up to the number of contacts at which
the maximum success rate is reached (refer to Figure 7.5).
Several conclusions can be drawn from this data. First, discrimination of the activity
of different fascicles is feasible, which confirms the information found in the literature [197,
14, 176] and the results of Chapter 6. Second, the use of the matrix cuff can significantly
improve the classification success rate. Lastly, optimal or near-optimal performance can be
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 140
Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 60
10
20
30
40
50
60
70
80
90
100
110Maximum Classification Success Rates of Matrix vs. Single−Ring Configurations
Suc
cess
Rat
e (%
)
MatrixSingle−ring
*
*
*
*
*
*
Figure 7.3: Maximum classification success rate achieved with the matrix and single-ring
configurations, for each rat. The standard deviations shown are based on the set of 10 results
obtained for each case during the 10-fold cross-validation process. The asterisk denotes a
statistically significant difference (p < 0.05).
achieved with fewer than 10 contacts. This implies that the superior performance of the
matrix cuff is not due to the absolute number of contacts, but rather to the possibility of
sampling the extracellular fields in locations that contain the most useful information. These
results were consistent across all of the animals, but the maximum classification success rate
varied widely, with a range of 83.9% to 100%. In addition, the contacts selected as providing
the most information were not consistent between animals. These variations could be due
to a number of factors, mainly related to the position of the cuff on the nerve, the quality
of the electrical connection established at each contact, and noise issues.
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 141
Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 60
20
40
60
80
100
120
Classification Success Rates Using the First 8 Contacts ofMatrix vs. Single−Ring Configurations
Suc
cess
Rat
e (%
)
Matrix (first 8 contacts)Single−ring (all 8 contacts)
*
*
*
*
*
Figure 7.4: Classification success rate achieved with the first 8 selected contacts of the
matrix and of the single-ring configurations, for each rat. The standard deviations shown
are based on the set of 10 results obtained for each case during the 10-fold cross-validation
process. The asterisk denotes a statistically significant difference (p < 0.05).
7.3.2 Influence of the stimulation artefact
An example of the comparison described in Section 7.2.2 is shown in Figure 7.7. The esti-
mated normalized artefact distribution is shown, as well as the expected performance vari-
ations if the classification was based on the artefact, and the actual performance variations
observed. The correlation between the expected and observed variations was 0.07, which
corresponds to a p-value of 0.88 when considering a null hypothesis of no correlation. Table
7.1 shows the correlations and p-values for all six animals. The p-value was considerably
larger than 0.05 in five of the six cases, such that we cannot conclude that the stimulation
artefact plays a significant role in the classification success rate. Although these results
do not allow us to state that the stimulation artefacts have absolutely no influence on the
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 142
0 10 20 30 40 50 6030
40
50
60
70
80
90
100
Classification Success Rate as a Function of the Number of Contacts Added
Number of Contacts
Suc
cess
Rat
e (%
)
Rat 1Rat 2Rat 3Rat 4Rat 5Rat 6
Figure 7.5: Classification success rate achieved using the matrix configuration, as a func-
tion of the number of contacts used. The first markers (o) indicate the point at which the ma-
trix configuration starts outperforming the maximum success rate achievable with the single-
ring configuration. The second markers (X) indicate the maximum success rate achieved with
the matrix configuration.
performance, they do show that the artefacts are not the dominant factor, and that the
comparisons between the different contact configurations are based on neural activity. The
case of Rat 3, in which p ≤ 0.05, suggests that there may have been an incomplete closure of
the cuff in that experiment, leading to a much more predominant stimulation artefact in the
recordings. All results for this rat should therefore be treated with caution, as mentioned
earlier in Chapter 5.
It should be mentioned that although Figure 7.7 shows a roughly linear variation of
the artefact, as expected, this pattern was not observed in all of the animals. The lack
of linearity in the other animals can be attributed to variations in the impedances of the
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 143
Figure 7.6: a) Order in which contacts were added in Rat 1 when using the matrix con-
figuration. Only the first 7 contacts are shown because that is the number required to reach
the maximum success rate for this animal (see Figure 7.5). b) Corresponding results when
using the single-ring configuration. In the case the maximum success rate was reached with
6 contacts.
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 144
Table 7.1: Correlation of the artefact and classification success rate variations between
contact rings.
Animal Correlation P-value
Rat1 0.07 0.880
Rat2 0.08 0.861
Rat3 0.85 0.015
Rat4 -0.57 0.177
Rat5 0.37 0.420
Rat6 0.33 0.475
contacts, as well as to small shifts in the cuff position during the experiments (recall that
the artefacts are estimated using an average of all the trials for a given animal). It is
for this reason that we examined the correlation between the variations in artefact and in
classification success rate, rather than checking for a pre-determined pattern in the success
rate variations.
Computing the classification success rates for every ring of contacts also allowed us
to confirm that none of them outperformed the success rate obtained using the matrix
configuration (results not shown).
7.3.3 Layout of the most informative contacts
Having established that a small number of contacts can be used to achieve high performance,
the question arises of whether it is possible to identify the optimal contacts, and potentially
incorporate this information in future cuff designs. We therefore examined the order in which
the contacts of the matrix configuration were selected, in other words which contacts proved
most informative for the purposes of fascicle classification accuracy.
Establishing a common set of useful contacts among all rats proved difficult, which is
not surprising given that the alignment of the contacts with the fascicles was not the same
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 145
from one experiment to the next. Nonetheless, it was observed that in all cases the selection
algorithm started by choosing contacts from several different radial positions along the nerve
(not necessarily aligned at the same longitudinal position). The number of contacts that were
added before any repetition of the radial position occurred varied between 4 and 6, with an
average of 5.33 over the six animals. This result indicates that there is value in sampling
different radial positions around the nerve, which of course is to be expected because it
allows different contacts to be close to different fascicles. The fact that the different radial
positions selected were not necessarily aligned longitudinally is also very important, because
it illustrates the value of having more than one contact to choose from when attempting
to record from a given fascicle. These observations are well illustrated by the Rat 1 results
shown in Figure 7.6.
7.4 Discussion
We demonstrated that by using a matrix cuff electrode it was possible to obtain a better
fascicle classification success rate than when using signals only from contacts in the middle
ring of the cuff. We further showed that the difference was not due to the sheer number of
contacts, since the matrix cuff could outperform the single-ring configuration even with a
small number of contacts. These results are in accordance with expectations, because they
support the idea that classification success rate can be improved by selecting the locations
around the nerve that contain the most useful information.
The locations of the most useful contacts cannot necessarily be determined a priori,
because they will not depend only on the locations of the fascicles. Rather, variations in
the impedances of the cuff contacts, the details of the interface of each one with the nerve
(i.e. distance, amount of interfering tissue, etc.) and the noise level are likely to play a large
role. In addition, even if the approximate placement of the fascicles can be estimated, their
relative positions will not be completely constant along the length of the cuff, particularly
if the device is long. As a result, how the selectivity will vary with the longitudinal position
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 146
is not known in advance. Furthermore, the optimal number of contacts will depend on the
number of fascicles that we are attempting to discriminate in a given nerve. Because of
these issues, the results presented here cannot be used to design a cuff with a small number
of optimally-placed contacts. Rather, they argue in favor of implanting a device with a
large number of contacts, then conducting an optimization procedure that will indicate
which subset of the contacts should actually be used. Having a large initial set of contacts
available is all the more beneficial when one considers the issue of chronic implantation.
With time, morphological changes will occur, in the form of connective tissue accumulation
and reshaping of the nerve itself [112, 46, 138, 139]. The optimal subset of contacts may
therefore not be constant. If the contact selection procedure could be conducted not only
during the initial implantation but on a regular basis, the nerve cuff’s performance could be
maintained at a higher level over time. Another aspect to this issue is illustrated by Figure
7.5, which shows that the accuracy not only can be maximized with a modest number of
contacts, but can actually decrease when too many contacts are added. We can hypothesize
that certain contacts contain very little classification information, either because of their
position or because of high impedance or noise. Including such contacts in the classification
procedure could therefore cause more confusion than improvement. This phenomenon argues
in favour of having a contact selection procedure regardless of the amount of information
bandwidth that can be accommodated.
The main limitation of our study is the presence of the stimulation artefact, which
casts doubt on the exact classification success rate that could be achieved in its absence.
Nonetheless, we have shown that its impact was limited. Similarly, the unintended variations
in the stimulus pulses (as described in the Data Collection section of Chapter 5) likewise
raise the possibility that the classification was partially based not on the spatial position
of the fascicles, but on differences in the neural activity generated in each one. The fact
that the data in each trial is normalized (as described in the Methods section) helps to
compensate for possible differences of this kind. Furthermore, the doubt created by this
issue pertains more to the actual values of the success rate achieved than to the difference
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 147
between the matrix and single-ring configurations, and as such has little bearing on the main
conclusion of our study (i.e. the benefit of choosing among numerous contact locations).
Another small but important limitation is that the algorithm that we used to select the
best contacts had the benefit of simplicity but was not necessarily optimal. Indeed, the
contacts were selected one at a time, rather than exploring the entire space of possible
contact configurations, which would have been computationally prohibitive. If different
contact selection algorithms were explored, they would most likely have some impact on
the classification success rates. Nonetheless, the simple algorithm was quite sufficient for
demonstrating that the matrix configuration was beneficial and that only a small number
of contacts was needed. A more significant drawback is that the results in this study are
based on recordings of compound action potentials, rather than spontaneous activity. The
larger amplitudes of these signals were helpful in establishing clear measurement patterns
corresponding to each nerve, achieving successful classification, and evaluating with greater
certainty the influence of the number and location of the contacts. The smaller signal-
to-noise ratio that can be expected in certain types of natural activity (e.g. [140]) would
likely result in poorer classification performance. Once again, however, this limitation does
not invalidate our conclusions regarding the varying usefulness of different contacts and the
benefits of carrying out a selection procedure.
Lastly, it is important to keep in mind that the findings described in this study deal
with a reasonably simple case, specifically the identification of the active fascicle when only
one fascicle is active and a training set is available. The more complex case of identifying
combinations of fascicles without a training set cannot be adequately handled with such
simple techniques, evoking the need for more complex methods such as source localization
algorithms. Nonetheless, the comparison of the matrix configuration with the single-ring
one has important practical applications. By using a matrix-type cuff and performing some
preliminary training recordings, it should be possible to improve the performance over current
devices while still using a small number of contacts, thereby avoiding wiring and power
consumption issues stemming from using large numbers of contacts (the combination of
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 148
nerve cuff electrodes with multiplexer circuits to access different contacts has previously been
explored in the literature [156]). Even when multiple fascicles are simultaneously active (as
will likely be the case in practice), the optimal number of contacts may not be the same for
all situations, but the contact selection method proposed here will still be useful by helping
to identify which contacts are most useful, by virtue of having a good interface with the
nerve and positions that allow them to discriminate among different fascicles.
Influence of the Number and Location of Recording Contacts on the Selectivity of a Nerve
Cuff Electrode 149
1 2 3 4 5 6 7
0.7
0.8
0.9
1Normalized Artefact
1 2 3 4 5 6 70
0.2
0.4Normalized Artefact (Tripole Reference)
1 2 3 4 5 6 760
70
80
90Classification Success Rate (%)
Figure 7.7: Comparison of artefact variations and classification success rate variations
between contact rings in Rat 1, as described in Section 7.2.2. The top plot shows the normal-
ized estimated variations of the stimulation artefact in the raw recordings, as a function of
the contact ring. The middle plot shows the variations once the use of the tripole reference
has been taken into account. The bottom plot shows the variations in the success rate as a
function of the contact ring, when each ring is used in turn to classify the recorded activity.
The middle and bottom plots are poorly correlated, arguing against the hypothesis that the
stimulation artefact plays an important role in the classification success rate.
Chapter 8
Discussion
8.1 Summary
This thesis was motivated by the lack of existing technology to monitor the electrical activity
of pathways within a peripheral nerve with a satisfactory combination of spatial specificity,
spatial coverage, minimal damage to the nerve, and applicability to different nerve sizes. We
sought to determine whether applying a bioelectric source localization approach to MCC
recordings could provide a useful trade-off between these different issues. Any attempt to
relate extraneural field measurements to the activity of specific locations in the nerve is
fundamentally a source localization problem. Applying to peripheral nerves the theory that
has been developed to address similar problems, most notably EEG source localization,
is therefore a logical step that had not yet been taken in the literature, partly because
cuff electrodes have only recently begun to incorporate sufficient recording contacts for the
approach to make sense. On the other hand, there are substantial obstacles to effectively
solving the source localization problem in practice. In particular, it is an ill-posed inverse
problem, a notoriously difficult type of situation to deal with. The problem requires an
accurate model of the region containing the sources, is sensitive to noise, and does not
have a unique solution, which creates the need for sensitive regularization techniques and
150
Discussion 151
hampers our ability to achieve very high spatial resolution. While these issues, particularly
the sensitivity to noise, may give the impression that this type of approach is poorly suited
to nerve cuff recordings, we must keep in mind that these obstacles do not stem from a
particular choice among several possible methods. Instead, they arise directly from the
nature of the task. In other words, we are not using a source localization approach because
of the characteristics of this method; we are using it because the task that we are addressing
is a source localization problem. On a more positive note, the approach can be applied to
instantaneous electric field recordings, and as a result has extremely high temporal resolution,
which is essential for correctly decoding neural activity.
The first issue we addressed was how to minimize the huge discrepancy between the
number of measurements and the number of unknowns in the source localization problem.
Our strategy was to reduce the redundancy present in the variables. Although a relatively
fine FEM model was used to solve the forward problem as accurately as possible, the mesh
elements were not all distinguishable from the point of view of the inverse problem. The
distinguishability of adjacent elements was quantified using comparisons of the corresponding
forward fields (leadfield columns), and elements that could not be discriminated from one
another were fused together, thereby reducing redundancy. In the idealized geometry used
for this part of the study, the number of variables was reduced by more than half through
this process. Although there was a minor improvement in numerical conditioning, the major
benefit was a reduction of computation times and storage requirements, important factors if
an on-line implementation is ever to be achieved. Furthermore, since the reduced leadfield
better reflects the inherent resolution of the problem, the proposed method has a secondary
application as a technique to analyze the difficulty of successful localization and establish an
upper bound on resolution for a given combination of nerve and electrode.
The next step was to conduct a simulation study to assess the potential of bioelectric
source localization in peripheral nerves. The influences of two factors were investigated:
using a leadfield based on an accurate versus an idealized model of the nerve anatomy, and
applying a spatio-temporal constraint based on the electrophysiology of myelinated fibers.
Discussion 152
Simulated measurements corresponding to either one or three pathways were generated using
the realistic nerve model, and the source localization performance was evaluated in terms
of localization error, spurious pathways, and missed pathways. The approach in its present
form was not found to be sufficiently reliable for sub-fascicular localization in practice, due
to mean localization errors in the 140 µm-180 µm range, high numbers of spurious pathways,
and low resolution. Nonetheless, the improved anatomical model and the spatio-temporal
constraint were shown to produce a marked reduction in the number of spurious pathways.
We concluded that methods by which the noise in nerve cuff recordings could be reduced
and anatomical information obtained in vivo should be investigated if the source localiza-
tion approach is to become viable. In the short term, it is more realistic to focus on reliably
identifying combinations of whole fascicles, rather than to seek to obtain sub-fascicular in-
formation.
Having analyzed the problem through simulations, we sought to validate our findings
through the collection of experimental data. The MCC was placed on the sciatic nerves of
rats, and recordings obtained while the tibial, peroneal, and sural nerves were stimulated
in every possible combination. Once again, localization was performed both with an ide-
alized leadfield and a more realistic one, as well as with and without the spatio-temporal
constraint. In this case, the more realistic leadfields were based on cross-sections of sciatic
nerve samples collected after each experiment. In contrast to the exact anatomical model
used in the simulations, these models are still only approximations. None of the four con-
straint combinations tried led to satisfactory performance, in the sense of allowing us to
infer the number and identity of the active fascicles from one of the estimates. Given the
high levels of error created by the stimulation artefact and inaccurate model, these results
are qualitatively in accordance with the predictions of the simulations. More specifically,
the predicted effect of the spatio-temporal constraint was validated, as it was able to reduce
spurious pathways in practice as well as in theory. The effects of having a more realistic
model were not consistent with the predictions of the simulations, but this is easily explained
by the fact that the “realistic” model used for the experimental data was a much coarser
Discussion 153
approximation than the one used in the simulations. However, although the nerve-specific
model used on the experimental data did not reduce the number of peaks, it did somewhat
decrease the overlap between the estimates for different single-fascicle cases.
In the face of the poor performance of the traditional source localization approach,
we investigated a novel alternative designed to reduce the dependency on anatomical model
accuracy. Instead of constructing the leadfield by solving the forward problem using a model
of the nerve, we constructed an “experimental” leadfield using a collection of patterns ac-
tually observed during a set of training trials. Each pattern was associated with a fascicle,
such that this method is inherently designed to discriminate between fascicles and cannot
provide sub-fascicle resolution. The results for single-fascicle cases were reasonably good:
the correct branch was reliably identified, although this was sometime accompanied by false
positives for the inactive branches. The success of the single-fascicle cases using this tech-
nique proves that the failure of the model-based source localization to reliably identify these
same cases was due to model inaccuracies (either in the anatomy or because of the use of a
distributed dipolar source model), rather than to a fundamental lack of information in the
measurements. On the other hand, results for multi-fascicle cases were not reliable. This can
be attributed to remaining sources of error, such as interference from the artefact, as well
as to the usual difficulty of recovering complex source configurations in the context of an
ill-posed inverse problem. Nonlinear interactions when multiple pathways are present may
also play a role.
In light of the source localization results using both the model-based and experimental
leadfields, the last part of this thesis was concerned with determining how much benefit the
matrix cuff electrode really provides compared to more traditional single-ring multi-contact
nerve cuffs. We focused on the simpler problem of discriminating single fascicles in the
presence of training data and using data from several time instants (i.e., the method is
not strictly instantaneous). We found that very high classification success rates could be
achieved, and that the matrix cuff significantly outperformed the single-ring configuration.
Additionally, we found that an improvement could be achieved even when using the same
Discussion 154
number of contacts in both configurations, simply by virtue of having more flexibility to
choose contact locations in the matrix cuff. The implications for the design and use of
cuff electrodes are that it is beneficial to create devices with high number of contacts, then
perform a calibration procedure after implantation to select a smaller set of contacts that
will actually be used. From the point of view of source localization, there are both pros and
cons to these results. On one hand, we showed that the matrix cuff really does provide more
information that previous designs, and thus has an advantage for source localization. On the
other hand, the results indicate clearly that some contacts are more useful than others, which
implies that the source localization may have less than 56 measurements to take advantage
of or, worse, that some contacts may actually degrade performance.
8.2 Comparison of the results
8.2.1 Resolution achievable
The basic question in this thesis is to determine the spatial resolution that can be achieved
by applying a source localization approach to MCC measurements. The studies presented
in the preceding chapters offer insight into different aspects of this question. The solution
space reduction study of Chapter 3 provided a method to quantify the number of different
locations in the nerve that are theoretically distinguishable. This is useful in that it provides
an upper bound on the resolution that could be achieved. It was found that while single-axon
resolution is not feasible, small groups of axons could be distinguished, meaning that sub-
fascicular resolution is theoretically possible for large fascicles, and that small fascicles can
be discriminated. Again, however, these results deal with the theoretical distinguishability
of regions of the nerve under simple conditions (e.g., only one source is active at a time).
In order to estimate the actual localization performance that could be achieved in more
realistic cases, we must use a different approach. This was the role of Chapter 4, which
contained a simulation study of the source localization, in both single- and multi-pathway
Discussion 155
cases. Spurious and missed pathways proved to be significant obstacles but, for the purposes
of the current discussion, the point of interest is that the localization error was in the
140 µm-180 µm range. This level of uncertainty corresponds to a resolution lower than the
one suggested by Chapter 3, confirming that the source localization resolution in a more
realistic case can be expected to be far from the theoretical upper bound previously derived.
The resolution predicted by the simulations suggests that while discriminating large fascicles
should be possible, smaller fascicles may be confused, and sub-fascicular resolution is unlikely.
The results of Chapter 5 provide a further level of complexity to the issue, by examining
the performance on experimental data. Although the design of the experiment was such
that only whole-fascicle resolution was investigated, even this proved challenging given the
multiple sources of error in the source localization: discriminating single fascicle cases proved
unsuccessful. This part of the study is therefore more informative about practical challenges
than it is about resolution. Chapters 6 and 7, taking into account all the results so far,
used methods which were explicitly limited to whole-fascicle discrimination and excluded
any possibility of sub-fascicular resolution. In both those studies, single-fascicle cases could
be successfully identified. This confirms that the recordings contain enough information at
least for fascicle-level resolution, including a small fascicle (the sural branch), but we must
keep in mind that many human nerves in which these techniques might be applied contain
a much larger number of fascicles, on the order of several dozen. Additionally, multi-fascicle
cases could not be successfully identified, which brings up a central issue in the resolution
question: the amount of localization error is dependent on the number and configuration
of the bioelectric sources. Therefore, it is not possible to obtain a precise estimate of the
resolution given only the shapes of the electrode and nerve. Any discussion of this issue
will inevitably be confined to qualitative observations, unless discussing a specific source
configuration (which would be of little interest). Nonetheless, the multiple studies in this
thesis provide, in the opinion of the author, some useful insight into what is achievable and
what is not.
Discussion 156
8.2.2 Validation of the simulations
As already discussed, the localization results on the experimental data validated the predicted
impact of the spatio-temporal constraint, but not of the nerve-specific leadfield, although
this latter discrepancy is easily explained by differences in the model accuracy. Another issue
related to the validation is that the localization error was predicted by the simulations to be
acceptable in the idealized case of low noise and useful constraints, but as noise increased,
the error increased and the constraints ceased to be beneficial. In the face of the inaccurate
experimental results, this raises the question of how much noise must be present before the
constraints stop being useful. Consulting Figures 4.2 and 4.4, we see that starting at 20%
noise, there is no longer any observable difference in the localization errors of the different
constraint combinations. Given the multiple sources of error during the source localization
on experimental data, it is very likely that the noise was above this threshold (recall that
“noise” in this context can include factors such as model inaccuracies). The simulations
therefore did not predict useful performance under the conditions that were present during
the experiments, and the localization results in Chapter 5 are not inconsistent with the
predictions.
8.2.3 Implication of the contact configuration study on the source
localization results
The results of Chapter 7 demonstrated that good single-fascicle discrimination can be achieved
with a small number of contacts, and also that some contacts may actually be detrimental to
the performance. We therefore need to consider the implications of this information on the
results of the other chapters. On one hand, it raises the possibility that performance could be
slightly increased by excluding some contacts, but at the same time, the numbers of contacts
identified as optimal in the context of the simple problem of Chapter 7 are fairly low (Figure
7.5), and cannot be expected to be sufficient for precise source localization. Initial attempts
were made to re-apply the source localization (using both model-based and experimental
Discussion 157
leadfields) using only the set of contacts selected in Chapter 7, but this decreased rather
than increased performance (results not shown). Although these attempts were made only
on a small portion of the data, they are in accordance with expectations: the methods used
in Chapters 5 and 6 are much more complex than those in Chapter 7, so it is not surprising
that the set of contacts that proved optimal in the latter was not appropriate in the former.
The results of Chapter 7 are still relevant in that they show that not all contacts are equally
useful, but it is clear that the best set of contacts is not the same depending on the algo-
rithm being applied, and that a contact selection procedure must be tailored to the specific
technique being used. Conducting analogous selection procedures for the source localization
methods is, however, much more demanding in terms of time and computational resources,
and was prohibitive in the context of this thesis. In order to overcome these problems, it
may be necessary to develop new contact selection procedures tailored to these techniques,
rather than to use “naıve” approaches like the one in Chapter 7.
8.3 Limitations of the study
8.3.1 Experimental issues
As described in Chapter 5, we encountered three experimental difficulties: movement of
the recording cuff electrodes due to a bulky connector, stimulation artefacts, and poorly
controlled stimulation pulses. The latter issue was relatively minor because the pulses were
still able to reliably produce CAPs, the exact magnitude of which was not crucial to the
study. The other two difficulties were more serious because they significantly increased the
amount of error in the measurements. This is particularly problematic given that the task at
hand is so sensitive to noise. We were therefore not able to test the source localization under
ideal conditions. On one hand, this is counter to our stated goal of evaluating the potential
of the method, but on the other hand, our experimental difficulties were closely related to
the type of difficulties that can be expected in clinical applications. Slight cuff movements
Discussion 158
can occur as a result of body motion, and interference can be expected from other source
of biopotentials (e.g. electromyogram signals). The source localization performance in our
experiments is therefore likely more reflective of the performance that could be expected in
practice. An additional clue as to the impact of these factors on the study is that the cuff
connector was held in place manually for Rats 2 to 6, but clamped in place for Rat 1. If the
motion of the cuff was the main reason for the poor performance, we would expect Rat 1 to
have significantly better performance than the others. Although the recordings for that rat
were somewhat cleaner visually, the source localization was not more successful, so we can
conclude that the slight motion of the cuff was not the dominant factor.
8.3.2 Use of CAPs
The recordings used for this study originate from CAPs elicited via direct electrical stimu-
lation of the nerve branches, and are therefore different from what could be expected during
natural activity. A CAP is produced by a large number of axons firing synchronously. In
contrast, natural activity is generally sparse and asynchronous, such that the total number
of APs occurring within the cuff at any given instant is small. This implies that a signal
recorded when a CAP traverses the cuff will be much larger than signals obtained during
natural activity. The larger amplitudes of our signals were helpful because they increased the
signal-to-noise ratio and clarified the differences between measurement patterns produced by
different branches. How these results translate to the smaller spontaneous neural activity is
again a noise issue and therefore the discussion provided for other sources of error in this
study applies to it as well. The notion that the neural signals will be smaller in practice that
those used here makes it all the more apparent that noise reduction strategies will be vital if
the source localization approach is ever to be applicable. In the mean time, our decision to
deal with the easier case of the CAPs was logical from an experimental point of view, in that
it allowed us to more easily and precisely control which combination of fascicles contained
bioelectric activity in a given trial. Furthermore, given the poor performance obtained it is
Discussion 159
obvious that it would be premature at the moment to apply the source localization approach
to natural neural activity. On the other hand, if the performance on CAPs were to reach
acceptable levels, applying the source localization approach to natural activity would be an
essential next step in validating the technique.
8.3.3 Use of simplified FE models
Although we investigated the localization performance using both idealized and more realistic
FE models, neither type was a completely accurate reflection of the nerve’s true anatomy.
Even in the models based on the nerve sample cross-sections, the trace was not perfectly
accurate and included some approximations (relating for example to the thickness of the
perineurium and the encapsulation tissue layer, as well as to the assumption that the nerve
had the same shape with and without the cuff). Even more significant was the fact that
the cross-sections were simply extruded in both directions to obtain the 3D FE model.
These longitudinally uniform models therefore do not take into consideration the cross-
sectional variations due to the progressive branching of the fascicles. This simplification
is detrimental to the accuracy of the leadfield, and therefore to the success of the source
localization. Constructing a completely accurate full 3D FE model is, however, a daunting
task that would require very sophisticated imaging techniques. At present no such methods
are available clinically, such that a source localization approach that is useable in practice
cannot rely on them. The use of simplified FE models in this thesis is therefore reflective of
the level of information that might be available in clinical application, and therefore more
appropriate for predicting the method’s performance. Similarly, the diameter of the nerve
in the idealized model proved to be too small compared to the measurements from the nerve
samples. As mentioned in Chapter 5, this is likely due to the fact that we deliberately used
large rats, whereas the idealized model was based on information in the literature that was
likely obtained from rats with more average sizes. Once again, however, the goal of having
an idealized model is to dispense with the need to tailor the method to individual nerves,
Discussion 160
and so having a discrepancy in the diameter would not be unexpected in practice.
8.3.4 Use of a low-resolution source localization algorithm
We chose to solve the inverse problem using the sLORETA algorithm. This choice was based
on a careful consideration of the alternatives and the specific characteristics of the peripheral
nerve problem (which differ from those of the EEG source localization problem). Nonetheless,
the fact remains that sLORETA is explicitly designed to produce low-resolution solutions,
which is the price it pays for higher accuracy. In a situation such as the peripheral nerve
problem, where extremely high spatial resolution is desirable, it is obvious that sLORETA
is not an ideal algorithm; it is simply better than the alternatives. There is a need for higher
resolution algorithms tailored to the needs of peripheral nerve source localization, but until
the accuracy of the method can be improved significantly over what was obtained in the
present study, improving the resolution will remain a much lower priority.
8.3.5 Use of peaks in the estimate as a measure of the number of
pathways
The procedure to determine the number of active pathways from the source localization es-
timate, for both simulated and experimental data, is described in Chapter 4. As mentioned
there, this method is not entirely without its flaws. In particular, the presence of an active
pathway may modify the shape of the estimate without producing a distinct peak, for exam-
ple by producing an elongated ridge. Several factors can influence the shape of the estimate
in multi-pathway cases, including the distance between the sources and their relative mag-
nitudes. Finding a method that can reliably determine the number of pathways in such a
widely varying range of situations is a very difficult task. Therefore, for the purposes of this
thesis, the number of peaks was judged to be an acceptable metric for gauging the success of
the localization and the influence of the constraints, even though it is not a perfect method.
The main impact of this limitation will be on the number of missed pathways: if a pathway
Discussion 161
is not detected because it did not create a clear peak, then the number of missed pathways
may be artificially high. The impact of this consideration is, however, minor enough that
our conclusions about the number of missed pathways (e.g. role of sLORETA blurring) still
hold true.
8.3.6 Focus on spatial over temporal resolution
A major reason why bioelectric source localization techniques have been drawing attention
is their high temporal resolution, which gives them access to crucial features of neural ac-
tivity that are unavailable to other slower modalities, such as fMRI or PET. The question
of temporal resolution therefore deserves a mention here, even though little attention was
devoted to it in this thesis. This is not strictly speaking a limitation of the study, because
this data is available: time series were collected from all recording contacts with a sampling
rate of 20 kHz. Until the neural activity can be associated with specific pathways, however,
the characteristics of the signal in time are of little interest because we do not know what
they refer to. If, on the other hand, only temporal information is of interest (as a reflection
of the amalgamated activity in the entire nerve), then it can be obtained with a considerably
simpler electrode and recording setup.
8.3.7 Focus on a specific electrode and nerve
All of the results in this thesis are concerned with performing source localization using a
1 mm-diameter matrix MCC on a rat sciatic nerve. The focus is therefore on a small nerve.
It is possible that in a larger nerve, or with a different electrode shape (e.g. a FINE), the
results may be different. Even if the localization error did not decrease, a similar error would
seem less significant if the overall region was larger, and may allow for better discrimination of
fascicles if they were spaced further apart. For a truly thorough assessment of the potential of
peripheral nerve source localization, different nerve sizes and electrode configurations should
therefore be investigated. In this thesis, however, we focused on the situations that we were
Discussion 162
in a position to test experimentally, in order to be able to validate our results. In addition,
small nerves are of interest because they are more difficult to access when using intraneural
peripheral nerve interfaces such as MEAs, and therefore are a natural target for methods
based on extraneural recordings.
8.4 Optimal number of contacts
The general task of obtaining spatial information about nerve activity from nerve cuff mea-
surements is very difficult. It stands to reason that as we seek to achieve better resolutions,
the amount of information needed will increase. A natural question throughout this study is
therefore how many contacts are needed to achieve the performance goals of a given applica-
tion. In the context of the source localization problem, the number of contacts that will yield
good performance depends on a variety of factors, including the complexity of the region in
the cuff and the source configuration. It is therefore a very thorny issue to try to determine
an optimal number of contacts, and if one were to do so for a very specific situation, it would
almost certainly not generalize to other similar situations. For these reasons, in this thesis
we did not directly investigate the question of which number of contacts is best. Many of
our observations, however, shed some light on different aspects of this issue.
First, our study of the solution space reduction method showed that the amount of
noise had a direct impact on how many contacts were useful for that application. When
there was very little noise the extra contacts were not very beneficial, and when there was
too much noise the benefit again decreased, but between those two extremes the number of
contacts could improve the resolution achievable. It is crucial to keep in mind, however, that
the measure of performance in that study was the number of different regions that could
theoretically be distinguished from one another. It therefore represents an upper bound on
the localization performance, but by no means suggest that this performance will be close to
that bound. There is therefore a distinction between the influence of the number of contacts
on the possible resolution, and on the accuracy of the localization.
Discussion 163
The second main indicator of the influence of the number of contacts was our compar-
ison of the matrix and single-ring cuff configurations. Based this time on experimental data,
and tackling a simplified problem of fascicle identification, we arrived at three conclusions.
First, the way to obtain the best performance was to start with a large number of contacts
and narrow that set down to a smaller subset of the most useful ones. The subset ultimately
selected varied from rat to rat, which further argues against the possibility of designing a cuff
with a set of contacts whose number and locations are optimal. Rather, this result suggests
that calibration procedures may be preferable, although of course the feasibility of this may
become more of an issue in problems more difficult than our simple fascicle discrimination
task. Second, we determined that high performance could be achieved with a small number
of contacts, although this is expected to be highly application-dependent. Third, we high-
lighted the fact that the location of the contacts is crucial, not just their number. Therefore,
varying for example the number of contacts in a single ring and examining the results is
not an appropriate way to investigate this issue. The importance of the locations further
complicates the goal of trying to identify an optimal number of contacts: the issue is not
only how many contacts do we need, but also where should we put them.
Although interesting in their own right and having concrete practical application,
neither of the two studies above directly addresses the fundamental question of this thesis,
which is whether or not bioelectric source localization can be applied successfully to periph-
eral nerves. The performance in that respect was poor, despite using a number of contacts
larger than what was shown to be useful in both the studies just discussed. Even taking into
account the fact, illustrated in Chapter 7, that too many contacts can in fact decrease perfor-
mance, there is clearly a gap between the performance of the source localization and that of
the two simpler problems. More than anything, this illustrates the difficulty of generalizing
any conclusions about the best number of contacts.
Taking all of these factors into consideration, it is the opinion of the authors that
the only way to satisfactorily answer the question “How many contacts do I need?” is by
conducting application-specific pilot studies that use a large number of contacts and inves-
Discussion 164
tigate the performance of various subsets. The general results should then be adapted to
each subject using a calibration procedure. Ultimately, higher number of contacts will not
necessarily always be better, but it is extremely difficult to arrive at specific conclusions a
priori for a given application.
8.5 Number of pathways to be localized
The simulations investigated the case of three simultaneous active pathways and, likewise,
the experiments involved three fascicles. In practice, the number of pathways simultaneously
active may be much higher. For example, the human femoral nerve contains upward of 20
fascicles [150]. The question is then how many pathways can be simultaneously localized
with the type of approach proposed in this thesis. Of course, at the current low level of
performance, the question is premature. Nonetheless, it will become important in the future
if the current obstacles are overcome.
Although the results presented in this thesis do not directly address the issue of a
maximum number of pathways, some of our observations can nevertheless allow us to make
some predictions. First, we have highlighted the fact that the accuracy of the localization
is dependent on the complexity of the source configuration. Second, we found that both
spurious and missed pathways are obstacles to the localization. Missed pathways will become
even more of an issue when there are more pathways, if algorithms are not developed that
avoid smoothing the solution while still retaining accuracy and robustness. Spurious peaks
in the solution may also be even more misleading if the true number of pathways is high,
such that small features in the estimate are actually reflective of activity of interest. Taking
into consideration all of these factors, it seems unlikely that situations with more than 5 or
6 simultaneously active pathways could be accurately and completely reconstructed using
this type of measurement setup. This limit will therefore have to be taken into account in
the development of neuroprosthetic applications. If there are more than 5 or 6 pathways
that are of interest (i.e., not including small pathways whose activity may be interpreted as
Discussion 165
noise without loss of functionality), then alternative strategies may become necessary. This
may include using more than one electrode and placing them more distally, after the nerve
has split into more branches, each of which has fewer fascicles. Alternatively, it may be that
intraneural recording devices such as MEAs will be found to be preferable for large nerves
with many fascicles, whereas MCCs will be preferable for smaller and more delicate nerves
with fewer fascicles.
A closely related issue, which is not specific to the method used to localize the bio-
electric source, arises once we have identified the approximate location of the activity. The
question is then to determine the functional relevance of this information in cases where
fibers with different roles are situated very close together. Identifying the conduction veloc-
ity (Section 2.3.3) can help alleviate this problem if the fibers of interest are of different sizes.
Nonetheless, cases will arise in which fibers have similar conduction velocities and locations,
but different functional implications for the control of a neuroprosthesis (as an example of
this situation, consider the case of a muscle spindle afferent and a Golgi tendon afferent). In
those situations, it may be feasible to classify the nature of the activity based on temporal
patterns. In other words, signal processing methods of the type that have been proposed
for nerve cuff recordings of whole-nerve activity (e.g. artificial neural networks [16]) could
be applied to the signals corresponding to only to a particular region in the nerve. We can
reasonably expect that the performance of these methods would only increase as the number
of different pathways affecting the measurements is reduced. In the case of efferent activity, if
the goal is to control an assistive device through a direct neural interface, then an additional
option would be to treat each of the pathways that can be reliably monitored as a binary
switch. In this way, a multi-switch control scheme could be established that would allow the
user to achieve sophisticated control, without any need to interpret the original functional
relevance of the signals in the pathways.
Discussion 166
8.6 Factors related to the use of a nerve cuff electrode
An important consideration when evaluating the results of this study is that a nerve cuff
electrode increases the uniformity of the electric fields inside of it. In other words, the spatial
variations in the potentials around the nerve are smaller than they would be if there was no
cuff [168, 98]. This is a consequence of the more restricted extracellular environment created
by the cuff, and suggests that were will be an upper limit to the number of recording con-
tacts that are useful. Sampling several point that are too close together to have appreciable
potential differences will not increase the amount of meaningful information obtained. Re-
calling the results of our contact selection study (Chapter 7), however, we know that a limit
on the number of useful contacts may not necessarily translate into a limit on the number of
contacts that are placed on the cuff initially. The increased field uniformity inside a cuff will
also impose a limit on the source localization resolution achievable from these recordings.
The results of the leadfield reduction technique presented in Chapter 3 are useful in this
respect because they provide a tool to help gauge what the attainable resolution might be.
Further, an inherent resolution limit bolsters our decision to use sLORETA, by making its
smoothing of the solution less of an issue. It is also essential to clarify that the increased
field uniformity by no means makes attempts to identify the locations of bioelectric sources
within the nerve hopeless. Among the results of this study, the most informative ones with
respect of this issue are the simulation results presented in Chapter 4. There, we saw that
under favorable conditions (low noise, accurate leadfield, and useful constraints), good per-
formance could be achieved. The FE model used in these simulations accurately reflects the
restricted extracellular medium in which the fields are propagating and therefore the results
are indicative of how big an obstacle the field uniformity really is. We can use the simulations
to draw conclusions about these theoretical limits without being hampered by the numerous
sources of error that were present in the experimental data. In addition, several published
studies (reviewed in Chapter 2) have shown that discrimination of the activity of different
pathways in a nerve is possible, which provides clear evidence that the use of a nerve cuff
Discussion 167
does not completely abolish the presence of useful spatial variations in the fields.
Although the restricted extracellular environment has the disadvantage of increasing
field uniformity, it also has the benefit of increasing the magnitude of the recorded signal.
This was in fact the core motivation behind the introduction of the first nerve cuff electrodes.
From the point of view of bioelectric source localization, the increase in signal-to-noise ratio
afforded by a cuff is very valuable. There is therefore a trade-off between spatial variability
of the potentials in the cuff and the signal-to-noise ratio, a trade-off which seems worthwhile
based on the simulation results presented in this study, and other published studies of nerve
cuff selectivity. Nerve cuff designs have also been introduced that are divided into several
chambers [61]. The selectivity of electrodes in different chambers is thus increased, while at
the same time retaining the benefit of larger signals. While a cuff with numerous contacts
may require numerous chambers and therefore pose a greater challenge in terms of manufac-
turing, the concept is sound and likely to improve the trade-off mentioned above. The nerve
cuff variant know as the FINE, which flattens the nerve to get better access to all fascicles,
has also been shown to improve selectivity, and therefore is a promising technology to com-
bine with the source localization approach. At the time of writing, however, no FINE had
been developed that had the grid-like configuration of contacts that was shown in Chapter
7 to be useful. When discussing the pros and cons of nerve cuffs we may be reminded of the
benefits of intraneural electrode designs, which can combine both high selectivity and high
signal-to-noise ratios, but of course those technologies have their own significant drawbacks
(e.g. damage to the nerve and unsuitability for smaller nerves in the case of MEAs, or lack
of spatial coverage in the case of LIFEs).
A final interesting issue is specific to the use of the MCC. In this study we used data
from all 56 contacts simultaneously, in such a way that a common reference was needed for
all of them. This highlighted a new difficulty that was not present in other nerve cuff studies
in the literature. In the traditional nerve cuff design, the tripole configuration is designed
to minimize the interference from sources outside the cuff, but only if the recording contact
is placed half-way between the two reference contacts (see Section 2.3.1 of the Literature
Discussion 168
Review). When there is an entire grid of contacts, like in the MCC, it is no longer possible
to reduce interference equally at all contacts while still maintaining a single reference for
all of them. In this thesis, we explore two reference options: a common-average reference
(as often used in EEG) was used for the source localization analysis, whereas a tripole
reference was used in Chapter 7 in order to maintain consistency between the different contact
configurations studied. We observed little difference in the noise reduction capabilities of
these two references. In both cases, the noise reduction was most effective in the centre ring
of contacts, as expected. The use of the MCC therefore intrinsically creates a situation in
which the signal-to-noise ratio at different contacts may be significantly different, which in
turn may affect the results of the source localization. Note that if an estimate of the noise
distribution is available, it can be taken into account in the inverse problem using the matrix
Cǫ of Equation 2.18, but a precise estimate may be difficult to obtain.
8.7 Implications for future cuff-based peripheral nerve
interfaces
A general remark regarding the use of a source localization approach in a peripheral nerve is
in order. In the EEG/MEG context, source localization techniques have both a much higher
temporal resolution and a lower spatial resolution than other modalities, such as fMRI. In the
peripheral nerve case, the high temporal resolution remains quite attractive, because of the
short time spans involved in trains of action potentials. The low spatial resolution remains
an area in obvious need of improvement but, in contrast with the EEG/MEG context, alter-
native methods for achieving spatial discrimination of the electrical activity within a nerve
without risking tissue damage are very limited. EEG/MEG source localization studies are
also hindered by the difficulty of assessing the correctness of the results, which is particularly
important in light of the ill-posed nature of the problem, but in peripheral nerves the tech-
nique can be validated in a more definite manner, either by stimulating known pathways (as
Discussion 169
done here) or by using simultaneous MEA recordings. In fact, applying source localization
techniques to MCC recordings is, from a mathematical point of view, the most complete and
flexible framework available to us for creating a peripheral nerve interface that can discrim-
inate between spatial locations in the nerve. Given the poor performance obtained in this
thesis, it is natural to ask what the implications are for future cuff-based peripheral nerve
interfaces, and whether there is a limit to the potential of this type of technology. Indeed,
although cuff-based neuroprostheses have been demonstrated to have some selectivity both
in recording and stimulation, there is no reason to assume that it will ever be possible to
achieve arbitrarily fine resolution. The study in Chapter 3 illustrates this fact quite clearly.
The first question to ask is therefore what neuroprosthetic applications are realistic
for multi-contact nerve cuff electrodes. The main factor will be the number of distinct
regions in the nerve that need to be distinguished. These regions may correspond to different
fascicles, or simply to different pathways in a large unifascicular nerve, but it seems unlikely
that their number will be far above single-digit quantities. However, this estimate only
takes into account discrimination based on spatial variations in the extraneural fields. This
method can be combined with other approaches that rely on temporal information. One
realistic technique would be to use the MCC to detect conduction velocity, then translate
this information into the type of fiber that is active in each of the spatially distinguishable
regions. In this manner, the amount of functional information that may be extracted from
an MCC-based neural interface may be significantly increased. Nevertheless, as long as the
spatial information remains coarse, there will be applications that will not be achievable.
At present, it seems that these types of interfaces can have a role to play for example in
monitoring bladder activity or leg position, but creating an interface with a complex artificial
hand, which involves many more degrees of freedom, is unlikely. As mentioned above, the
size of the nerve may also have an impact on the efficacy of the method and therefore on the
applications for which it is appropriate.
The second question is how the performance of MCC-based interfaces can be brought
from the current low level described in this thesis to the level where the applications dis-
Discussion 170
cussed in the previous paragraph can actually be implemented. Our results highlight several
possible avenues for improvement. Primarily, the results of Chapter 6 suggest that further
research into the experimental leadfield approach would be beneficial. In an inverse problem
context, this could take the form of appropriate constraints and regularization techniques
to improve the performance in multi-fascicular situations. More generally, a promising ap-
proach for identifying combinations of branches may be to establish nonlinear input-output
maps between the active pathways and the cuff measurements [93]. In this way, the com-
plex spatio-temporal interactions between bioelectric sources in the nerve may be captured
more accurately than what is possible within the linear framework of the distributed dipoles
approach used here. As for improving model-based source localization in peripheral nerves,
research should be conducted into how to obtain more accurate anatomical models of the
nerve. The nerve’s shape changes during chronic cuff implantation, and encapsulation tissue
forms, meaning that what is really required is a diagnostic method that can be applied in vivo
to obtain an image of the region inside the cuff, as often as required. This information can
then be used to update the model used in the source localization and ensure that it remains
accurate. One example of a method that could be investigated for this purpose is to adapt
the electrical impedance imaging technique [9] to peripheral nerves. Alternatively, ultra-
sound has been shown to be able to image fascicles within a nerve [77], so that if transducer
technology advanced to the point of being implantable around a nerve, this avenue could
also be explored. Our results have also confirmed that applying physiologically meaningful
constraints can have a positive impact on performance. Investigating new constraints for
this purpose will remain premature until a reasonably accurate model of the nerve is avail-
able, but may be warranted in the future. In this respect, using several implanted electrodes
may open new possibilities for constraining the solution. For instance, having simple cuff
electrodes placed distally on the branches of a nerve could indicate which branch is active.
This information could then be used as a constraint when source localization is applied to
measurements from a larger MCC placed more proximally on the main nerve trunk, in order
to achieve finer resolution within that branch. Lastly, higher-order multipole expansions
Discussion 171
could be used as a source model, in order to determine if the dipole model is too simple and
therefore a source of error.
Chapter 9
Conclusions
Having completed our study of bioelectric source localization in peripheral nerves, our main
conclusions can be summarized as follows:
• The theoretical resolution that can be obtained using an MCC, when a single source
is present, is on the order of a small group of fibers or a small fascicle. The number of
variables in the inverse problem of source localization can be matched to this resolution,
eliminating unnecessary computation time and storage requirements.
• Approximate localization of active pathways in peripheral nerves by applying tradi-
tional source localization algorithms to MCC recordings is theoretically possible for a
small number of pathways, under near-ideal conditions of very low noise, informative
constraints about the solution, and accurate nerve and source models. Even under
those conditions, distinguishing closely spaced pathways may not be possible.
• As soon as the noise exceeds very modest levels (i.e., an NSR of 10%), simulations
showed that performance will be hampered by localization errors in the 140 µm-180 µm
range and high numbers of spurious pathways, in addition to the low resolution. There-
fore, under realistic conditions, the proposed approach is not expected to be reliable.
This was confirmed by applying source localization to experimental recordings from172
Conclusions 173
rat sciatic nerves, and noting that the results could not be used to accurately identify
the branches that were stimulated.
• The application of a physiologically meaningful constraint can reduce the number of
spurious pathways both in theory and practice, but had little effect on the localization
error. The use of an accurate model of the nerve anatomy was similarly useful in theory,
but in practice models custom-made based on cross-sections of the nerves used in the
experiments were not successful in improving performance. This suggests that minor
improvements in model accuracy are not sufficient and that very accurate models are
required. As a result, model inaccuracies such as incorrect or insufficient anatomical
detail as well as the use of the distributed dipoles source model are very likely to
prevent our approach, as initially proposed, from being clinically practical.
• An alternative and novel formulation of the source localization problem, in which the
leadfield is constructed using training data instead of a model, outperforms the model-
based approach and is able to correctly identify single-branch cases, confirming that
MCC recordings can be used for this task. The failure of this method to correctly
identify multi-branch cases can be ascribed to a combination of experimental difficul-
ties, the inherent difficulties of ill-posed inverse problems, and nonlinear interactions
within the nerve. More work is warranted to characterize the influences of these dif-
ferent factors and compensate for them. Based on the results presented in this thesis,
methods that rely on experimentally derived relationships between the active branches
and the measurements are a more promising approach to peripheral nerve interfacing
than model-based alternatives.
• A “matrix” type cuff, in which contacts are laid out in a grid, contains more useful
information than a cuff with a single ring of contacts. This can be attributed to differ-
ences between the interface of each contact with the nerve, and their position relative
to the fascicles. The matrix cuff can take advantage of its better spatial sampling and
Conclusions 174
use only the most useful contacts, therefore achieving superior performance without
necessarily needing to use a larger number of contacts. Consequently, applications
that use MCCs would benefit from implanting devices with large numbers of contacts
spread over the inside of the cuff and then performing a contact selection procedure.
The best number of contacts to use depends on the application.
Bibliography
[1] S. P. Ahlfors, G. V. Simpson, A. M. Dale, J. W. Belliveau, A. K. Liu, A. Korvenoja,
J. Virtanen, M. Huotilainen, R. B. Tootell, H. J. Aronen, and R. J. Ilmoniemi. Spa-
tiotemporal activity of a cortical network for processing visual motion revealed by
MEG and fMRI. J. Neurophys., 82(5):2545–2555, Nov 1999.
[2] L. N. Andreasen and J. J. Struijk. Artefact reduction with alternative cuff configura-
tions. IEEE Trans. Biomed. Eng., 50(10):1160–1166, Oct 2003.
[3] S. Andreassen, R. B. Stein, and M. N. Oguztoreli. Application of optimal multichannel
filtering to simulated nerve signals. Biol. Cybern., 32(1):25–33, Feb 2 1979.
[4] G. E. Backus and J. F. Gilbert. The resolving power of gross earth data. Geophys J
R Astron Soc, 16:169–205, 1968.
[5] S. Baillet, J. C. Mosher, and R. M. Leahy. Electromagnetic brain mapping. IEEE
Signal Proc. Mag., Nov:14–30, 2001.
[6] A. Branner, R. B. Stein, E. Fernandez, Y. Aoyagi, and R. A. Normann. Long-term
stimulation and recording with a penetrating microelectrode array in cat sciatic nerve.
IEEE Trans. Biomed. Eng., 51(1):146–157, Jan 2004.
[7] A. Branner, R. B. Stein, and R. A. Normann. Selective stimulation of cat sciatic nerve
using an array of varying-length microelectrodes. J. Neurophys., 85(4):1585–1594, Apr
2001.175
BIBLIOGRAPHY 176
[8] D. H. Brooks, G. F. Ahmad, R. S. MacLeod, and G. M. Maratos. Inverse electrocar-
diography by simultaneous imposition of multiple constraints. IEEE Trans. Biomed.
Eng., 46(1):3–18, Jan 1999.
[9] B. H. Brown. Electrical impedance tomography (EIT): a review. J. Med. Eng. Technol.,
27(3):97–108, May-Jun 2003.
[10] H. Buchner, G. Knoll, M. Fuchs, A. Rienacker, R. Beckmann, M. Wagner, J. Silny, and
J. Pesch. Inverse localization of electric dipole current sources in finite element models
of the human head. Electroencephalogr. Clin. Neurophysiol, 102(4):267–278, Apr 1997.
[11] J. S. Carp, X. Y. Chen, H. Sheikh, and J. R. Wolpaw. Effects of chronic nerve cuff and
intramuscular electrodes on rat triceps surae motor units. Neurosci. Lett., 312(1):1–4,
Oct 12 2001.
[12] E. Cavallaro, S. Micera, P. Dario, W. Jensen, and T. Sinkjaer. On the intersubject
generalization ability in extracting kinematic information from afferent nervous signals.
IEEE Trans. Biomed. Eng., 50(9):1063–1073, Sep 2003.
[13] M. Cheney, D. Isaacson, J. C. Newell, S. Simske, and J. Goble. NOSER: An algorithm
for solving the inverse conductivity problem. Int. J. Imag. Syst. Tech., 2(2):66–75,
1990.
[14] H. S. Cheng, M. S. Ju, and C. C. Lin. Estimation of peroneal and tibial afferent activity
from a multichannel cuff placed on the sciatic nerve. Muscle Nerve, 32(5):589–599, Nov
2005.
[15] A. Q. Choi, J. K. Cavanaugh, and D. M. Durand. Selectivity of multiple-contact nerve
cuff electrodes: a simulation analysis. IEEE Trans. Biomed. Eng., 48(2):165–172, Feb
2001.
BIBLIOGRAPHY 177
[16] C. Choi, J. Carpaneto, N. Lago, J. Kim, P. Dario, X. Navarro, and S. Micera. Clas-
sification of afferent signals recorded with a single cuff electrode. Proc. IEEE EMBS
29th Ann Int Conf., 1:2385–2388, 2007.
[17] P. R. Christensen, Y. Chen, K. D. Strange, K. Yoshida, and J. A. Hoffer. Multi-
channel recordings from peripheral nerves: 4. Evaluation of selectivity using mechanical
stimulation of individual digits. In Proc. 2nd Ann. Conf. IFESS, Burnaby, BC, Canada,
1997.
[18] G. Creasey, J. Elefteriades, A. DiMarco, P. Talonen, M. Bijak, W. Girsch, and C. Kan-
tor. Electrical stimulation to restore respiration. J. Rehabil. Res. Dev., 33(2):123–132,
Apr 1996.
[19] B. N. Cuffin. EEG localization accuracy improvements using realistically shaped head
models. IEEE Trans. Biomed. Eng., 43(3):299–303, Mar 1996.
[20] S. S. Dalal, A. G. Guggisberg, E. Edwards, K. Sekihara, A. M. Findlay, R. T. Canolty,
M. S. Berger, R. T. Knight, N. M. Barbaro, H. E. Kirsch, and S. S. Nagarajan. Five-
dimensional neuroimaging: localization of the time-frequency dynamics of cortical ac-
tivity. NeuroImage, 40(4):1686–1700, May 1 2008.
[21] A. M. Dale, A. K. Liu, B. R. Fischl, R. L. Buckner, J. W. Belliveau, J. D. Lewine, and
E. Halgren. Dynamic statistical parametric mapping: combining fMRI and MEG for
high-resolution imaging of cortical activity. Neuron, 26(1):55–67, Apr 2000.
[22] A. M. Dale and M. I. Sereno. Improved localization of cortical activity by combining
EEG and MEG with MRI cortical surface reconstruction. J. Cognitive Neurosci.,
5(2):162–176, 1993.
[23] F. Darvas, D. Pantazis, E. Kucukaltun-Yildirim, and R. M. Leahy. Mapping human
brain function with MEG and EEG: methods and validation. NeuroImage, 23 Suppl
1:S289–99, 2004.
BIBLIOGRAPHY 178
[24] F. Darvas, U. Schmitt, A. K. Louis, M. Fuchs, G. Knoll, and H. Buchner. Spatio-
temporal current density reconstruction (stCDR) from EEG/MEG-data. Brain To-
pogr., 13(3):195–207, Spring 2001.
[25] A. Devor, A. K. Dunn, M. L. Andermann, I. Ulbert, D. A. Boas, and A. M. Dale.
Coupling of total hemoglobin concentration, oxygenation, and neural activity in rat
somatosensory cortex. Neuron, 39(2):353–359, Jul 17 2003.
[26] G. S. Dhillon and K. W. Horch. Direct neural sensory feedback and control of a
prosthetic arm. IEEE Trans. Neural Sys. Rehab. Eng., 13(4):468–472, Dec 2005.
[27] G. S. Dhillon, S. M. Lawrence, D. T. Hutchinson, and K. W. Horch. Residual function
in peripheral nerve stumps of amputees: implications for neural control of artificial
limbs. J. Hand Surg., 29(4):605–15; discussion 616–8, Jul 2004.
[28] D. M. Durand, H. J. Park, and B. Wodlinger. Localization and control of activity
in peripheral nerves. In Proc. IEEE EMBS 30th Ann Int Conf., pages 3352–3354,
Vancouver, BC, Canada, 2008.
[29] D. J. Edell. A peripheral nerve information transducer for amputees: long-term multi-
channel recordings from rabbit peripheral nerves. IEEE Trans. Biomed. Eng., 33:203–
214, 1986.
[30] J. J. Foxe, M. E. McCourt, and D. C. Javitt. Right hemisphere control of visuospatial
attention: line-bisection judgments evaluated with high-density electrical mapping and
source analysis. NeuroImage, 19(3):710–726, Jul 2003.
[31] K. J. Friston, W. Penny, C. Phillips, S. Kiebel, G. Hinton, and J. Ashburner. Classical
and Bayesian inference in neuroimaging: theory. NeuroImage, 16(2):465–483, Jun
2002.
[32] M. Fuchs, M. Wagner, T. Kohler, and H. A. Wischmann. Linear and nonlinear current
density reconstructions. J. Clin. Neurophysiol., 16(3):267–295, May 1999.
BIBLIOGRAPHY 179
[33] A. Galka, O. Yamashita, T. Ozaki, R. Biscay, and P. Valdes-Sosa. A solution to the
dynamical inverse problem of EEG generation using spatiotemporal Kalman filtering.
NeuroImage, 23(2):435–453, Oct 2004.
[34] G. H. Golub and C. F. V. Loan. Matrix computations. Johns Hopkins University Press,
Baltimore, 1996.
[35] S. L. Gonzalez Andino, R. Grave de Peralta Menendez, C. M. Lantz, O. Blank, C. M.
Michel, and T. Landis. Non-stationary distributed source approximation: an alterna-
tive to improve localization procedures. Hum. Brain Mapp., 14(2):81–95, Oct 2001.
[36] E. V. Goodall, L. M. Kosterman, J. Holsheimer, and J. J. Struijk. Modeling study of
activation and propagation delays during stimulation of peripheral nerve fibers with a
tripolar cuff electrode. IEEE Trans. Rehab. Eng., 3:272–282, 1995.
[37] I. F. Gorodnitsky, J. S. George, and B. D. Rao. Neuromagnetic source imaging with
FOCUSS: a recursive weighted minimum norm algorithm. Electroencephalogr. Clin.
Neurophysiol, 95(4):231–251, Oct 1995.
[38] I. F. Gorodnitsky and B. D. Rao. Sparse signal reconstruction from limited data using
FOCUSS: a re-weighted minimum norm algorithm. IEEE Trans. Sig. Proc., 45(3):600–
616, 1997.
[39] R. Grave de Peralta Menendez, S. L. Gonzalez Andino, S. Morand, C. M. Michel,
and T. Landis. Imaging the electrical activity of the brain: ELECTRA. Hum. Brain
Mapp., 9(1):1–12, 2000.
[40] R. Grave de Peralta Menendez, O. Hauk, S. Gonzalez Andino, H. Vogt, and C. Michel.
Linear inverse solutions with optimal resolution kernels applied to electromagnetic
tomography. Hum. Brain Mapp., 5:454–467, 1997.
BIBLIOGRAPHY 180
[41] R. Grave de Peralta Menendez, M. M. Murray, and S. L. Gonzalez Andino. Improving
the performance of linear inverse solutions by inverting the resolution matrix. IEEE
Trans. Biomed. Eng., 51(9):1680–1683, 2004.
[42] R. Grave de Peralta Menendez, M. M. Murray, C. M. Michel, R. Martuzzi, and S. L.
Gonzalez Andino. Electrical neuroimaging based on biophysical constraints. NeuroIm-
age, 21(2):527–539, Feb 2004.
[43] R. E. Greenblatt, A. Ossadtchi, and M. E. Pflieger. Local linear estimators for the
bioelectromagnetic inverse problem. IEEE Trans. Signal Proces., 53(9):3403–3412,
2005.
[44] F. Greensite. The temporal prior in bioelectromagnetic source imaging problems. IEEE
Trans. Biomed. Eng., 50(10):1152–1159, Oct 2003.
[45] W. M. Grill and J. T. Mortimer. Quantification of recruitment properties of multiple
contact cuff electrodes. IEEE Trans. Neural Sys. Rehab. Eng., 4(2):49–62, Jun 1996.
[46] W. M. Grill and J. T. Mortimer. Neural and connective tissue response to long-
term implantation of multiple contact nerve cuff electrodes. J. Biomed. Mater. Res.,
50(2):215–226, May 2000.
[47] J. Hadamard. La theorie des equations aux derivees partielles. Editions Scientifiques,
Beijing, 1964.
[48] H. Hallez, B. Vanrumste, R. Grech, J. Muscat, W. D. Clercq, A. Vergult, Y. D’Asseler,
K. P. Camilleri, S. G. Fabri, S. V. Huffel, and I. Lemahieu. Review on solving the
forward problem in EEG source analysis. J. Neuroeng. Rehabil., 4:46, Nov 30 2007.
[49] M. S. Hamalainen and R. J. Ilmoniemi. Interpreting magnetic fields of the brain:
minimum norm estimates. Med. Biol. Eng. Comput., 32(1):35–42, Jan 1994.
BIBLIOGRAPHY 181
[50] M. S. Hamalainen and J. Sarvas. Realistic conductivity geometry model of the human
head for interpretation of neuromagnetic data. IEEE Trans. Biomed. Eng., 36(2):165–
171, Feb 1989.
[51] M. Hansen, M. K. Haugland, and F. Sepulveda. Feasibility of using peroneal nerve
recordings for deriving stimulation timing in a foot drop correction system. Neuro-
modulation, 6(1):68–77, 2003.
[52] M. Hansen, M. K. Haugland, and T. Sinkjaer. Evaluating robustness of gait event
detection based on machine learning and natural sensors. IEEE Trans. Neural Sys.
Rehab. Eng., 12(1):81–88, Mar 2004.
[53] P. C. Hansen. Analysis of discrete ill-posed problems by means of the L-curve. SIAM
Rev., 34:561–580, 1992.
[54] P. C. Hansen. Rank-deficient and discrete ill-posed problems : numerical aspects of
linear inversion. Siam, Philadelphia, Pa., 1998.
[55] M. K. Haugland and T. Sinkjaer. Cutaneous whole nerve recordings used for correction
of footdrop in hemiplegic man. IEEE Trans. Rehab. Eng., 3(4):307–317, 1995.
[56] O. Hauk. Keep it simple: a case for using classical minimum norm estimation in the
analysis of EEG and MEG data. NeuroImage, 21(4):1612–1621, Apr 2004.
[57] W. J. Heetderks and W. J. Williams. Partition of gross peripheral nerve activity into
single unit responses by correlation techniques. Science, 188(4186):373–375, Apr 25
1975.
[58] H. Helmholtz. Ueber einige gesetze der vertheilung elektrischer strme in krperlichen
leitern, mit anwendung auf die thierischen elektrischen versuche. Pogg. Ann. Phys.
Chem., 89:211–233, 353–377, 1853.
BIBLIOGRAPHY 182
[59] H. Helmholtz. Some laws concerning the distribution of electric currents in volume
conductors with applications to experiments on animal electricity. Proc. IEEE, 92:868–
870, 2004.
[60] A. E. Hess, J. L. Dunning, D. J. Tyler, and C. A. Zorman. A polynorbornene-based
microelectrode array for neural interfacing. In Proc. 14th Int. Conf. Solid-State Sens.,
Act. Microsys., pages 1235–1238, Lyon, France, June 10-14 2007.
[61] J. A. Hoffer, Y. Chen, K. Strange, and P. Christensen. Nerve cuff having one or more
isolated chambers. United States Patent 5,824,027, October 20 1998.
[62] K. W. Horch and P. R. Burgess. Peripheral nervous system. In K. W. Horch and
G. S. Dhillon, editors, Neuroprosthetics: Theory and Practice, pages 342–447. World
Scientific Publishing Company, Singapore, 2004.
[63] M. Huang, C. J. Aine, S. Supek, E. Best, D. Ranken, and E. R. Flynn. Multi-
start downhill simplex method for spatio-temporal source localization in magnetoen-
cephalography. Electroencephalogr. Clin. Neurophysiol, 108(1):32–44, Jan 1998.
[64] A. Inmann and M. Haugland. Functional evaluation of natural sensory feedback incor-
porated in a hand grasp neuroprosthesis. Med. Eng. Phys., 26(6):439–447, Jul 2004.
[65] A. Inmann and M. Haugland. Implementation of natural sensory feedback in a portable
control system for a hand grasp neuroprosthesis. Med. Eng. Phys., 26(6):449–458, Jul
2004.
[66] D. Isaacson. Distinguishability of conductivities by electric current computed tomog-
raphy. IEEE Trans. Med. Imag., MI-5(2):91–95, 1986.
[67] S. Jezernik and W. M. Grill. Optimal filtering of whole nerve signals. J. Neurosci.
Meth., 106(1):101–110, Mar 30 2001.
BIBLIOGRAPHY 183
[68] S. C. Jun, J. S. George, W. Kim, J. Pare-Blagoev, S. Plis, D. M. Ranken, and D. M.
Schmidt. Bayesian brain source imaging based on combined MEG/EEG and fMRI
using MCMC. NeuroImage, 40(4):1581–1594, May 1 2008.
[69] E. R. Kandel and J. H. Schwartz. Principles of neural science. McGraw-Hill, New
York, 2000.
[70] D. Khosla, M. Singh, and M. Don. Spatio-temporal EEG source localization using
simulated annealing. IEEE Trans. Biomed. Eng., 44(11):1075–1091, Nov 1997.
[71] K. Kobayashi, T. Akiyama, T. Nakahori, H. Yoshinaga, and J. Gotman. Systematic
source estimation of spikes by a combination of independent component analysis and
RAP-MUSIC. I: Principles and simulation study. Clin. Neurophysiol., 113(5):713–724,
May 2002.
[72] Z. J. Koles, J. C. Lind, and A. C. Soong. Spatio-temporal decomposition of the EEG: a
general approach to the isolation and localization of sources. Electroencephalogr. Clin.
Neurophysiol., 95(4):219–230, Oct 1995.
[73] Z. J. Koles and A. C. Soong. EEG source localization: implementing the
spatio-temporal decomposition approach. Electroencephalogr. Clin. Neurophysiol.,
107(5):343–352, Nov 1998.
[74] A. Korvenoja, J. Huttunen, E. Salli, H. Pohjonen, S. Martinkauppi, J. M. Palva,
L. Lauronen, J. Virtanen, R. J. Ilmoniemi, and H. J. Aronen. Activation of multiple
cortical areas in response to somatosensory stimulation: combined magnetoencephalo-
graphic and functional magnetic resonance imaging. Hum. Brain Mapp., 8(1):13–27,
1999.
[75] A. Kostov, M. Hansen, M. Haugland, and T. Sinkjaer. Adaptive restriction rules
provide functional and safe stimulation pattern for foot drop correction. Artif. Organs,
23(5):443–446, May 1999.
BIBLIOGRAPHY 184
[76] C. Krarup, G. E. Loeb, and G. H. Pezeshkpour. Conduction studies in peripheral
cat nerve using implanted electrodes: III. The effects of prolonged constriction on the
distal nerve segment. Muscle Nerve, 12(11):915–928, Nov 1989.
[77] H. Kubiena, M. Hormann, W. Michlits, M. Tschabitscher, K. Groszschmidt, and
M. Frey. Intraoperative imaging of the brachial plexus by high-resolution ultrasound.
J. Reconstr. Microsurg., 21(7):429–433, Oct 2005.
[78] G. A. Kurstjens, A. Borau, A. Rodriguez, N. J. Rijkhoff, and T. Sinkjaer. Intraopera-
tive recording of electroneurographic signals from cuff electrodes on extradural sacral
roots in spinal cord injured patients. J. Urology, 174(4 Pt 1):1482–1487, Oct 2005.
[79] N. Lago, D. Ceballos, F. J. Rodriguez, T. Stieglitz, and X. Navarro. Long term as-
sessment of axonal regeneration through polyimide regenerative electrodes to interface
the peripheral nerve. Biomaterials, 26(14):2021–2031, May 2005.
[80] J. O. Larsen, M. Thomsen, M. Haugland, and T. Sinkjaer. Degeneration and regener-
ation in rabbit peripheral nerve with long-term nerve cuff electrode implant: a stereo-
logical study of myelinated and unmyelinated axons. Acta Neuropathol., 96(4):365–378,
Oct 1998.
[81] S. M. Lawrence, G. S. Dhillon, and K. W. Horch. Fabrication and characteristics of an
implantable, polymer-based, intrafascicular electrode. J. Neurosci. Meth., 131(1-2):9–
26, Dec 30 2003.
[82] S. M. Lawrence, G. S. Dhillon, W. Jensen, K. Yoshida, and K. W. Horch. Acute
peripheral nerve recording characteristics of polymer-based longitudinal intrafascicular
electrodes. IEEE Trans. Neural Sys. Rehab. Eng., 12(3):345–348, Sep 2004.
[83] T. Lefurge, E. Goodall, K. Horch, L. Stensaas, and A. Schoenberg. Chronically im-
planted intrafascicular recording electrodes. Ann. Biomed. Eng., 19(2):197–207, 1991.
BIBLIOGRAPHY 185
[84] D. Lehmann and C. M. Michel. Intracerebral dipole source localization for FFT power
maps. Electroencephalogr. Clin. Neurophysiol., 76(3):271–276, Sep 1990.
[85] B. K. Lichtenberg and C. J. De Luca. Distinguishability of functionally distinct evoked
neuroelectric signals on the surface of a nerve. IEEE Trans. Biomed. Eng., 26(4):228–
237, Apr 1979.
[86] F. H. Lin, T. Witzel, M. S. Hamalainen, A. M. Dale, J. W. Belliveau, and S. M.
Stufflebeam. Spectral spatiotemporal imaging of cortical oscillations and interactions
in the human brain. NeuroImage, 23(2):582–595, Oct 2004.
[87] A. K. Liu, J. W. Belliveau, and A. M. Dale. Spatiotemporal imaging of human brain
activity using functional MRI constrained magnetoencephalography data: Monte Carlo
simulations. Proc. Natl. Acad. Sci. USA, 95(15):8945–8950, Jul 21 1998.
[88] H. Liu, P. H. Schimpf, G. Dong, X. Gao, F. Yang, and S. Gao. Standardized shrink-
ing LORETA-FOCUSS (SSLOFO): a new algorithm for spatio-temporal EEG source
reconstruction. IEEE Trans. Biomed. Eng., 52(10):1681–1691, Oct 2005.
[89] G. E. Loeb, W. B. Marks, and P. G. Beatty. Analysis and microelectronic design
of tubular electrode arrays intended for chronic, multiple single-unit recording from
captured nerve fibres. Med. Biol. Eng. Comput., 15(2):195–201, Mar 1977.
[90] N. K. Logothetis, J. Pauls, M. Augath, T. Trinath, and A. Oeltermann. Neurophysi-
ological investigation of the basis of the fMRI signal. Nature, 412(6843):150–157, Jul
12 2001.
[91] B. Lutkenhoner. Frequency-domain localization of intracerebral dipolar sources. Elec-
troencephalogr. Clin. Neurophysiol., 82(2):112–118, Feb 1992.
[92] J. Malmivuo and R. Plonsey. Bioelectromagnetism : principles and applications of
bioelectric and biomagnetic fields. Oxford University Press, New York, 1995.
BIBLIOGRAPHY 186
[93] V. Z. Marmarelis. Modeling methodology for nonlinear physiological systems. Annals
of Biomedical Engineering, 25(2):239–251, Mar-Apr 1997.
[94] K. Matsuura and Y. Okabe. Selective minimum-norm solution of the biomagnetic
inverse problem. IEEE Trans. Biomed. Eng., 42(6):608–615, Jun 1995.
[95] K. Matsuura and Y. Okabe. A robust reconstruction of sparse biomagnetic sources.
IEEE Trans. Biomed. Eng., 44(8):720–726, Aug 1997.
[96] K. E. Matzel, U. Stadelmaier, M. Hohenfellner, and W. Hohenberger. Chronic sacral
spinal nerve stimulation for fecal incontinence: long-term results with foramen and
cuff electrodes. Dis. Colon Rectum, 44(1):59–66, Jan 2001.
[97] R. S. McLachlan. Vagus nerve stimulation for intractable epilepsy: a review. J. Clin.
Neurophysiol., 14(5):358–368, Sep 1997.
[98] J. H. Meier, W. L. C. Rutten, and H. B. K. Boom. Extracellular potentials from
active myelinated fibers inside insulated and noninsulated peripheral nerve. IEEE
Trans. Biomed. Eng., 45(9):1146–1154, 1998.
[99] J. W. H. Meijs, O. W. Weier, M. J. Peters, and A. V. Oosterom. On the numerical
accuracy of the boundary element method. IEEE Trans. Biomed. Eng., 36(10):1038–
1049, 1989.
[100] E. Menninghaus, B. Lutkenhoner, and S. L. Gonzalez. Localization of a dipolar source
in a skull phantom: realistic versus spherical model. IEEE Trans. Biomed. Eng.,
41(10):986–989, Oct 1994.
[101] V. Menon, J. M. Ford, K. O. Lim, G. H. Glover, and A. Pfefferbaum. Combined
event-related fMRI and EEG evidence for temporal-parietal cortex activation during
target detection. Neuroreport, 8(14):3029–3037, Sep 29 1997.
BIBLIOGRAPHY 187
[102] S. Micera, W. Jensen, F. Sepulveda, R. R. Riso, and T. Sinkjaer. Neuro-fuzzy extrac-
tion of angular information from muscle afferents for ankle control during standing in
paraplegic subjects: an animal model. IEEE Trans. Biomed. Eng., 48(7):787–794, Jul
2001.
[103] S. Micera, X. Navarro, J. Carpaneto, L. Citi, O. Tonet, P. M. Rossini, M. C. Carrozza,
K. P. Hoffmann, M. Vivo, K. Yoshida, and P. Dario. On the use of longitudinal
intrafascicular peripheral interfaces for the control of cybernetic hand prostheses in
amputees. IEEE Trans. Neural Sys. Rehab. Eng., 16(5):453–472, Oct 2008.
[104] C. M. Michel, M. M. Murray, G. Lantz, S. Gonzalez, L. Spinelli, and R. Grave de
Peralta. EEG source imaging. Clin. Neurophysiol., 115(10):2195–2222, Oct 2004.
[105] C. E. Miller and C. S. Henriquez. Finite element analysis of bioelectric phenomena.
Crit. Rev. Biomed. Eng., 18(3):207–233, 1990.
[106] F. Miwakeichi, E. Martinez-Montes, P. A. Valdes-Sosa, N. Nishiyama, H. Mizuhara,
and Y. Yamaguchi. Decomposing EEG data into space-time-frequency components
using parallel factor analysis. NeuroImage, 22(3):1035–1045, Jul 2004.
[107] M. A. Moffitt and W. M. Grill. Electrical localization of neural activity in the dorsal
horn of the spinal cord: a modeling study. Ann. Biomed. Eng., 32(12):1694–1709, Dec
2004.
[108] J. C. Mosher and R. M. Leahy. Recursive MUSIC: a framework for EEG and MEG
source localization. IEEE Trans. Biomed. Eng., 45(11):1342–1354, Nov 1998.
[109] J. C. Mosher, R. M. Leahy, and P. S. Lewis. EEG and MEG: forward solutions for
inverse methods. IEEE Trans. Biomed. Eng., 46(3):245–259, Mar 1999.
[110] J. C. Mosher, P. S. Lewis, and R. M. Leahy. Multiple dipole modeling and localization
from spatio-temporal MEG data. IEEE Trans. Biomed. Eng., 39(6):541–557, Jun 1992.
BIBLIOGRAPHY 188
[111] C. Nangini, B. Ross, F. Tam, and S. J. Graham. Magnetoencephalographic study
of vibrotactile evoked transient and steady-state responses in human somatosensory
cortex. NeuroImage, 33(1):252–262, Oct 15 2006.
[112] G. G. Naples, J. T. Mortimer, A. Scheiner, and J. D. Sweeney. A spiral nerve cuff
electrode for peripheral nerve stimulation. IEEE Trans. Biomed. Eng., 35(11):905–916,
Nov 1988.
[113] G. G. Naples, J. T. Mortimer, and T. G. H. Yuen. In W. F. Agnew and D. B.
McCreery, editors, Overview of peripheral nerve electrode design and implantation.,
Neural Prostheses: Fundamental Studies, pages 107–144. Prentice Hall, New Jersey,
1990.
[114] X. Navarro, S. Calvet, F. J. Rodriguez, T. Stieglitz, C. Blau, M. Buti, E. Valderrama,
and J. U. Meyer. Stimulation and recording from regenerated peripheral nerves through
polyimide sieve electrodes. J. Periph. Nerv. Sys., 3(2):91–101, 1998.
[115] X. Navarro, T. B. Krueger, N. Lago, S. Micera, T. Stieglitz, and P. Dario. A critical re-
view of interfaces with the peripheral nervous system for the control of neuroprostheses
and hybrid bionic systems. J. Periph. Nerv. Syst., 10(3):229–258, Sep 2005.
[116] X. Navarro, E. Valderrama, T. Stieglitz, and M. Schuttler. Selective fascicular stimula-
tion of the rat sciatic nerve with multipolar polyimide cuff electrodes. Restor. Neurol.
Neurosci., 18:9–21, 2001.
[117] H. S. Oster and Y. Rudy. The use of temporal information in the regularization of the
inverse problem of electrocardiography. IEEE Trans. Biomed. Eng., 39(1):65–75, Jan
1992.
[118] D. Papo, A. Douiri, F. Bouchet, J. C. Bourzeix, J. P. Caverni, and P. M. Baudonniere.
Time-frequency intracranial source localization of feedback-related EEG activity in
hypothesis testing. Cereb. Cortex, 17(6):1314–1322, Jun 2007.
BIBLIOGRAPHY 189
[119] S. Parrini, J. Delbeke, E. Romero, V. Legat, and C. Veraart. Hybrid finite elements
and spectral method for computation of the electric potential generated by a nerve
cuff electrode. Med. Biol. Eng. Comput., 37(6):733–736, Nov 1999.
[120] R. D. Pascual-Marqui. Reply to comments by Hamalainen, Ilmoniemi and Nunez.
ISBET Newsletter, 6:16–28, 1995.
[121] R. D. Pascual-Marqui. Reply to comments made by R. Grave de Peralta Menendez
and S.L. Gonzalez Andino. IJBEM, 1(2), 1999.
[122] R. D. Pascual-Marqui. Review of methods for solving the EEG inverse problem.
IJBEM, 1(1):75–86, 1999.
[123] R. D. Pascual-Marqui. Standardized low-resolution brain electromagnetic tomography
(sLORETA): technical details. Methods Find. Exp. Clin. Pharmacol., 24 Suppl D:5–12,
2002.
[124] R. D. Pascual-Marqui, C. M. Michel, and D. Lehmann. Low resolution electromagnetic
tomography: a new method for localizing electrical activity in the brain. Int. J.
Psychophysiol., 18(1):49–65, Oct 1994.
[125] J. Perez-Orive and D. M. Durand. Modeling study of peripheral nerve recording selec-
tivity. IEEE Trans. Rehab. Eng., 8(3):320–329, Sep 2000.
[126] C. Phillips, J. Mattout, M. D. Rugg, P. Maquet, and K. J. Friston. An empiri-
cal Bayesian solution to the source reconstruction problem in EEG. NeuroImage,
24(4):997–1011, Feb 15 2005.
[127] C. Phillips, M. D. Rugg, and K. J. Friston. Anatomically informed basis functions for
EEG source localization: combining functional and anatomical constraints. NeuroIm-
age, 16(3 Pt 1):678–695, Jul 2002.
BIBLIOGRAPHY 190
[128] C. Phillips, M. D. Rugg, and K. J. Friston. Systematic regularization of linear inverse
solutions of the EEG source localization problem. NeuroImage, 17(1):287–301, Sep
2002.
[129] R. Plonsey. Action potential sources and their volume conductor fields. Proc. IEEE,
65(5):601–611, 1977.
[130] R. Plonsey and R. C. Barr. Bioelectricity : a quantitative approach. Plenum Press,
New York, 1988.
[131] R. Plonsey and D. B. Heppner. Considerations of quasi-stationarity in electrophysio-
logical systems. Bull. Math. Biophys., 29(4):657–664, Dec 1967.
[132] M. Rahal, J. Taylor, and N. Donaldson. The effect of nerve cuff geometry on in-
terference reduction: a study by computer modeling. IEEE Trans. Biomed. Eng.,
47(1):136–138, Jan 2000.
[133] M. Rahal, J. Winter, J. Taylor, and N. Donaldson. An improved configuration for the
reduction of EMG in electrode cuff recordings: a theoretical approach. IEEE Trans.
Biomed. Eng., 47(9):1281–1284, Sep 2000.
[134] B. D. Rao, K. Engan, S. F. Cotter, J. Palmer, and K. Kreutz-Delgado. Subset selection
in noise based on diversity measure minimization. IEEE Trans. Sig. Proc., 51(3):760–
770, 2003.
[135] R. Rieger, M. Schuettler, D. Pal, C. Clarke, P. Langlois, J. Taylor, and N. Donaldson.
Very low-noise ENG amplifier system using CMOS technology. IEEE Trans. Neural
Sys. Rehab. Eng., 14(4):427–437, Dec 2006.
[136] R. Rieger, J. Taylor, E. Comi, N. Donaldson, M. Russold, C. M. Mahony, J. A.
McLaughlin, E. McAdams, A. Demosthenous, and J. C. Jarvis. Experimental de-
termination of compound action potential direction and propagation velocity from
multi-electrode nerve cuffs. Med. Eng. Phys., 26(6):531–534, Jul 2004.
BIBLIOGRAPHY 191
[137] S. E. Robinson and J. Vrba. Functional neuroimaging by synthetic aperture magne-
tometry (SAM). In T. Yoshimoto, M. Kotani, S. Kuriki, H. Karibe, and N. Nakasoto,
editors, Recent advances in biomagnetism., pages 302–305. Tohoku University Press,
Senai, 1999.
[138] F. J. Rodriguez, D. Ceballos, M. Schuttler, A. Valero, E. Valderrama, T. Stieglitz, and
X. Navarro. Polyimide cuff electrodes for peripheral nerve stimulation. J. Neurosci.
Meth., 98(2):105–118, Jun 1 2000.
[139] E. Romero, J. F. Denef, J. Delbeke, A. Robert, and C. Veraart. Neural morphological
effects of long-term implantation of the self-sizing spiral cuff nerve electrode. Med.
Biol. Eng. Comput., 39(1):90–100, Jan 2001.
[140] J. Rozman, B. Zorko, and M. Bunc. Selective recording of electroneurograms from the
sciatic nerve of a dog with multi-electrode spiral cuffs. Jpn. J. Physiol., 50(5):509–514,
Oct 2000.
[141] J. Rozman, B. Zorko, A. Seliskar, and M. Bunc. Selective recording of neuroelectric
activity from the peripheral nerve. Pflug. Arch. Eur. J. Phy., 440(5 Suppl):R157–9,
2000.
[142] W. L. Rutten. Selective electrical interfaces with the nervous system. Annu. Rev.
Biomed. Eng., 4:407–452, 2002.
[143] W. L. Rutten, T. A. Frieswijk, J. P. Smit, T. H. Rozijn, and J. H. Meier. 3D neuro-
electronic interface devices for neuromuscular control: design studies and realisation
steps. Biosens. Bioelectron., 10(1-2):141–153, 1995.
[144] W. L. Rutten, J. P. Smit, T. A. Frieswijk, J. A. Bielen, A. L. Brouwer, J. R. Buitenweg,
and C. Heida. Neuro-electronic interfacing with multielectrode arrays. IEEE Eng. Med.
Biol., 18(3):47–55, May-Jun 1999.
BIBLIOGRAPHY 192
[145] M. Sahin and D. M. Durand. Selective recording with a multi-contact nerve cuff
electrode. In Proc. 18th Ann. Int. Conf. IEEE EMBS, pages 369–370, Amsterdam,
Netherlands, 1996.
[146] M. Sahin, D. M. Durand, and M. A. Haxhiu. Improved nerve cuff electrode recordings
by sub-threshold anodic currents. In Proc. 17th Ann. Conf. IEEE EMBS, pages 1107–
1108, Montreal, Canada, 1995.
[147] M. A. Sato, T. Yoshioka, S. Kajihara, K. Toyama, N. Goda, K. Doya, and M. Kawato.
Hierarchical bayesian estimation for MEG inverse problem. NeuroImage, 23(3):806–
826, Nov 2004.
[148] M. Scherg, T. Bast, and P. Berg. Multiple source analysis of interictal spikes: goals,
requirements, and clinical value. J. Clin. Neurophysiol., 16(3):214–224, May 1999.
[149] M. Scherg and D. V. Cramon. Two bilateral sources of the late AEP as identified by
a spatio-temporal dipole model. Electroencephalogr. Clin. Neurophysiol., 62(1):32–44,
Jan 1985.
[150] M. A. Schiefer, R. J. Triolo, and D. J. Tyler. A model of selective activation of the
femoral nerve with a flat interface nerve electrode for a lower extremity neuroprosthesis.
IEEE Trans. Neural Sys. Rehab. Eng., 16(2):195–204, Apr 2008.
[151] H. Schmalbruch. Fiber composition of the rat sciatic nerve. Anat. Rec., 215(1):71–81,
May 1986.
[152] D. M. Schmidt, J. S. George, and C. C. Wood. Bayesian inference applied to the
electromagnetic inverse problem. Hum. Brain Mapp., 7(3):195–212, 1999.
[153] U. Schmitt and A. K. Louis. Efficient algorithms for the regularization of dynamic
inverse problems: I. Theory. Inverse Probl., 18:645–658, 2002.
BIBLIOGRAPHY 193
[154] U. Schmitt, A. K. Louis, F. Darvas, H. Buchner, and M. Fuchs. Numerical aspects of
spatio-temporal current density reconstruction from EEG-/MEG-data. IEEE Trans.
Med. Imaging, 20(4):314–324, Apr 2001.
[155] U. Schmitt, A. K. Louis, C. Wolters, and M. Vauhkonen. Efficient algorithms for the
regularization of dynamic inverse problems: II. Applications. Inverse Probl., 18:659–
676, 2002.
[156] M. Schuettler, K. P. Koch, T. Stieglitz, O. Scholz, W. Haberer, R. Keller, and J. U.
Meyer. Multichannel neural cuff electrodes with integrated multiplexer circuit. In 1st
Ann. Int. IEEE EMBS Spec. Topic Conf. Microtech. in Med. and Biol., Lyon, France,
October 12-14 2000.
[157] M. Schuettler, I. F. Triantis, B. Rubehn, and T. Stieglitz. Matrix cuff electrodes for
fibre and fascicle selective peripheral nerve recording and stimulation. In Proc. 12th
Ann. Conf. IFESS, Philadelphia, PA, 2007.
[158] SCIRun: A Scientific Computing Problem Solving Environment. Scientific Computing
and Imaging Institute (SCI), http://software.sci.utah.edu/scirun.html, 2002.
[159] S. H. Scott. Optimal feedback control and the neural basis of volitional motor control.
Nat. Rev. Neurosci., 5(7):532–546, Jul 2004.
[160] M. Seger, G. Fischer, R. Modre, B. Messnarz, F. Hanser, and B. Tilg. Lead field com-
putation for the electrocardiographic inverse problem – finite elements versus boundary
elements. Comput. Methods Programs Biomed., 77(3):241–252, Mar 2005.
[161] K. Sekihara, S. Nagarajan, D. Poeppel, and Y. Miyashita. Time-frequency MEG-
MUSIC algorithm. IEEE Trans. Med. Imaging, 18(1):92–97, Jan 1999.
[162] K. Sekihara, M. Sahani, and S. S. Nagarajan. Localization bias and spatial resolution of
adaptive and non-adaptive spatial filters for MEG source reconstruction. NeuroImage,
25(4):1056–1067, May 1 2005.
BIBLIOGRAPHY 194
[163] D. Sharon, M. S. Hamalainen, R. B. Tootell, E. Halgren, and J. W. Belliveau. The
advantage of combining MEG and EEG: Comparison to fMRI in focally stimulated
visual cortex. NeuroImage, 36(4):1225–1235, Jul 15 2007.
[164] G. Shou, M. Jiang, L. Xia, W. Wei, F. Liu, and S. Crozier. A comparison of different
choices for the regularization parameter in inverse electrocardiography models. In Proc.
28th Ann. Int. Conf. IEEE EMBS, Aug 30-Sept 2006.
[165] C. Silva, J. C. Maltez, E. Trindade, A. Arriaga, and E. Ducla-Soares. Evaluation of
L1 and L2 minimum norm performances on EEG localizations. Clin. Neurophysiol.,
115(7):1657–1668, Jul 2004.
[166] R. B. Stein, D. Charles, L. Davis, J. Jhamandas, A. Mannard, and T. R. Nichols.
Principles underlying new methods for chronic neural recording. Can. J. Neurol. Sci.,
2(3):235–244, Aug 1975.
[167] T. Stieglitz, H. Beutel, M. Schuettler, and J. W. Meyer. Micromachined, polyimide-
based devices for flexible neural interfaces. Biomed. Microdev., 2:283–294, 2000.
[168] J. J. Struijk. The extracellular potential of a myelinated nerve fiber in an unbounded
medium and in nerve cuff models. Biophys. J., 72(6):2457–2469, Jun 1997.
[169] J. J. Struijk, M. K. Haugland, and M. Thomsen. Fascicle selective recording with
a nerve cuff electrode. In Proc. 18th Ann. Int. Conf. IEEE EMBS, pages 361–362,
Amsterdam, 31 Oct-3 Nov 1996.
[170] J. J. Struijk and M. Thomsen. Tripolar nerve cuff recordings: stimulus artifact, EMG,
and the recorded nerve signal. In Proc. 17th Ann. Int. Conf. IEEE EMBS, 1995.
[171] J. J. Struijk, M. Thomsen, J. O. Larsen, and T. Sinkjaer. Cuff electrodes for long-term
recording of natural sensory information. IEEE Eng. Med. Biol., 18(3):91–98, May-Jun
1999.
BIBLIOGRAPHY 195
[172] J. D. Sweeney, J. T. Mortimer, and D. Durand. Modeling of mammalian myelinated
nerve for functional neuromuscular stimulation. In Proc. 9th Ann. Int. Conf. IEEE
EMBS, pages 1577–1578, 1987.
[173] J. Taylor, N. Donaldson, and J. Winter. Multiple-electrode nerve cuffs for low-velocity
and velocity-selective neural recording. Med. Biol. Eng. Comput., 42(5):634–643, Sep
2004.
[174] C. Tesche and M. Kajola. A comparison of the localization of spontaneous neuromag-
netic activity in the frequency and time domains. Electroencephalogr. Clin. Neurophys-
iol., 87(6):408–416, Dec 1993.
[175] W. Tesfayesus and D. M. Durand. Blind source separation of neural recordings and
control signals. In Proc. 28th Ann. Int. Conf. IEEE EMBS, Aug. 2006.
[176] W. Tesfayesus and D. M. Durand. Blind source separation of peripheral nerve record-
ings. J. Neural Eng., 4(3):S157–67, Sep 2007.
[177] W. Tesfayesus, P. Yoo, M. Moffitt, and D. M. Durand. Blind source separation of
neural recordings and control signals. In Proc. 26th Ann. Int. Conf. IEEE EMBS,
Sept. 2004.
[178] R. D. Throne, L. G. Olson, and J. R. Windle. A new method for incorporating weighted
temporal and spatial smoothing in the inverse problem of electrocardiography. IEEE
Trans. Biomed. Eng., 49(9):1054–1059, Sep 2002.
[179] A. N. Tikhonov and V. Y. Arsenin. Solutions of ill-posed problems. Winston; dis-
tributed solely by Halsted Press, Washington; New York, 1977.
[180] N. J. Trujillo-Barreto, E. Aubert-Vazquez, and W. D. Penny. Bayesian M/EEG source
reconstruction with spatio-temporal priors. NeuroImage, 39(1):318–335, Jan 1 2008.
BIBLIOGRAPHY 196
[181] N. J. Trujillo-Barreto, E. Aubert-Vazquez, and P. A. Valdes-Sosa. Bayesian model
averaging in EEG/MEG imaging. NeuroImage, 21(4):1300–1319, Apr 2004.
[182] D. J. Tyler and D. M. Durand. A slowly penetrating interfascicular nerve electrode for
selective activation of peripheral nerves. IEEE Trans. Rehab. Eng., 5(1):51–61, 1997.
[183] D. J. Tyler and D. M. Durand. Functionally selective peripheral nerve stimulation with
a flat interface nerve electrode. IEEE Trans. Neural Syst. Rehabil. Eng., 10(4):294–303,
Dec 2002.
[184] D. J. Tyler and D. M. Durand. Chronic response of the rat sciatic nerve to the flat
interface nerve electrode. Ann. Biomed. Eng., 31(6):633–642, Jun 2003.
[185] B. Upshaw and T. Sinkjaer. Digital signal processing algorithms for the detection
of afferent nerve activity recorded from cuff electrodes. IEEE Trans. Rehab. Eng.,
6(2):172–181, Jun 1998.
[186] K. Uutela, M. Hamalainen, and R. Salmelin. Global optimization in the localization
of neuromagnetic sources. IEEE Trans. Biomed. Eng., 45(6):716–723, Jun 1998.
[187] K. Uutela, M. Hamalainen, and E. Somersalo. Visualization of magnetoencephalo-
graphic data using minimum current estimates. NeuroImage, 10(2):173–180, Aug 1999.
[188] B. D. Van Veen, W. van Drongelen, M. Yuchtman, and A. Suzuki. Localization of
brain electrical activity via linearly constrained minimum variance spatial filtering.
IEEE Trans. Biomed. Eng., 44(9):867–880, Sep 1997.
[189] D. Vitacco, D. Brandeis, R. Pascual-Marqui, and E. Martin. Correspondence of event-
related potential tomography and functional magnetic resonance imaging during lan-
guage processing. Hum. Brain Mapp., 17(1):4–12, Sep 2002.
[190] J. Vrba and S. E. Robinson. Signal processing in magnetoencephalography. Methods,
25(2):249–271, Oct 2001.
BIBLIOGRAPHY 197
[191] K. Warwick, M. Gasson, B. Hutt, I. Goodhew, P. Kyberd, B. Andrews, P. Teddy, and
A. Shad. The application of implant technology for cybernetic systems. Arch. Neurol.,
60(10):1369–1373, Oct 2003.
[192] R. L. Waters, D. R. McNeal, W. Faloon, and B. Clifford. Functional electrical stimu-
lation of the peroneal nerve for hemiplegia. Long-term clinical follow-up. J Bone Joint
Surg. Am., 67(5):792–793, Jun 1985.
[193] D. Weinstein, L. Zhukov, and C. Johnson. Lead-field bases for electroencephalography
source imaging. Ann. Biomed. Eng., 28(9):1059–1065, 2000.
[194] P. Xu, Y. Tian, H. Chen, and D. Yao. Lp norm iterative sparse solution for EEG
source localization. IEEE Trans. Biomed. Eng., 54(3):400–409, 2007.
[195] O. Yamashita, A. Galka, T. Ozaki, R. Biscay, and P. Valdes-Sosa. Recursive penalized
least squares solution for dynamical inverse problems of EEG generation. Hum. Brain
Mapp., 21(4):221–235, Apr 2004.
[196] Y. Yan, P. L. Nunez, and R. T. Hart. Finite-element model of the human head: scalp
potentials due to dipole sources. Med. Biol. Eng. Comput., 29(5):475–481, Sep 1991.
[197] P. B. Yoo and D. M. Durand. Selective recording of the canine hypoglossal nerve using
a multicontact flat interface nerve electrode. IEEE Trans. Biomed. Eng., 52(8):1461–
1469, Aug 2005.
[198] K. Yoshida and K. Horch. Closed-loop control of ankle position using muscle affer-
ent feedback with functional neuromuscular stimulation. IEEE Trans. Biomed. Eng.,
43(2):167–176, Feb 1996.
[199] K. Yoshida and J. Struijk. The theory of peripheral nerve recording. In K. W. Horch
and G. S. Dhillon, editors, Neuroprosthetics: Theory and Practice, pages 342–447.
World Scientific Publishing Company, Singapore, 2004.
BIBLIOGRAPHY 198
[200] J. Zariffa, M. K. Nagai, Z. J. Daskalakis, and M. R. Popovic. Bioelectric source
localization in the rat sciatic nerve: Initial assessment using an idealized nerve model.
In Proc. World Congress on Med. Physics and Biomed. Eng., Munich, Germany, 2009.
[201] J. Zariffa, M. K. Nagai, Z. J. Daskalakis, and M. R. Popovic. Comparison of model-
based and experimental leadfields in the peripheral nerve source localization problem.
submitted, 2009.
[202] J. Zariffa, M. K. Nagai, Z. J. Daskalakis, and M. R. Popovic. Influence of the number
and location of recording contacts on the selectivity of a nerve cuff electrode. IEEE
Trans. Neural Sys. Rehab. Eng., in press, 2009.
[203] J. Zariffa and M. R. Popovic. Finite element mesh reduction guidelines for the inverse
problem of source localization in peripheral nerves. In Proc. 30th CMBEC. CMBES,
Jun 2007.
[204] J. Zariffa and M. R. Popovic. Application of EEG source localization algorithms to
the monitoring of active pathways in peripheral nerves. In Proc. IEEE EMBS 30th
Ann Int Conf., pages 4216–4219, Vancouver, BC, Canada, August 21-24 2008.
[205] J. Zariffa and M. R. Popovic. Dimensionality reduction of a bioelectric source lo-
calization problem using forward field similarities. In Proc. Neural Interfaces Conf.,
Cleveland, OH, USA, June 16-18 2008.
[206] J. Zariffa and M. R. Popovic. Solution space reduction in the peripheral nerve source
localization problem using forward field similarities. J. Neural Eng., 5(2):191–202, Jun
2008.
[207] J. Zariffa and M. R. Popovic. Localization of active pathways in peripheral nerves: A
simulation study. IEEE Trans. Neural Sys. Rehab. Eng., 17(1):53–62, 2009.