3D Reconstruction of Long Bones Utilising Magnetic ... · Abstract IV usability of MRI as a...
Transcript of 3D Reconstruction of Long Bones Utilising Magnetic ... · Abstract IV usability of MRI as a...
3D Reconstruction of Long Bones
Utilising Magnetic Resonance Imaging
(MRI)
Thesis submitted by
Kanchana Rathnayaka
Rathnayaka Mudiyanselage
MBBS
This thesis is submitted in fulfilment of the requirements for the
degree of Doctor of Philosophy
Institute of Health and Biomedical Innovation
School of Engineering Systems
Faculty of Built Environment and Engineering
Queensland University of Technology
Brisbane, Australia
2011
Abstract
III
Abstract
The design of pre-contoured fracture fixation implants (plates and nails) that
correctly fit the anatomy of a patient utilises 3D models of long bones with accurate
geometric representation. 3D data is usually available from computed tomography
(CT) scans of human cadavers that generally represent the above 60 year old age
group. Thus, despite the fact that half of the seriously injured population comes from
the 30 year age group and below, virtually no data exists from these younger age
groups to inform the design of implants that optimally fit patients from these groups.
Hence, relevant bone data from these age groups is required. The current gold
standard for acquiring such data–CT–involves ionising radiation and cannot be used
to scan healthy human volunteers. Magnetic resonance imaging (MRI) has been
shown to be a potential alternative in the previous studies conducted using small
bones (tarsal bones) and parts of the long bones. However, in order to use MRI
effectively for 3D reconstruction of human long bones, further validations using long
bones and appropriate reference standards are required.
Accurate reconstruction of 3D models from CT or MRI data sets requires an accurate
image segmentation method. Currently available sophisticated segmentation methods
involve complex programming and mathematics that researchers are not trained to
perform. Therefore, an accurate but relatively simple segmentation method is
required for segmentation of CT and MRI data. Furthermore, some of the limitations
of 1.5T MRI such as very long scanning times and poor contrast in articular regions
can potentially be reduced by using higher field 3T MRI imaging. However, a
quantification of the signal to noise ratio (SNR) gain at the bone - soft tissue
interface should be performed; this is not reported in the literature. As MRI scanning
of long bones has very long scanning times, the acquired images are more prone to
motion artefacts due to random movements of the subject‟s limbs. One of the
artefacts observed is the step artefact that is believed to occur from the random
movements of the volunteer during a scan. This needs to be corrected before the
models can be used for implant design.
As the first aim, this study investigated two segmentation methods: intensity
thresholding and Canny edge detection as accurate but simple segmentation methods
for segmentation of MRI and CT data. The second aim was to investigate the
Abstract
IV
usability of MRI as a radiation free imaging alternative to CT for reconstruction of
3D models of long bones. The third aim was to use 3T MRI to improve the poor
contrast in articular regions and long scanning times of current MRI. The fourth and
final aim was to minimise the step artefact using 3D modelling techniques.
The segmentation methods were investigated using CT scans of five ovine femora.
The single level thresholding was performed using a visually selected threshold level
to segment the complete femur. For multilevel thresholding, multiple threshold levels
calculated from the threshold selection method were used for the proximal,
diaphyseal and distal regions of the femur. Canny edge detection was used by
delineating the outer and inner contour of 2D images and then combining them to
generate the 3D model. Models generated from these methods were compared to the
reference standard generated using the mechanical contact scans of the denuded
bone. The second aim was achieved using CT and MRI scans of five ovine femora
and segmenting them using the multilevel threshold method. A surface geometric
comparison was conducted between CT based, MRI based and reference models. To
quantitatively compare the 1.5T images to the 3T MRI images, the right lower limbs
of five healthy volunteers were scanned using scanners from the same manufacturer.
The images obtained using the identical protocols were compared by means of SNR
and contrast to noise ratio (CNR) of muscle, bone marrow and bone. In order to
correct the step artefact in the final 3D models, the step was simulated in five ovine
femora scanned with a 3T MRI scanner. The step was corrected using the iterative
closest point (ICP) algorithm based aligning method.
The present study demonstrated that the multi-threshold approach in combination
with the threshold selection method can generate 3D models from long bones with an
average deviation of 0.18 mm. The same was 0.24 mm of the single threshold
method. There was a significant statistical difference between the accuracy of models
generated by the two methods. In comparison, the Canny edge detection method
generated average deviation of 0.20 mm. MRI based models exhibited 0.23 mm
average deviation in comparison to the 0.18 mm average deviation of CT based
models. The differences were not statistically significant. 3T MRI improved the
contrast in the bone–muscle interfaces of most anatomical regions of femora and
tibiae, potentially improving the inaccuracies conferred by poor contrast of the
articular regions. Using the robust ICP algorithm to align the 3D surfaces, the step
Abstract
V
artefact that occurred by the volunteer moving the leg was corrected, generating
errors of 0.32 ± 0.02 mm when compared with the reference standard.
The study concludes that magnetic resonance imaging, together with simple
multilevel thresholding segmentation, is able to produce 3D models of long bones
with accurate geometric representations. The method is, therefore, a potential
alternative to the current gold standard CT imaging.
Keywords
VI
Keywords
Magnetic resonance imaging
Computed tomography
Image segmentation
3D models
Long bones
Thresholding
Edge detection
Multi thresholding
Higher field MRI
Musculoskeletal MRI
Motion artefacts
Validation
Contents
VII
Contents
Abstract ................................................................................................................. III
Keywords ................................................................................................................. VI
List of figures ............................................................................................................. XIII
List of tables ................................................................................................................ XV
Publications, presentations and awards .................................................................... XVI
Authorship .............................................................................................................. XIX
Acknowledgement ...................................................................................................... XXI
Abbreviations ............................................................................................................ XXII
Chapter 1. Introduction .............................................................................................. 1
Chapter 2. Quantitative imaging of the skeletal system for 3D reconstruction
(Background) ............................................................................................ 7
2.1 Introduction ................................................................................................... 7
2.2 Computed tomography (CT) .......................................................................... 8
2.2.1 Basic principles of CT ........................................................................ 8
2.2.2 Radiation exposure during CT imaging ............................................... 9
2.3 Magnetic resonance imaging (MRI) .............................................................10
2.3.1 Basic principles of MRI .....................................................................10
2.3.2 How tissue contrast is determined ......................................................12
2.3.3 Selection of slice position and thickness ............................................13
2.3.4 Pulse sequences .................................................................................14
2.3.5 MRI safety .........................................................................................14
2.3.6 Signal to noise ratio of an MRI system ...............................................14
2.3.7 Artefacts of MRI ................................................................................15
2.3.7.1 Motion artefacts ..........................................................................16
2.3.7.2 Magnetic susceptibility difference artefact ..................................16
2.3.7.3 Chemical shift ............................................................................17
Contents
VIII
2.3.8 MRI for imaging of the skeletal system ............................................. 17
2.3.9 Advantages and current limitations of MRI ....................................... 18
2.3.9.1 Longer scanning times of MRI ................................................... 18
2.3.9.2 Poor contrast in certain anatomical regions ................................. 18
2.3.9.3 Non-uniformity of the external magnetic field ............................ 19
2.3.9.4 Limited accessibility .................................................................. 19
2.4 Summary ..................................................................................................... 20
Chapter 3. Image processing and surface reconstruction ........................................ 21
3.1 Introduction ................................................................................................. 21
3.2 Acquisition of data for 3D modelling of bones ............................................. 22
3.2.1 Effect of in plane resolution and slice thickness on accuracy of
reconstructed 3D models ................................................................... 23
3.3 Image segmentation ..................................................................................... 24
3.3.1 Manual segmentation ........................................................................ 25
3.3.2 Intensity thresholding ........................................................................ 25
3.3.2.1 Selecting an appropriate threshold level...................................... 26
3.3.2.2 Multilevel thresholding .............................................................. 26
3.3.3 Edge detection ................................................................................... 28
3.3.4 Region growing ................................................................................. 28
3.3.5 Sophisticated segmentation methods.................................................. 29
3.4 Surface generation ....................................................................................... 29
3.5 Registration (aligning) and comparison of surfaces ...................................... 30
3.6 A reference standard for validating 3D models of bones ............................... 30
3.7 Aims of the study ......................................................................................... 32
3.8 Methods ....................................................................................................... 32
3.8.1 Samples ............................................................................................. 32
3.8.2 Image segmentation ........................................................................... 32
Contents
IX
3.8.3 Reference model for validation of the outer 3D models ......................33
3.8.3.1 Removal of the soft tissues from long bones ...............................33
3.8.3.2 Scanning of the bone‟s outer surface using the contact scanner ...34
3.8.3.3 Reconstruction of the 3D model from scanned surfaces ..............37
3.8.4 Reference model for validation of the medullary canal .......................39
3.8.5 Basic 3D modelling techniques using Rapidform 2006 ......................41
3.8.5.1 Registration of 3D surfaces using Rapidform 2006 .....................41
3.8.5.2 Comparison of the aligned 3D models ........................................43
3.8.5.3 Dividing the 3D models of bones into anatomical regions ...........44
3.9 Results .........................................................................................................44
3.10 Summary, discussion and conclusion ............................................................45
3.11 Paper 1: Effect of CT image segmentation methods on the accuracy of long
bone 3D reconstructions (published) .............................................................48
Chapter 4. Application of 3D modelling techniques for orthopaedic implant design
and validation ..........................................................................................57
4.1 Introduction ..................................................................................................57
4.2 3D models for implant design and validation ................................................58
4.3 Aims of the study .........................................................................................59
4.4 Methods .......................................................................................................59
4.5 Results .........................................................................................................59
4.6 Summary, discussion and conclusion ............................................................60
4.7 Paper 2: Quantitative fit assessment of tibial nail designs using 3D computer
modelling (published) ...................................................................................61
Chapter 5. Magnetic resonance imaging for 3D reconstruction of long bones ........67
5.1 Introduction ..................................................................................................67
5.2 Imaging of skeletal system with MRI ...........................................................68
5.3 Aims of the study .........................................................................................71
5.4 Methods .......................................................................................................71
Contents
X
5.5 Results ......................................................................................................... 72
5.6 Summary, discussion and conclusion ........................................................... 72
5.7 Paper 3: Quantification of the accuracy of MRI generated 3D models of long
bones compared to CT generated 3D models (in press) ................................ 74
Chapter 6. Higher field strength MRI scanning of long bones for generation of 3D
models ...................................................................................................... 83
6.1 Introduction ................................................................................................. 83
6.2 Theoretical consideration of increased SNR at 3T ........................................ 84
6.3 3T MRI for musculoskeletal system imaging ............................................... 84
6.3.1 Spin relaxation times and flip angle ................................................... 85
6.3.2 Fat suppression .................................................................................. 86
6.3.3 Magnetic susceptibility at 3T MRI..................................................... 87
6.3.4 Chemical shift at 3T .......................................................................... 87
6.3.5 MRI safety at 3T ............................................................................... 88
6.4 Aims of the study ......................................................................................... 88
6.5 Methods ....................................................................................................... 88
6.5.1 Samples ............................................................................................. 88
6.5.2 Measuring the quality of MR images ................................................. 88
6.5.3 Quantification of spin relaxation times .............................................. 90
6.5.4 Comparison of 1.5T and 3T imaging of musculoskeletal system ........ 91
6.6 Results ......................................................................................................... 93
6.7 Summary, discussion and conclusion ........................................................... 93
6.8 Paper 4: 3T MRI improves bone-soft tissue image contrast compared with
1.5T MRI (Submitted – under review).......................................................... 96
Chapter 7. Step artefact caused by Magnetic Resonance Imaging of long bone ... 121
7.1 Introduction ............................................................................................... 121
7.2 Motion artefact of MRI .............................................................................. 122
7.3 Aims of the study ....................................................................................... 123
Contents
XI
7.4 Methods ..................................................................................................... 123
7.5 Results ....................................................................................................... 124
7.6 Summary, discussion and conclusion .......................................................... 124
7.7 Paper 5: Correction of step artefact associated with MRI scanning of long
bones (Submitted – under review) .............................................................. 126
Chapter 8. Summary, conclusion and future directions......................................... 145
8.1 Summary and conclusion ............................................................................ 145
8.2 Future directions......................................................................................... 148
Appendix 1 Ethical approval for the study in Chapter 6......................................... 151
Appendix 2 Participant information and Consent form used in Chapter 6 ............ 154
Appendix 3 Animal tissue use notification ............................................................... 157
References ................................................................................................................ 159
List of figures
XIII
List of figures
Figure 2.1: Arrangement of the x-ray source, detector and the object in a CT scanner ...... 9
Figure 2.2: A spin possesses a tiny magnetic field aligned with the axis of rotation .........11
Figure 2.3: Spins aligned with the external magnetic field B0 ..........................................11
Figure 2.4: An MRI image of the coronal section of the proximal femur .........................19
Figure 2.5: The uniform regions of the external magnetic field of a MRI scanner (The
uniform region is shaded) .....................................................................................19
Figure 3.1: Average intensity values of the outer bone contours as detected by the Canny
filter for each axial CT image ...............................................................................27
Figure 3.2: The process of removing soft tissues from the sheep femur before scanning
with the contact mechanical scanner: a - gross dissection with the scalpel, ............34
Figure 3.3: Scanning of the bone's outer surface of the diaphyseal region using the MDX
20 contact scanner (The bone is positioned on the stage using glue tags)...............35
Figure 3.4: Bone is cut in three parts in order to scan the articular surfaces which cannot
be reached by the scanner on the intact bone .........................................................36
Figure 3.5: Positioning of the proximal articular segment of the femur in order to scan the
articular surface ....................................................................................................37
Figure 3.6: The reconstructed model before the scanning of articular surfaces (This model
was used as a guide to scan the articular regions) ..................................................37
Figure 3.7: Scanned surface with unusable data...............................................................38
Figure 3.8: The surface after removing the unusable data ................................................38
Figure 3.9: Two adjacent surfaces are fine registered ......................................................39
Figure 3.10: The final 3D model reconstructed by merging the surfaces ..........................39
Figure 3.11: a - The original microCT image (a cross section from the diaphysis); and b -
the image after applying a 20 × 20 median filter ...................................................40
List of figures
XIV
Figure 3.12: The initial aligning of the CT based 3D model to the reference model using
Trackball prior to the application of fine registration function .............................. 42
Figure 3.13: A CT based model (red) is aligned to the reference model (blue) in
Rapidform 2006 using the fine registration function ............................................. 42
Figure 3.14: Comparison of the aligned CT model to the reference model in Rapidform
2006 ..................................................................................................................... 43
Figure 3.15: Five anatomical regions used for the comparison: 1 - femoral head, 2 -
proximal region, 3 - diaphysis, 4 - distal region, 5 - distal articular region ............ 44
Figure 3.16: Reference planes and curves used for the splitting of the model into five
anatomical regions ............................................................................................... 44
Figure 5.1: Cross sections of CT (left) and MRI (right) from the same anatomical location
of a sample ........................................................................................................... 69
Figure 6.1: Positioning of the volunteer in the MRI scanner and the position of the matrix
coils that cover the lower limbs and the pelvis ...................................................... 92
Figure 6.2: Positioning of the field of view (FOV) on volunteer‟s leg ............................. 93
Figure 7.1: The step artefact caused by volunteer moving the leg between two successive
scanning stages................................................................................................... 121
Figure 7.2: MRI scanning of human lower limb with five scanning segments to scan the
complete limb .................................................................................................... 123
List of tables
XV
List of tables
Table 3.1 Specifications of the MDX 20 contact 3D scanner ...........................................35
Table 3.2 Scanner parameters used for microCT scanning ...............................................40
Table 6.1 TR and TE values used for the MRI scanning at 1.5T and 3T ...........................90
Table 6.2 Different flip angles used for scanning .............................................................90
Table 6.3 The protocols used for MRI scanning ...............................................................92
Publications, presentations and awards
XVI
Publications, presentations and awards
Journal Publications
1. Rathnayaka K, Schuetz MA, Sahama T & Schmutz B. Correction of step
artefact associated with MRI scanning of long bones, Submitted to Medical
Engineering and Physics.
2. Rathnayaka K, Coulthard A, Momot K, Volp A, Sahama T, Schuetz MA &
Schmutz B. 3T MRI improves bone-soft tissue image contrast compared
with 1.5T MRI, submitted to Magnetic Resonance Imaging.
3. Rathnayaka K, Momot K I, Noser H, Volp A, Schuetz M, Sahama T &
Schmutz B. Quantification of the accuracy of MRI generated 3D models of
long bones compared to CT generated 3D models. Medical Engineering &
Physics. 2011, in press, DOI:10.1016/j.medengphy.2011.07.027.
4. Rathnayaka K, Schmutz B, Sahama T and Schuetz M A. Effects of CT image
segmentation methods on the accuracy of long bone 3D reconstructions
Medical Engineering & Physics. 2011, 33(2): 226-233.
5. Schmutz B, Rathnayaka K, Wullschleger ME, Meek J, Schuetz MA.
Quantitative fit assessment of tibial nail designs using 3D computer
modeling. Injury. 2010; 41(2): 216-219.
Conference presentations1
1. Rathnayaka K, Cowin G, Schuetz MA, Sahama T, Schmutz B. Correction of
the step artefact in 3D bone models caused by the random movement of the
lower limb during MRI. 17th Annual Scientific Meeting, Australian & New
Zealand Orthopaedic Research Society. Brisbane, Australia, 1-2 September,
2011. (Oral presentation)
2. Rathnayaka K, Coulthard A, Momot K, Volp A, Sahama T, Schuetz M,
Schmutz B. Improved image contrast of the bone-muscle interface with 3T
MRI compared to 1.5T MRI. 6th World Congress on Biomechanics.
Singapore, 1-7 August, 2010. (Poster presentation)
3. Schmutz B, Rathnayaka K, Wullschleger M, Meek J, Schuetz M.
Quantitative fit assessment of tibial nail designs using 3 D computer
modeling. German Society for Orthopaedic and Trauma Surgery. Berlin,
Germany, 21-24 October, 2009. (Oral Presentation)
1 The conference abstracts have not been included in the thesis as the contents of them are covered by the journal articles
Publications, presentations and awards
XVII
4. Rathnayaka K, Sahama T, Schuetz MA, Schmutz B. Validation of 3D models
of the outer and inner surfaces of an ovine femur. 15th Annual Scientific
Meeting, Australian & New Zealand Orthopaedic Research Society.
Adelaide, Australia, 9-10 October, 2009. (Oral presentation)
5. Rathnayaka K, Momot K, Volp A, Noser H, Sahama T, Schuetz MA,
Schmutz B. Quantification of the accuracy of MRI generated 3D models of
long bones. 4th Asian Pacific conference of biomechanics. University of
Canterbury, Christchurch, New Zealand, 14–17 April, 2009. (Oral
presentation)
6. Rathnayaka K, Schmutz B, Sahama T, Schuetz MA. Effects of image
segmentation methods on the accuracy of long bone 3D reconstructions.
14th Annual Scientific Meeting, Australian & New Zealand Orthopaedic
Research Society. Brisbane, Australia, 17-18 November, 2008. (Poster
presentation)
7. Mohd Radizi S, Rathnayaka K, Pratap J, Mishra S, Schuetz MA, Schmutz B,
The effects of CT convolution kernels on the geometry of 3D bone models.
14th Annual Scientific Meeting, Australian & New Zealand Orthopaedic
Research Society. Brisbane, Australia, 17-18 November, 2008. (Poster
presentation)
Awards and Scholarships
1. Outstanding HDR student of the month, Faculty of Built Environment and
Engineering, Queensland University of Technology, December 2010.
2. Student travel grant awarded by 6th
World congress on Biomechanics.
Singapore, 1-7 August, 2010.
3. Joint winner of the Wilhelm-Roux-Preis 2009, at Annual conference of the
German Society for Orthopaedic and Trauma Surgery. Berlin, Germany, 21-
24 October, 2009.
4. Student Travel grant awarded by 15th
Annual Scientific Meeting of Australian
& New Zealand Orthopaedic Research Society. Adelaide, Australia, 9-10
October, 2009.
5. Runner-up for best poster presentation, IHBI inspires postgraduate student
conference. Gold Coast, Australia, 2-4 December, 2008.
6. QUT, Faculty of Built Environment and Engineering living allowance PhD
scholarship 2008-2011.
Authorship
XIX
Authorship
I declare that the work contained in this thesis has not been previously submitted to
meet the requirements for an award at this or any other higher education institution.
To the best of my knowledge and belief, the thesis contains no materials previously
published or written by another person except where due reference is made in the
text.
……………………………
Kanchana Rathnayaka Date:……………….
Rathnayaka Mudiyanselage
Acknowledgement
XXI
Acknowledgement
Firstly, I offer my sincere thanks to my supervisors: Dr Beat Schmutz for the
invaluable support, guidance and advice given throughout the PhD and for helping to
establish my directions; Prof. Michael Schuetz, my principal supervisor, for
encouragement and guidance given; and Dr Tony Sahama for introducing me to the
trauma research group at Queensland University of Technology and for the support
given throughout the PhD study.
I offer my special thanks also to: Dr Konstantin Momot for helping me by sharing his
knowledge of MRI physics and by reading manuscripts, especially during the second
and fourth parts of the research project; Prof. Alan Coulthard for collaborating with
me for the fourth part of the project; Dr Gary Cowin for helping me with MRI
scanning for the third part of the project; Mr Andrew Volp and Mr Russell Porter at
Princes Alexandra Hospital; Mr Raymond Buckley at Royal Brisbane and Women‟s
Hospital for MRI scanning of the samples and volunteers of the study; Mr Jit Pratap
at Princes Alexandra Hospital; and Ms Margaret Day at University of Queensland for
CT scanning of samples.
I must offer my sincere gratitude to all those who volunteered as subjects for the
study and spent their valuable time on my project, and to the National Imaging
Facility for providing me with 100% subsidised access to the 3T MRI scanner.
Thanks to all the researchers who donated ovine limbs from their studies and who
helped me to obtain them at the Medical Engineering Research Facility (MERF). I
would also like to thank the High Performance Computer (HPC) Unit and its
personnel at QUT for their help with the 3D modelling and use of the super
computers. Thanks to all the members of the trauma research group and all the
friends who helped me with various aspects of this research, especially with feedback
on writing and presentations. Finally, the laboratory and directorate staff at IHBI and
MERF also provided kind help during this project and I offer them my gratitude.
Abbreviations
XXII
Abbreviations 13
C Carbon 15
N Nitrogen 1H Hydrogen
31P Phosphorus
3D Three dimensional
3T Three tesla
B0 Main magnetic field
BW Bandwidth
CAS Computer assisted surgery
CNR Contrast to noise ratio
CT Computed tomography
FA Flip angle
FOV Field of view
H2O Water
HU Hounsfield units
ICP Iterative closest point
M0 Net magnetisation vector
MHz Mega Hertz
MR Magnetic resonance
MRI Magnetic resonance imaging
Mt Transverse component of net magnetisation vector
Mz Longitudinal component of net magnetisation vector
NAV Number of signal averages
NMR Nuclear magnetic resonance
NPA Number of acquired partitions
NPE Number of acquired phase encodes
PMMA Poly-methyl methacrylate (Dental acrylic)
RF Radiofrequency
ROI Region of interest
SD Standard deviation
SNR Signal to noise ratio
SNRGER Signal to noise ratio for gradient echo sequence
SNRSE Signal to noise ratio for spin echo sequence
T1 Longitudinal relaxation time
T2 Transverse relaxation time
TE Echo time
TMS Tetramethylsilane
TR Repetition time
V Voxel volume
Chapter 1: Introduction
1
Chapter 1 Introduction
The introduction of x-ray computed tomographic (CT) scanning and magnetic
resonance imaging (MRI) in the 1970s allowed medical personnel and researchers to
visualise the internal anatomical structures of the human body in three dimensions.
This allowed clinicians and researchers to reconstruct anatomical structures as
computer based three dimensional (3D) models and perform various experiments that
cannot be performed on living subjects. Thus, accurate reconstruction of 3D models
of anatomical structures from CT and MRI became a major research interest. Even
though the main mode of imaging bones is CT, the involvement of ionising radiation
leads clinicians and researchers to avoid CT whenever possible. Thus, a trend
towards the frequent use of MRI is developing among these groups, not only due to
the non-involvement of ionising radiation in MRI, but also due to its ability to
provide better quality images of soft tissue.
Reconstruction of a three dimensional computer model of an anatomical structure
using either CT or MRI imaging methods involves a number of complex processes:
data acquisition; segmentation of the region of interest (ROI) and surface generation
from the segmented volume. Each of these processes plays a crucial role in
determining the geometric accuracy of the reconstructed 3D model. Since the
geometric accuracy of 3D models is of high importance for most of their applications
(e. g. implant design and simulation of surgery), these processes in reconstructing 3D
models have drawn major attention from researchers [1-3]. While all steps play a
crucial role in determining the accuracy of 3D models, image segmentation is one of
the steps which has a higher human involvement and is thus vulnerable to errors.
Even though existing sophisticated segmentation methods are capable of minimising
the human intervention, most of these methods involve complex programming and
mathematics which many of the researchers are not trained to perform [2, 4-7].
Chapter 1: Introduction
2
Furthermore, these algorithms are designed to perform segmentation in a specific
anatomical region and, therefore, are not easily extended to the segmentation of a
different region due to their complex nature. Thus, a simple but accurate method for
medical image segmentation is a necessity.
Reconstruction of a 3D model of a small bone (phalanges or metatarsal bones) is
relatively easy when compared to the reconstruction of a 3D model of a long bone
that has a complex geometry. Thus, most of the studies that investigated
segmentation methods have utilised small bones. Nevertheless, 3D reconstruction of
long bones is important as most of the fracture fixation plates and intramedullary
nails are used for fixation of long bones. When 3D models of long bones are
reconstructed, the diaphyseal as well as the distal and proximal regions are equally
important. Most of the fracture fixation plates and intramedullary nails extend to the
proximal and distal regions (e.g. expert tibia nail used in chapter 4). The
intramedullary nail insertion point is usually in the proximal or distal region, thus,
accuracy of these regions are important to determine the entry point of the nail.
Furthermore, design of implants such as joint replacements needs highly accurate 3D
models of the proximal and distal articular regions. Therefore, the research projects
contained in the thesis will focus on all anatomical regions of long bones.
The decision to use either CT or MRI is mainly determined by the anatomical
structures being scanned. While CT visualises the bone tissue with better contrast,
MRI visualises soft tissues with better contrast as its main source of signal is
hydrogen nuclei which are abundant in soft tissues. The radiation exposure of CT
limits its utilisation to clinical cases and cadaver specimens. As most of the available
cadavers are more than 60 years old, the data acquisition is also limited to this age
group. However, approximately 51% of land transport trauma patients (or 11.4% of
total injury hospitalisations) in Australia during the 2006-2007 period were under 30
years of age. Furthermore, the study conducted by Noble et al 1995 [8] shows that
the femoral isthmus expands in old female population compared to the young
population. The study also showed that the medullary canal expands and the cortex
becomes thinner in old females and the CCD angle (femoral neck-shaft angle)
change with the age. These changes will impact the anatomical fitting of plates and
intramedullary nails designed using 3D models reconstructed from old bones. In
addition, osteophytes in old bones can significantly affect the anatomical fitting of
Chapter 1: Introduction
3
fracture fixation plates especially in ends of the bones. Thus, the acquisition of bone
data from this age group to inform the design of anatomically shaped fracture
fixation implants (plates and nails) for its trauma patients is of utmost importance
[9]. As MRI does not utilise ionising radiation, it is a potential alternative to CT for
acquiring bone data of volunteers from younger age groups.
Even though MRI visualises soft tissues with a high contrast, due to the extremely
short transverse relaxation times, bones generally do not generate a signal in MRI
[10-12]. However, using the signal generated by the soft tissues, bone geometry can
be delineated from the surrounding soft tissues and this has been demonstrated in the
literature [1, 13-17]. MRI has been used for the scanning of bones mainly in the case
of diagnosing metastatic disease, as MRI visualises metastasis with better quality
[18].
The use of MRI for 3D reconstruction of bones has been reported in computer
assisted surgery (CAS) and in foot bone motion quantification where the 3D models
of vertebrae and tarsal bones have been reconstructed [19-21]. Most of these studies
have used MRI for small bones with relatively simple geometry, and a proper
validation of the models has not been performed. Lee et al. used MRI to generate a
3D model of a porcine femur; however, the model has not been validated using an
accurate validation standard [1]. Therefore, before using MRI for 3D reconstruction
of long bones, a quantitative validation with an accurate reference standard is
necessary.
Some of the current limitations of the MRI scanning of long bones are long scanning
times and the difficulty of segmenting certain anatomical regions, conferred by poor
contrast between those anatomical regions and surrounding soft tissues. Since the
signal to noise ratio (SNR) of an MRI system is approximately directly proportional
to the main magnetic field of the scanner, higher field strength (3T) scanners promise
to offer an improved signal which can be converted to faster scanning times or better
image quality compared to the currently available (1.5T) scanners [22, 23]. The
improved image quality of 3T scanners has been demonstrated in a few studies for
computer assisted surgery and kinematic analysis of foot bone motion [24, 25].
However, the contrast at the bone muscle interface, which is more important for
segmentation of bones, has not been quantified and compared in those studies.
Chapter 1: Introduction
4
Furthermore, different contrast levels which occur in different anatomical regions of
a long bone need to be studied in detail to see the improvement in contrast at those
regions.
MR imaging of anatomical structures is challenged by various artefacts. Among
them, the motion artefacts due to random movements are of main concern in MRI
imaging of long bones due to their effects on the geometric accuracy of 3D models
reconstructed. In addition to the long scanning time of MRI, the non-uniformity of
the main magnetic field limits the effective scanning length, resulting in a long bone
being scanned in several stages. One of the adverse effects of this, which has been
observed in an initial study conducted by the supervisory team, is the displacement
artefact caused by the volunteer moving the leg between two scanning stages [26].
Thus, there is a step that can be seen on the final 3D model generated from such a
data set. This artefact may not be critical for clinical use of the images; however,
when the precise measurements are performed for implant design, these artefacts can
have a major effect on their accuracy. Therefore, minimisation or correction of these
artefacts can improve the accuracy of implants designed using those models.
This thesis presents the studies carried out to investigate: a simple and accurate
method for medical image segmentation; the feasibility of MRI as an alternative to
CT for scanning of long bones; the usability of higher field strength MRI to
overcome some of the problems with low field strength scanners; and the correction
of the step artefact that occurred from MRI scanning of long bones. Chapter 2
provides the basic physics involved in CT and MRI, while Chapter 3 provides the
background of image segmentation, 3D reconstruction and the investigation carried
out to develop and validate a simple and accurate image segmentation method.
Chapter 4 presents the application of 3D modelling techniques in implant validation,
utilising 3D models of long bones for fit quantification of two anatomically shaped
intramedullary nails. Chapter 5 presents the investigation carried out to formally
validate the MRI based 3D models of long bones against the CT based models.
Chapter 6 provides the details of the quantitative comparison between 1.5T MRI and
3T MRI. Chapter 7 presents the correction of the step artefact that occurred due to
the random movements of the volunteer during MRI scanning. Chapter 8 presents a
summary, discussion and future directions of the thesis.
Chapter 1: Introduction
5
The aims of the study in brief are as follows:
Investigation of the accuracy of multilevel intensity thresholding and Canny
edge detection for segmentation of CT images
Quantification of the accuracy of 3D models based on MRI compared to the
3D models based on CT
Quantitative comparison of the image quality at 1.5T MRI to 3T MRI
Correction of the step artefact that occurs due to the random movement of the
lower limb during MR imaging
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
7
Chapter 2 Quantitative imaging of the skeletal
system for 3D reconstruction
(Background)
2.1 Introduction
A number of methods are available for the imaging of various anatomical structures
of the human body, such as: plain x-ray, computed tomography (CT), Dual energy x-
ray absorptiometry (DEXA), magnetic resonance imaging (MRI) and ultrasound
(US). Even though quantitative imaging of the skeletal system is possible with most
of the above scanning methods, accurate spatially-resolved information of the
anatomical structures can only be acquired using CT or MRI. Thus, CT and MR
imaging methods have taken an integral part in research and in clinical applications
where the 3D reconstruction of the anatomical structures is required. The most
commonly used imaging technique for quantitative imaging of the skeletal system is
CT; however, MRI has also been reported as a potential imaging technique for this
purpose.
CT has become the gold standard of imaging the skeletal system for 3D
reconstruction because CT produces images with better contrast at the bone–soft
tissue interface. CT images can also be acquired within a very short period of time,
thus, essentially avoiding the motion artefacts caused by moving body parts or
tissues. While CT involves ionising radiation that prevents its use on healthy human
volunteers for research purposes, it can be used for in vitro research studies for
scanning of bones. In the present study, CT will be used to validate two image
segmentation methods and for validation of MRI based models of ovine femora.
Section 2.2 of this chapter provides the basic principles of CT and discusses its
advantages and disadvantages.
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
8
MRI utilises the principles of nuclear magnetic resonance (NMR) of hydrogen nuclei
to generate a signal from the tissue. Even though the bone tissues do not generate a
significantly large signal, by utilising the signal generated from the surrounding soft
tissues, MRI can be used in the 3D reconstruction of bones. Thus, MRI can
potentially be used to image healthy human volunteers for research purposes, without
having to expose them to the ionising radiation of CT. Section 2.3 discusses the basic
principles, advantages and current problems of MRI in detail. Since MRI has been
utilised for most of the studies presented in this thesis, it will be discussed in more
detail than CT.
2.2 Computed tomography (CT)
Computed tomography (CT) was the first method of imaging anatomical structures
inside the body without having the problem of the superimposition of anatomical
structures that was a major drawback of plain X-ray images. Since its introduction to
clinical use in 1970 [27], CT has become the most commonly used imaging
technique in the clinical setting. It has also become the standard practice for imaging
of trauma patients for accurate diagnosis of bone fractures in emergency situations
[28, 29].
2.2.1 Basic principles of CT
CT images are acquired by recording the object‟s attenuation of the radiation which
is emitted from an x-ray source (x-ray tube). A CT image is reconstructed from a
large number of projections of the object, taken around a single axis of rotation using
an x-ray beam. Depending on its x-ray absorption properties, when the x-ray beam
passes through the object, a projected image is generated on the detector (image
sensor). These images are integrated using a computer based algorithm to produce
axial image slices. The projections are obtained by rotating the detector and the x-ray
source simultaneously around the object (Figure 2.1).
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
9
Figure 2.1: Arrangement of the x-ray source, detector and the object in a basic CT
scanner (A large number of projections of the object will be obtained by rotating the
source and the detector simultaneously around the object.)
Early generation CT scanners imaged a patient slice by slice with specific slice
spacing. Once an image slice is obtained, the table with the patient moves a set
distance and the next slice is obtained. With the development of helical CT,
continuous imaging is performed by moving the patient continuously through the
gantry in combination with the continuous rotation of the x-ray source and detector
system. The obtained data volume is later reconstructed to image slices with specific
slice spacing. This also allows for the reconstruction of images in anatomical planes
other than the traditional axial image slices. Modern spiral scanners with multiple
rows of x-ray detectors (multi-slice scanners or multi-row scanners) can image a
subject within a very short time period (a few seconds), thus almost eliminating
motion artefacts.
Due to the high accuracy obtained for the bone geometry, CT has become the gold
standard for imaging of the bones for reconstructing 3D models, mainly for the
development of implants and clinical applications. CT can also be used for
measurement of relative tissue density and can be presented as Hounsfield Units
(HU) for comparison with other or reference tissues.
2.2.2 Radiation exposure during CT imaging
The use of diagnostic CT has increased dramatically over the last 20 years and it is
the gold standard for bone imaging. However, CT uses a high dose of radiation and
concerns have been raised regarding cumulative radiation exposure and associated
lifetime risk as there is epidemiological evidence of a small risk of radiation
associated cancer at doses comparable to a few CT scans [30-33]. For example, the
radiation exposure of a standard thoracic CT is equivalent to 400 standard chest x-ray
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
10
radiographs (8mSv), and that of a pelvic CT is equivalent to 250 chest radiographs
(250 mSv) [30, 34, 35]. According to a report by the Royal College of Radiologists
in the UK, CT scans probably contribute almost half of the collective dose of
radiation from all x-ray examinations [36]. This has become a major problem, as CT
scanning of a healthy human volunteer for research purposes is ethically not
justifiable due to this high radiation exposure.
As a solution, protocols that use low radiation doses while maintaining a higher
image quality are under investigation [37-39]. Some slice selection strategies (e.g.
use of fewer slices for simple geometric shapes such as diaphyseal region) have also
been investigated to reduce the radiation dose [38]. However, due to the fact that the
radiation exposure of CT cannot be eliminated completely, some countries do not
approve the scanning of volunteers with these protocols. Therefore, researchers are
moving towards using an imaging technique such as MRI that does not utilise
ionising radiation.
2.3 Magnetic resonance imaging (MRI)
2.3.1 Basic principles of MRI
Magnetic resonance imaging (MRI) utilises the nuclear magnetic resonance (NMR)
of 1H nuclei as the source of signal. There are a number of elements that demonstrate
NMR capabilities, such as 1H,
13C,
15N,
31P. Human tissue is largely composed of
water (H2O) and, thus, 1H is the most abundant NMR capable nuclei in the human
tissue. Throughout this discussion, 1H nuclei are also referred to as „spins‟, as
1H
nuclei have the quantum mechanical property termed „nuclear spin‟.
If a single 1H nucleus is considered, it possesses a magnetic moment, which is a
quantum mechanical property, parallel to its axis (Figure 2.2). In the absence of an
external magnetic field, the axes of the spins are randomly aligned in a given tissue
sample and the vector sum of the magnetisation is equal to zero (Figure 2.2).
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
11
Figure 2.2: A spin possesses a tiny magnetic field aligned with the axis of rotation
(left); randomly aligned axes of spins in the absence of an external magnetic field
(right) [40]2
To measure NMR of 1H nuclei (or any NMR capable nucleus), an external magnetic
field (also referred to as „the main magnetic field‟ or „B0‟) is applied to the sample,
thus making randomly aligned spins partially align with the externally applied
magnetic field (in the opposite direction to B0) (Figure 2.3). Thus, the sample now
possesses a net magnetisation vector (M0) parallel to B0. M0 can be split into its
component vectors: Mz which is parallel to B0, and Mt which is perpendicular to B0.
At rest, Mz = M0 and Mt = 0 (Figure 2.3).
Figure 2.3: Spins aligned with the external magnetic field B0 and M0 and its two
components, Mz and Mt.
2 Adapted from: Brown, M.A. and Semelka, R.C. MRI Basic principles and applications, 4th ed. 2010, New Jersey: John Wiley
& Sons
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
12
The frequency at which the spins precess under an external magnetic field is
proportional to the strength of the external magnetic field and is expressed by the
Larmor equation (Equation 2.1):
2
00
Bv 2.1
Where 0v is the Larmor frequency in megahertz (MHz), B0 is the external magnetic
field strength in tesla (T) and is a constant known as gyromagnetic ratio [40].
If an external radiofrequency (RF) wave with a frequency same as the Larmor
frequency of the spins (~64 MHz at B0 = 1.5T) is applied to the sample, some of the
spins shift from a low energy orientation to a high energy orientation. This moves M0
of the spins towards a direction perpendicular to B0 (if a 90º pulse is applied),
generating a net transverse magnetisation (Mt), and leading Mz to decline. At this
stage, the Larmor precession of the spins will induce a voltage in the receiving coil
(RF coil) which is measured as the MR signal. The intensity of the signal generated
in the receiver coil is proportional to the transverse magnetisation (Mt); therefore, the
initial magnitude of the signal depends on the value of the Mz immediately prior to
the RF pulse.
When excited, the angle at which M0 is oriented relative to B0 is the flip angle (FA),
which is one of the parameters that should be changed accordingly to get an optimal
contrast. When the RF wave is shut off, Mz starts to recover, and the inverse of the
rate constant of recovery is called the „longitudinal relaxation time‟ (T1). At the same
time, Mt starts to decay and the exponential rate constant of decay is called
„transverse relaxation time‟ (T2). Both T1 and T2 take different values for different
tissue types [41].
2.3.2 How tissue contrast is determined
In MRI, tissue contrast is related to the differences of rate of magnetisation decay.
The three factors that determined the tissue contrast in the present work were T1, T2
and the proton density of the tissue. The differences between spin relaxation times
and the proton density in different tissues serve as the basis for image contrast. The
contrast can be manipulated by selecting different scan parameters, namely repetition
time (TR) and echo time (TE). TR is the time period between two successive
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
13
acquisitions and TE is the time between delivery of the RF pulse and signal
detection. T1 contrast can be selected by choosing short repetition times (TR) and
such images are called „T1 weighted images‟ where the contrast is mainly determined
by T1 of the particular tissue. T2 contrast can be modulated by changing echo time
(TE) and the images of which the contrast is mainly determined by T2 are called „T2
weighted images‟. In both types of images, there is a contribution from T1 and T2,
however, the effect from one is minimised while the other is maximised. The
contrast can also be determined by the proton density of the tissue and the images
acquired this way are called „proton density weighted images‟.
2.3.3 Selection of slice position and thickness
The slice position, slice thickness and the Phase and Read directions are determined
by the respective gradient pulses and (in the case of the slice position) the RF
frequency offset. When a magnetic field gradient is applied on top of the existing
main field B0 in x, y, or z directions, the spins at different locations along the
gradient experience slightly different magnetic fields. Thus, the spins at different
locations along the gradient precess at different Larmor frequencies, which are given
by the following equation (2.2):
)( 0 ii rGBv 2.2
Where iv is the frequency of the spin at position ir , G is the gradient vector
representing the total gradient amplitude and the direction, B0 is the main magnetic
field and is the gyromagnetic ratio [40].
Three linear mutually perpendicular gradients are used: phase encoding gradient,
readout gradient and slice selection gradient. The phase-encoding gradient encodes
the locations of the nuclei in the direction of that gradient using the phase
accumulated by the nuclei during the gradient pulse. The readout (or frequency-
encoding) gradient encodes the locations of the nuclei in the direction of that gradient
using the position-dependent precession frequency during acquisition of the echo.
The receiver coils detect the entire spectrum of the different precession frequencies
during the readout gradient, which ensures that the field of view in the Read direction
covers the entire sample.
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
14
The slice selection gradient is used to achieve the localisation of the RF excitation to
a region in the space. The RF pulse applied has two parts: a central frequency and a
narrow bandwidth of frequencies (1-2 kHz). When such RF pulse is applied to the
sample in the presence of the slice selection gradient, a narrow region of tissue
achieves the resonance state. Thus the bandwidth of the applied RF wave determines
the thickness of the image slice.
2.3.4 Pulse sequences
The pulse sequence is a sequence of instructions to the hardware for switching the
RF pulse and gradient pulses on and off and for sampling the signal, keeping a
specific time period between each of them. This allows for the acquisition of data in
the desired manner by manipulating the relevant parameters (TR, TE, and FA). Spin
echo sequence and gradient echo sequence are two commonly used sequences for
clinical imaging. The FLASH (Fast Low Angle Shot) sequence used for this study is
based on the gradient echo sequence.
2.3.5 MRI safety
MRI is relatively safe compared to CT; however, the RF power deposition in the
conductive tissue results in heating of the tissue inside the body. To prevent hazards
from the heat, the specific absorption rates (SAR) of energy dissipation are
monitored using hardware level or software level monitors [42]. There are no known
direct biological hazards to patients from exposure to strong magnetic fields.
However, there is a high risk of the strong magnetic field of the scanner affecting
metallic implants and cardiac pacemakers. Thus, MRI is contraindicated for patients
with cardiac pacemakers, metallic debris in eyes or other ferromagnetic materials in
the body. Patients with implants that do not have a risk of detaching, or which do not
contain ferromagnetic materials (e.g. hip replacements, stents made of nickel-
titanium alloy) can be safely scanned with MRI [27].
2.3.6 Signal to noise ratio of an MRI system
Signal to noise ratio (SNR) is an important measure that can be used to quantify the
quality of a MRI system (Equation 2.3). In the case of conducting tissues, the
intrinsic SNR of a MRI system is approximately proportional to the strength of the
external magnetic field and the volume of tissue being scanned, and depends on
tissue parameters (e.g. T1 & T2). The following equations (2.4 & 2.5) show the
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
15
relationship between SNR and other parameters of a MRI system for a spin echo
sequence and gradient echo sequence [43]:
Noise
SignalSNR 2.3
21 )1(0
TTETTRAVPAPESE ee
BW
NNNVBSNR 2.4
2
1
1
cos)1(
)1(sin0
TTE
TTR
TTR
AVPAPEGER e
e
e
BW
NNNVBSNR 2.5
Where SNRSE = signal to noise ratio for spin echo sequence, SNRGER = signal to noise
ratio for a gradient echo sequence (FLASH), B0 = external magnetic field, V = voxel
volume, NPE = number of acquired phase encode lines, NPA = number of acquired
partitions, NAV = number of signals averaged, BW = receiver bandwidth per pixel, TR
= repetition time, T1 = longitudinal relaxation time, TE = echo time, T2 = transverse
relaxation time and θ = flip angle.
In both equations, the term under the square root is the total time for acquiring data.
Therefore, intrinsic SNR is directly proportional to the strength of the external
magnetic field, the voxel volume, the square root of total sampling time and contrast
related parameters. Thus, from the above relationship, it is clear that the external
magnetic field, voxel size, number of averages, flip angle, T1, T2, TR and TE all have
an influence on SNR of a MRI system. In addition, sensitivity to magnetic
susceptibility and chemical shift difference between fat and water also influence the
SNR of a MRI system.
2.3.7 Artefacts of MRI
When the pixels in the MR image do not represent the actual anatomical structure
being scanned, this region of the image is referred to as an „artefact‟. These artefacts
appear among the general structures as signals that do not correspond to the actual
tissue at the location. They may or may not be easily recognised from the normal
anatomy, particularly if they are low in intensity.
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
16
2.3.7.1 Motion artefacts
Motion (also referred to as movement) artefacts occur as a result of movement of the
tissue (heart, lung) or parts of the body (limbs) which are being scanned during the
data acquisition. Motion artefacts can either be due to periodic movements (e.g.
blood flow, respiration and heart beat) or random movements which mainly occur
due to the person‟s inability to keep the body parts still during the long scanning
time. These movements result in misregistration of pixels along the phase-encoding
direction [40, 44]. The artefact occurs by tissue that is excited at one location
producing signals that are mapped to a different location during detection [40]. The
nature and the extent of the artefact depend on the extent of movement and the
protocol used for scanning.
The most common motion artefact caused by periodic movements is due to blood
flow in the vessels of the tissue being scanned [40]. If the blood flow is in a direction
perpendicular to the slice plane, the artefact is localised to the vessel diameter. If the
flow is along the slice plane, a more diffuse artefact is seen.
Motion artefacts from random movements occur due to muscle contraction from
nerve excitations. They can also occur as a result of the volunteer or patient
randomly moving the body part being scanned due to the longer scanning times (e.g.
keeping a lower limb still for 65 minutes is nearly impossible). Since the complete
lower limb has to be scanned in several segments, volunteers tend to move the leg
between segments and this causes a step in the final image stack.
Motion artefacts that occur due to periodic movements such as breathing movements
can be minimised by using specially designed protocols which synchronise the data
acquisition with the breathing movements, or by post processing techniques.
Elimination of the artefacts occurring due to random movements is, however, more
difficult to achieve through such methods.
2.3.7.2 Magnetic susceptibility difference artefact
Magnetic susceptibility ( ) is the response of a substance to the applied magnetic
field. There are three levels of responses that have been described: diamagnetic,
paramagnetic and ferromagnetic. The diamagnetic response arises from the electrons
surrounding the nuclei, while the paramagnetic response arises from molecules that
have unpaired electrons. Both these responses are relatively weak responses and
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
17
materials with such responses are safe to be used in MRI. However, the
ferromagnetic response is found in certain ferrous metals and the magnetic
susceptibility due to this response is very large. The relationship between magnetic
susceptibility, external magnetic field and net magnetisation vector is expressed as
the equation below:
00 BM 2.6
Where M0 = net magnetisation vector, = magnetic susceptibility and B0 = external
magnetic field [40].
The artefact is generated due to the different magnetic susceptibility of two adjacent
tissue types. Cortical bone has a low magnetic susceptibility, while soft tissues have
larger magnetic susceptibility. Thus, at the interface between soft tissue and bone, a
considerable change in the local magnetic field present causes a significant signal
loss.
2.3.7.3 Chemical shift
Chemical shift is the difference in precessional frequency conferred by the magnetic
shielding effect of the electron clouds that surround protons within tissues, relative to
that of a standard reference compound (in the case of protons tetramethylsilane (
TMS)). Basically, in MRI there are two sources of 1H nuclei, water and fat. Water
has two H atoms bonded to one oxygen atom, while fat has many H atoms bonded to
a long-chain carbon framework. Due to this difference, protons from water have a
different local magnetic field than protons from fat which is called „magnetic
shielding‟. This magnetic shielding effect causes the protons from two sources to
precess at different frequencies. This, in turn, causes fat and water protons from the
same tissue location map to different positions in the reconstructed image. The
difference of precessional frequency between water and fat at 1.5T is approximately
220Hz.
2.3.8 MRI for imaging of the skeletal system
MRI is designed to scan soft tissues utilizing 1H nuclei as the source of signal and,
thus, is not routinely used for imaging of bones. However, by using the signal
generated from the surrounding soft tissue, bone outer geometry can be quantified
from MRI images. This will be discussed in more detail in Chapter 4.
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
18
2.3.9 Advantages and current limitations of MRI
Absence of ionising radiation is an important advantage of MRI over CT, as this
allows researchers to scan healthy human volunteers without exposing them to a high
dose of ionising radiation. However, MRI has some limitations compared to CT
when used for scanning of long bones, such as longer scanning times, poor contrast
in certain anatomical regions, non-uniformity of the magnetic field, limited
availability, and higher cost per scan.
2.3.9.1 Longer scanning times of MRI
Longer scanning time is the most important limitation of MRI when it is used for
scanning of clinical cases as well as for research. As an example, in this study,
scanning of a human lower limb with a modern 64 slice helical CT scanner takes less
than ten seconds of scanning time, while an MRI scanner takes more than one hour
for the same scan. This longer scanning time of MRI makes the images of moving
(breathing) body parts vulnerable to motion artefacts.
2.3.9.2 Poor contrast in certain anatomical regions
The next important limitation of MRI is the poor contrast of MRI images in certain
anatomical regions of the bone (Figure 2.4). In the human body or other
mammalians (sheep), the diaphyseal region of long bones is covered mostly with
muscles. However, the distal and proximal regions of the bone, on the other hand, are
mostly covered with ligaments, joint fluid, joint capsule and cartilage. These
different soft tissue types have different MRI properties and, depending on the
chosen scanning parameters, some generate poor or no signal, thus making them
indistinguishable from cortical bones (e.g. ligaments, cartilage). Thus, the
demarcation between such soft tissues and the cortical bone cannot be clearly defined
and a complete 3D model is generated by making an educated guess or by
interpolating the available data. This educated guessing or interpolation of the
regions introduces errors to the 3D models.
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
19
Figure 2.4: Left: an MRI image of the coronal section of the proximal femur
(showing that the shaft region has a good contrast between cortical bone and the
muscles, while the regions indicated by arrows are not clearly defined), right: the
corresponding CT image (showing the well defined boundary of cortical bone)
2.3.9.3 Non-uniformity of the external magnetic field
The external magnetic field used by MRI scanners is not uniform throughout its
length (Figure 2.5). Due to this non-uniformity of the magnetic field, the signal of
the MRI images tends to distort towards the ends of the magnetic field, thus limiting
the effective scanning length of the scanner to about 30 - 40 cm. Therefore, long
samples (such as human lower limbs) have to be scanned in several stages; this
involves moving the table to position different parts of the sample in the centre of the
magnet.
Figure 2.5: The uniform regions of the external magnetic field of a MRI scanner
(The uniform region is shaded)
2.3.9.4 Limited accessibility
The accessibility of MRI scanners for research is mainly determined by the cost and
the clinical work load of the scanner. The cost of a MRI scan is considerably higher
than the cost of a CT scan. Due to the longer scanning times, MRI scanners in
clinical use are heavily booked for scanning of patients. Few scanners are dedicated
for research purposes. With the increased clinical use of 3T MR imaging, more
Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)
20
scanners will become available and the availability of scanners for research purposes
will potentially be increased.
2.4 Summary
The two available scanning methods for quantitative 3D imaging of long bones are
CT and MRI. CT and MRI both provide accurate information for quantifying
anatomical structures in a 3D environment. Both imaging methods have certain uses
in clinical application, with CT mainly being used for bone imaging and MRI for soft
tissue imaging.
CT uses ionising radiation for its scanning and, therefore, its use in research is
generally limited to scanning of cadaver specimens or clinical cases. While CT has a
number of advantages such as a high contrast in bone–muscle interface and faster
imaging times, it cannot be used for scanning of human volunteers.
MRI utilises NMR of the 1H nuclei as the source of signal for imaging. Hence, the
theoretical use of MRI is limited to imaging of soft tissues. However, MRI has the
advantage of not using ionising radiation and is therefore well suited for scanning of
healthy human volunteers for research purposes. MRI has some limitations such as
very long scanning times, poor contrast in certain anatomical regions and shorter
scanning length due to non-uniformity of the magnetic field. Limitations such as
longer scanning times and poor contrast in certain anatomical regions can be
overcome to some extent by using an external magnetic field with a higher strength,
as later demonstrated by the study conducted in Chapter 6.
Chapter 3: Image processing and surface reconstruction
21
Chapter 3 Image processing and surface
reconstruction
3.1 Introduction
The reconstruction of 3D models of bones using CT imaging has become an interest
of current medical engineering researchers, as 3D models of bones are increasingly
being utilised for various practices of clinical medicine and medical research; for
example, in the design of orthopaedic implants [45-50], in the computer aided
planning of surgery [51-54], in fracture healing models [55, 56] and in finite element
methods for fracture load analysis and bone strength analysis [57-61]. This interest is
not only due to the wide utilisation of 3D models, but also because the generation of
an accurate 3D model is a complex process. This process involves several stages
where accuracy of the 3D model can be highly affected by various factors in each
stage.
The process of reconstructing a 3D model from a bone can be categorised in terms of
data acquisition, image segmentation and surface generation [62]. Data acquisition is
conducted using a tomographic imaging method such as CT or MRI. Image
segmentation is the process of separating the identified ROI. The surface generation
of the segmented volume is performed automatically using an algorithm, and is one
step that determines the sub-voxel level accuracy of 3D models. While all steps play
a crucial role in generating an accurate 3D model, it is the segmentation step that is
most user-dependent and thus vulnerable to operator introduced inaccuracies. Thus,
an accurate image segmentation method is of utmost importance for generating 3D
models with correct geometric representation of the actual bones.
Among the various segmentation methods available, intensity thresholding and edge
detection are two simple image segmentation methods commonly used in medical
Chapter 3: Image processing and surface reconstruction
22
image segmentation. This study investigates intensity thresholding and Canny edge
detection as simple but accurate segmentation methods for long bone image
segmentation that can be used by the general research community who do not have a
background in the complex programming and mathematics involved in segmentation.
The next section discusses the relevant literature and the processes involved in
reconstructing 3D models of bones from a CT or MRI data set. From Section 3.8
onwards the description will be focused on the 3D modelling methods used in this
study. The image segmentation methods investigated and the surface generation and
3D model manipulation techniques described in this chapter are used in all the
projects that are included in this thesis.
3.2 Acquisition of data for 3D modelling of bones
In imaging, data acquisition is the process of obtaining a digital representation of the
anatomical structures. While this can be achieved using various acquisition
techniques, the acquisition of data from living subjects for 3D reconstruction of
bones is done by using CT or MRI imaging, as other methods cannot generate 3D
spatially resolved information of the anatomical structures. CT and MRI can also be
used for acquisition of the image data from cadaver bone specimens, and CT is the
gold standard for this process. The soft tissue free cadaver bones can be scanned with
contact mechanical scanners or optical 3D scanners. In certain countries CT cannot
be used to scan healthy humans for research due to the high amount of radiation
involved in CT.
The accuracy of the data acquired from bones depends on the type of imaging
method, the accuracy of the hardware used and the set imaging parameters.
Adequately calibrated hardware and optimally set scanning parameters are necessary
for accurate acquisition of data from any anatomical structure. The calibration of the
hardware has usually been conducted at the factory and the recalibration is conducted
periodically by scanning phantoms. The scanning parameters vary with the imaging
modality (e.g. MRI or CT) and can be adjusted depending on the structures to be
visualised.
For reconstruction of 3D models, the data should be acquired as spatially resolved
information of the anatomical structures. Tomographic imaging techniques such as
Chapter 3: Image processing and surface reconstruction
23
CT and MRI are capable of obtaining such data of anatomical structures. Generally
in CT or MRI, 3D volumetric data is presented as axial image slices with a certain
user defined thickness. A single image slice is composed of a two dimensional array
of elements called „pixels‟ and with the thickness added, these elements are called
„voxels‟. The voxel basically represent the average signal or intensity of the tissues
contained in it. The size of a voxel is determined by the field of view (FOV), size of
the image matrix and the slice thickness.
3.2.1 Effect of in plane resolution and slice thickness on accuracy of
reconstructed 3D models
In order to obtain an accurate representation of the anatomical structures being
scanned, a voxel should be sufficiently small in size. When the voxel size becomes
larger, it contains the average signal/attenuation from larger tissue volume and more
tissue types. Hence, larger voxel size (low resolution) results in higher inaccuracies
in 3D models due to inadequate representation of the anatomical structures. This
mainly affects the scanning of thin structures (e.g. distal and proximal regions of the
cortex of a long bone) where the thickness is less than the voxel size. In the case of
CT, this results in overestimation of the thickness of the structure and
underestimation of its density [63, 64]. This produces 3D models that do not
accurately represent the surface geometry affecting the implants generated using
such models [65, 66]. In addition, the bone density properties acquired from such
data adversely affect the accuracy of the FE models.
This inaccurate representation of anatomical structures by pixels is called „the partial
volume effect‟. The partial volume effect appears when one element (voxel) is filled
by tissues with different attenuation properties for which the mean attenuation is
calculated [67]. Appearance of this effect in the bone–muscle interface makes the
separation of the bone a relatively difficult process requiring more robust
segmentation methods. Generally, this effect can be minimised using a smaller voxel
size [68]. However, the complete elimination of this effect is not possible. In
addition, acquisition of the image data with smaller voxel sizes is done at the expense
of imaging time in the case of MRI, or of exposing the subject to a high radiation
dose in the case of CT.
Chapter 3: Image processing and surface reconstruction
24
The slice spacing or thickness also has an effect on the reconstructed 3D models,
especially when relatively complex shapes have to be reconstructed (e.g. femoral
head) compared to simple shapes (e.g. diaphysis of a long bone) [69-71]. This effect
can be minimised using a smaller slice spacing or thickness, especially where the
geometric shapes are complex [69]; again, however, this is at the expense of imaging
time in MRI or radiation exposure in CT.
3.3 Image segmentation
Image segmentation is the process of separating or partitioning the image into
meaningful entities by defining boundaries between features and objects of an image
based on intensity or texture criteria [72-74]. In medical image segmentation, prior
knowledge of anatomy is used to identify the structures being considered [5, 75]. In
the process of generating 3D models, image segmentation is a crucial step in
determining the accuracy of the segmented region. The accuracy required of the
segmented region varies depending on the purpose. For example, designing an
anatomically pre-contoured fracture fixation plate does not require the same level of
accuracy as is required for the finite element analysis of a bone model for stress
analysis of an intramedullary nail-bone construct.
In image segmentation, automatic processing is sometimes desirable as this can
minimise operator involvement and reduce manual processing time. This is not
always attainable due to the limitations imposed by the image acquisition and the
complexity of the anatomical structures [73]. The articular regions of the long bones
are often covered by a mixture of different types of tissues including cartilages, joint
capsules, synovial fluid, ligaments, fat tissue, tendons and muscle. These different
tissue types have different imaging properties and some of them are nearly
impossible to be differentiated from the bone. For instance, the articular cartilage is
not visible in CT images, while in MRI it is visible but often cannot be differentiated
from the bone [15, 17]. Thus, depending on the segmentation method, considerable
manual processing time is required to segment the articular regions of a long bone.
Therefore, some of the segmentation methods are prone to operator introduced
errors.
Due to the difficulties of segmentation conferred by the partial volume effect and the
complex anatomical structures, a large number of segmentation techniques that can
Chapter 3: Image processing and surface reconstruction
25
be used for segmentation of medical images of various anatomical regions have been
reported [6, 76-82]. These vary from simple to complex methods that involve highly
sophisticated mathematical algorithms as well as programming techniques [2, 4, 6,
76, 81, 83-85]. Among the various methods, manual segmentation and thresholding
are relatively simple segmentation techniques commonly used for medical image
segmentation compared to the region growing, artificial neural networks (ANN) and
fuzzy logic based techniques where highly sophisticated programming and
mathematics have been used.
3.3.1 Manual segmentation
Manual segmentation is by far the simplest segmentation method available for
medical image segmentation [83, 86]. The region of interest is manually delineated
using a simple image editing or painting software program. Hence, there is no need
for complex programming or software packages. The tracing of the ROI is usually
carried out by a person with a good knowledge of both the anatomy of the desired
region and image segmentation.
This method is prone to inter- and intra-operator variability and the accuracy of the
segmented region always depends on the knowledge and experience of the person
who performs the segmentation [87]. The method is also more labour intensive and
time consuming than the other segmentation methods available. Manual
segmentation also has a poor repeatability compared to other segmentation methods
and is not suitable for applications where high accuracy and repeatability is expected.
3.3.2 Intensity thresholding
Intensity thresholding is a commonly used segmentation method for medical image
segmentation that has been implemented in most of the commercially available
image processing software packages [74, 88]. With this technique, the group of
pixels (ROI) that has the intensity value above a set threshold level is assigned to one
class while the rest (the background) is assigned to another class. Thus, a binary
image that contains the ROI and the background is generated. In its basic form, this
technique often relies on the user visually selecting a threshold level, thus making the
method vulnerable to user dependent errors and less repeatable [89]. In addition, one
threshold level does not accurately segment a complete long bone, as different
regions of the bone have different intensity levels (Figure 3.1).
Chapter 3: Image processing and surface reconstruction
26
3.3.2.1 Selecting an appropriate threshold level
The simplest method of selecting an appropriate threshold level for intensity
thresholding is the visual selection of the threshold level. This is usually achieved by
selecting a threshold level which reasonably selects the ROI without under- or
overestimating it. Accuracy of the selected threshold level varies depending on
several factors such as the window level setting of the display, and the knowledge
and experience of the person. Therefore, this method is not appropriate in
applications for which higher accuracy and repeatability are expected.
Due to these drawbacks of visually selecting a threshold level and the unavailability
of a standard method of selecting an appropriate level, various methods have been
investigated [89]. Histogram based selection of threshold level [90-93] and clustering
of grey levels of the boundary [94] are two of these methods. Most of the methods
are not highly repeatable, and some involve complex programming that limits their
use by a person with little knowledge of programming. Therefore, a repeatable and
simple method of selecting a threshold level is required.
3.3.2.2 Multilevel thresholding
Intensity thresholding is generally conducted using one threshold level to segment
the complete bone or the region (global thresholding). However, global thresholding
often fails to segment the complete bone accurately due to intensity inhomogeneity
of the different regions of the bone [92, 95]. For example, the proximal, diaphyseal
and distal regions of a long bone (femur or humerus) have different threshold levels,
as illustrated in the graph below (Figure 3.1). Thus, the use of one threshold level to
segment the complete bone will necessarily under- or overestimate the regions with
different threshold levels.
Chapter 3: Image processing and surface reconstruction
27
Figure 3.1: Average intensity values of the outer bone contours as detected by the
Canny filter for each axial CT image (350 slices) of the bone (The intensity values at
the arrow locations were used to calculate the average thresholding values used for
segmentation of each anatomical region) (HU = Hounsfield Units)
The graph was obtained by plotting the threshold values calculated for each image
slice against the slice number. The Canny edge detector based threshold selection
method developed as a part of this research project was used to calculate the
threshold value for each image slice. The graph shows that the threshold level for the
diaphysis is fairly constant throughout its length, but the image slices of the proximal
and distal regions have relatively low non steady threshold levels. For this reason,
using a single threshold level to segment the complete bone will lead to inevitable
inaccuracies of the segmented bone model.
Thresholding the bone using more than one threshold level for regions with different
threshold levels will segment the complete bone accurately and this has been
successfully tested on small bones [95]. However, studies using multiple threshold
levels for the segmentation of the complete long bones have not yet been reported in
the literature and this will be investigated in the present study.
Chapter 3: Image processing and surface reconstruction
28
3.3.3 Edge detection
Edge detection algorithms identify the rapid change of intensity level in a small
neighbourhood of pixels in the image [96]. This is a popular segmentation method
which has been used in cardiac and other medical image segmentation [76, 97]. As
the algorithm considers local change of the intensity of a small region, this would be
ideal for segmentation of an object with different intensity levels in different regions
such as a long bone (e.g. femur). Positioning of the edge relative to the actual
boundary of the object basically depends on the sensitivity of the edge detection
algorithm used. Some of the algorithms allow users to change the sensitivity of the
algorithm by choosing a threshold level.
Edge detection has a higher repeatability compared to other segmentation methods as
human intervention can be kept to a minimum level. However, this method is
susceptible to artefacts and, more often, intensity changes due to noise are also
detected as edges. The edge detection algorithms are necessarily complex programs;
however, most of the algorithms are built into many image processing software
packages (e.g. Matlab and IDL) and can be used easily. Among the number of edge
detection algorithms available (such as Roberts and Sobel), Canny is an accurate,
reliable and faster edge detector for image segmentation [98-101]. Therefore, the
Canny edge detector was selected to investigate segmentation of long bones in the
present study.
3.3.4 Region growing
Region growing is a method of segmenting image regions or features that are
connected, using pixel neighbourhood operations [102]. Starting from a user defined
seed point, the region grows around it, extracting all the pixels connected to the seed
point until the set criteria are met [74]. Intensity thresholding is often used in
combination with region growing to segment the image features that are connected. It
has also been used in skeletal system image segmentation [103]. The accuracy of the
segmented region depends on the set criteria and this method often fails when used to
segment complex structures such as long bones.
Chapter 3: Image processing and surface reconstruction
29
3.3.5 Sophisticated segmentation methods
Most of the segmentation methods discussed above are basically implemented in 2D,
in which each of the image slices is processed individually. Thus, a considerable
amount of labour and time is required to segment a complete 3D volume. Therefore,
fast and automatic segmentation methods have been investigated over the past few
years. As a result, there are a number of robust image segmentation methods
available for medical image segmentation. These methods carry out the segmentation
process automatically, minimising the human intervention. In some of the
techniques, simple methods such as intensity thresholding or edge detection have
been used with modifications to automate the process, while other methods involve
techniques such as artificial neural networks and fuzzy logic [6, 7, 81, 104-106].
These segmentation techniques utilise advanced programming techniques and
mathematical algorithms, making them unavailable to the general research
community with little knowledge of complex mathematics and advanced
programming techniques. In addition, most of these techniques have been tested on
smaller bones or part of a long bone which has relatively simple geometry compared
to a human long bone [4, 107, 108]. Thus, these methods have the potential to fail or
produce inaccurate results when used to segment a long bone with complex
geometry, where image segmentation is particularly difficult due to restrictions
imposed by image acquisition and anatomical structure variations. Therefore, further
investigations using long bones are necessary before applying these methods on
segmentation of long bones.
3.4 Surface generation
The generation of triangular meshed 3D surfaces from the segmented boundary
voxels is as important as the image segmentation, as the sub-voxel level accuracy of
the 3D surfaces is mainly determined by this process [109]. This process also
determines the number of triangles, their consistency, and their accuracy on the
surface. The surface generation is usually carried out using one of the algorithms
available [110, 111]. The marching cube algorithm (or its derivatives) is one of the
popular algorithms that have been used in most of the commercially available
software packages [62, 112]. In these packages, the surface generation is usually
Chapter 3: Image processing and surface reconstruction
30
carried out automatically, with the user being limited to setting the level of
smoothing applied.
3.5 Registration (aligning) and comparison of surfaces
Aligning of 3D surfaces is often used in research that involves 3D model
manipulation. This aligning of the surfaces is first performed manually and then a
surface matching algorithm is used to fine-align the surfaces. The iterative closest
point (ICP) algorithm is a commonly used robust method for registration of 3D
objects [1]. This algorithm has been used successfully in the literature with a high
accuracy [113]. Lee et al. [1] conducted a registration test using the ICP algorithm in
which a part of the bone model which had separated from the original model was
matched perfectly to its original full model. The algorithm has also been
implemented in most of the 3D modelling software packages and, therefore, is easily
accessible.
After the registration of the 3D surfaces, the comparison is usually carried out by
calculating the deviation of the surface of interest from the reference surface. This is
conducted using a point to point comparison method where the normal distance from
a point of the surface of interest to a corresponding point of the reference surface is
calculated.
3.6 A reference standard for validating 3D models of bones
Validation of 3D models plays an important role in the studies that involve 3D
models of bones, especially when live subjects have been used, where the physical
bone is not readily accessible. A number of methods have been established over
time; however, none of these is accepted as a standard method for validation of 3D
models of bones.
Amongst the methods used, models manually segmented by anatomy experts have
been used to validate the 3D models [4]; however, the accuracy of this method is
highly dependent on the experience and knowledge of the person who conducts the
segmentation. 3D laser scanning of the bone‟s surface has been used in several
studies [3, 69, 83, 114]. In this method, an outer coating has been applied on the
bone‟s surface; however, this coating might introduce errors to the scanned surface
unless applied evenly. Mechanical digitisers or digitising arms are other options for
Chapter 3: Image processing and surface reconstruction
31
digitising the bone surface. There are studies that report on the use of mechanical
digitising arms; however, these devices are not capable of generating an evenly
distributed mesh unless they move automatically [115].
One of the important limitations of laser and mechanical digitisers is that none of
these methods can be used to validate the internal medullary canal of a long bone. As
the medullary canal is 1-2 cm in diameter, the scanner head of a digital scanning arm
or the laser scanner cannot reach the inside of this canal. Even though scanning of
the cut-opened bone canal is possible, the bone loss from the bone saw (0.5 -1.0 mm)
is inevitable and this can lead to errors in the final 3D model. In the present study, an
attempt was also made to model the medullary canal using dental acrylic (PMMA);
however, this was not possible as the material shrinks when it solidifies and also
generates air bubbles, thus causing some regions of the PMMA mould to lose contact
with the bone.
Goyal et al. used MicroScribe digitiser–a mechanical arm with a stylus–to capture
3D points from tibial surface fitted with a plate [46]. It has an accuracy of up to 0.23
mm and sampling rate of 1000Hz. The study used the scanning arm only to record
the position of the plate and tibiae and did not generate the complete 3D model.
Gelaude et al. used a laser strip scanner which measures the distance to an object
from the scanner head [3]. In combination with a coordinate measuring system, this
was used to generate a reference standard for the soft tissue free human femora,
obtaining an accuracy of 0.70 ± 0.55 mm when compared with CT derived models of
the same samples. DeVries et al. also used a laser scanner (Roland LPX-250) to
validate phalanx 3D models. The scanner was used with a resolution of 0.2 mm and
there was an average 0.2 mm deviation from the manually segmented CT based 3D
models.
Considering the advantages and disadvantages of the methods described, the
following two methods were used to validate the 3D models of long bones
reconstructed from CT and MRI data of ovine femora. A contact mechanical scanner,
which automatically moves along the object being scanned, is a good option for
accurate digitisation of a denuded bone‟s outer surface. The scanner moves
automatically along a pre-defined mesh, essentially generating an evenly distributed
mesh that cannot be achieved with mechanical digitising arms. A MicroCT scanner is
Chapter 3: Image processing and surface reconstruction
32
capable of scanning an object with a very high resolution (e.g. 30 µm isotropic voxel
size). This also has the ability to digitise the inner medullary canal, as well as
complex geometric shapes that cannot be digitised with methods such as laser or
mechanical digitising arms. Therefore, this is an ideal method for the validation of
the medullary canal of long bones.
3.7 Aims of the study
This study specifically aimed to:
Investigate the accuracy of multilevel intensity thresholding as a method of
segmenting CT data of long bones in combination with a new threshold level
selection method
Investigate the accuracy of Canny edge detection for segmentation of CT data
of long bones
Compare the accuracy of multilevel intensity thresholding and Canny edge
detection to single level thresholding
3.8 Methods
3.8.1 Samples
Five intact cadaveric sheep hind limbs, amputated from the pelvis, were obtained
from four Merino-Cross sheep. However, the statistical analysis of the sample size
for 80% power shows that 28 samples are needed to detect a difference of 0.06 mm
with SD = 0.04 and 0.015. Due to the long processing time of the samples, it is not
practicable to use these sample sizes and therefore, a sample size of 5 has been used.
Using the sample size of 5 the difference that can be detected is 0.108 mm with the
same standard deviations.
3.8.2 Image segmentation
Three image segmentation methods were investigated in this research project for
segmentation of CT data: single-level intensity thresholding; multilevel intensity
thresholding and Canny edge detection. Single level thresholding was performed for
the purpose of comparing this method with the other two methods.
Multilevel thresholding was used to overcome the problem of over- or under-
estimating the regions with different threshold levels when a single threshold level is
Chapter 3: Image processing and surface reconstruction
33
used. A method of selecting an appropriate threshold level was also used with the
multilevel thresholding. This threshold selection method was used to reduce the user
dependent errors of visually selecting a threshold level. The threshold selection
method was based on the calculation of average intensity of the edge that is detected
by the Canny edge detection filter.
In multilevel thresholding, the Canny edge detector was used to determine the
threshold level utilising Canny edge detector‟s higher repeatability as a superior
method to visual selection of the threshold level. Thus, the multilevel thresholding
and Canny edge detector methods were expected to be similarly accurate for
segmentation of CT image data.
The Canny edge detection filter was used to delineate the outer and inner cortex from
the bone as the third segmentation method. Canny edge detection was performed in
2D axial images, and then the outer and inner edges of the bone cortex were
delineated using a customised Matlab script. These edges were later combined to
reconstruct the 3D models of the outer and inner cortex of the femur. A detailed
section of the segmentation methods used in this part of the research project is
available in the paper presented at the end of this chapter.
3.8.3 Reference model for validation of the outer 3D models
Validation of the outer 3D models was carried out using a contact mechanical
scanner (MDX-20 Roland) to digitise the outer surface of the bone. The complete
process involved the prior removal of the soft tissues from the bone and then
scanning of the outer surface in several steps, generating a number of surfaces.
Finally, the reference 3D model was reconstructed by merging the scanned surfaces.
3.8.3.1 Removal of the soft tissues from long bones
Various methods have been reportedly used for removing soft tissues from bones,
such as boiling or use of chemicals to dissolve the tissues [3]. These methods have
the risk of changing the outer geometry of the bone and therefore were not used in
this study. The removal of soft tissues before the scanning with the contact
mechanical scanner was achieved by dissecting the limb with a scalpel. After the
bone has been harvested, the scalpel blade was used to carefully remove the soft
tissue attached to the bone, without damaging the bone‟s outer geometry (Figure
3.2).
Chapter 3: Image processing and surface reconstruction
34
Figure 3.2: The process of removing soft tissues from the sheep femur before
scanning with the contact mechanical scanner: a - gross dissection with the scalpel,
b - removing soft tissues attached to the bone, and c - soft tissue free bone
3.8.3.2 Scanning of the bone’s outer surface using the contact scanner
A mechanical 3D contact scanner (Roland DG Corporation, MDX 20, Japan)
(Figure 3.3) was used to digitise the surface of the denuded bone. The MDX 20
scanner scans an object in the horizontal plane (x, y Plane), moving its head on “x”
direction while moving the stage in “y” direction. A needle connected to the head
containing an active piezo sensor moves vertically (z direction) perpendicular to the
x y plane until it touches the surface of the object and records the x, y and z
coordinates of the position of the needle. Then, the head moves towards x direction
at a set distance and records the position of the needle. Finally, the scanner collects a
point cloud with an x, y and z coordinate for each point.
The manufacturer‟s specifications of the scanner are given below (Table 3.1). The
active piezo sensor, to which the needle is connected, is highly sensitive and ensures
that the needle stops before it damages the surface of the object being scanned.
Chapter 3: Image processing and surface reconstruction
35
Figure 3.3: Scanning of the bone's outer surface of the diaphyseal region using the
MDX 20 contact scanner (The bone is positioned on the stage using glue tags)
Table 3.1 Specifications of the MDX 20 contact 3D scanner3
Property Value
Sensor Roland Active Piezo Sensor (R.A.P.S.) Probe length 60
mm (2-5/16 in.), tip bulb diameter 0.08 mm (0.00315 in.)
Scanning method Contacting, mesh-point height-sensing
Scanning pitch
X/Y-axis directions -0.05 to 5.00 mm (0.002 to 0.20 in.)
(Settable in steps of 0.05 mm (0.002 in.))
Z-axis direction - 0.025 mm (0.000984 in.)
Scanning speed 4-15 mm/sec. (1/8-9/16 in./sec.)
Exportable file formats DXF, VRML, STL, 3DMF, IGES, Greyscale, Point
Group and BMP
XY table size 220 (X) x 160 (Y) mm ( 8-5/8 x 6-1/4 in.)
Dr.PICZA software package (Intellecta Technology Pty Ltd, Adelaide, Australia)
installed on a personal computer was used for the operation of the scanner and for the
acquisition of the x, y and z coordinates from the scanner. In the present study, the
bone outer surface was digitised with a resolution of 0.3 mm × 0.3 mm in the
scanning plane (x, y plane) and a step size of 0.025 mm in the vertical direction (z
direction). The scanning of the surfaces of the bones was performed in two stages,
scanning of the diaphysis and the articular regions.
3 The information was drawn from the manufacturer‟s website.
Chapter 3: Image processing and surface reconstruction
36
To digitise the diaphysis of the bone, the soft tissue free bone was positioned
horizontally on the stage of the scanner (Figure 3.3). The bone was firmly fixed on
the stage using two sets of glue tags (Figure 3.3). Mechanical clamps were not used
so as to prevent damage to the bone‟s surface. Scanning of the shaft was carried out
in five steps, rotating the bone around its long axis by approximately 70° to generate
a total of five surfaces. Five surfaces were used in order to keep a good overlap
between two consecutive surfaces for their alignment using an ICP algorithm based
method.
The bone was then cut into three parts in order to scan the articular surfaces as these
regions cannot be reached while the bone is intact (Figure 3.4). The bone was
divided such that the distal and proximal parts were not longer than 50 mm, as the
maximum height of an object that can be scanned by the scanner is 55 mm. Then, the
proximal or distal part was positioned vertically on the stage to scan the articular
surfaces (Figure 3.5). The number of scans required for completely digitising the
bone‟s articular surfaces was determined by the complexity of the geometry of these
surfaces. In general, five to ten scans were carried out in each of the articular
surfaces. Before the scanning of articular surfaces started, the 3D model of the shaft
was reconstructed (Figure 3.6) and used as a guide to locate the areas to be scanned.
The scanned surfaces were exported as STL files for further reconstructing the 3D
model in Rapidform 2006.
Figure 3.4: Bone is cut in three parts in order to scan the articular surfaces which
cannot be reached by the scanner on the intact bone
Chapter 3: Image processing and surface reconstruction
37
Figure 3.5: Positioning of the proximal articular segment of the femur in order to
scan the articular surface
Figure 3.6: The reconstructed model before the scanning of articular surfaces (This
model was used as a guide to scan the articular regions)
3.8.3.3 Reconstruction of the 3D model from scanned surfaces
The 3D reference model was reconstructed from the scanned surfaces using the
reverse engineering software package Rapidform 2006. This was carried out in three
steps:
1. Removal of unusable data from the scanned surfaces
2. Aligning of the consecutive surfaces and
3. Merging of the aligned surfaces to generate the final 3D model.
Chapter 3: Image processing and surface reconstruction
38
As the first step, unusable triangles were removed from the scanned surfaces. The
scanner could acquire geometric data only on the horizontal plane (Figure 3.7 &
Figure 3.8). Thus, the geometric data in the horizontal plane is displayed with
equilateral triangles, while unusable data is usually represented by triangles which
have long faces. As a result, the surfaces contained triangles with variable length.
The triangles which are longer than 1 mm (roughly) were determined to be the
unusable data. Using functions built into Rapidform 2006, these unusable triangles
were removed permanently from the surfaces. In addition, the triangles which made
up the parts of the stage and the blue tags used to hold the sample to the stage were
also removed from the surfaces.
Figure 3.7: Scanned surface with unusable data
Figure 3.8: The surface after removing the unusable data
In the second step, two cleaned consecutive surfaces were aligned using the methods
described in Section 3.8.4.1(Figure 3.9). Once the alignment of the first two surfaces
Chapter 3: Image processing and surface reconstruction
39
was performed, the second surface was kept locked and the third surface was aligned
to the second.
Figure 3.9: Two adjacent surfaces are fine registered
Finally, all aligned surfaces were merged using the „Merge Surfaces‟ function built
into Rapidform 2006 (Figure 3.10). Once the merging of the surfaces was
completed, re-meshing of the triangular surface was carried out to obtain equilateral
triangles and an evenly distributed mesh.
Figure 3.10: The final 3D model reconstructed by merging the surfaces
3.8.4 Reference model for validation of the medullary canal
As the contact or laser scanner is unable to reach the medullary canal of ovine
femora, a microCT scanner was used to generate a reference standard. The scanning
was conducted only for the diaphyseal region due to the limitation of the sample
length that the scanner could accommodate.
The medullary canal of the soft tissue free bone diaphysis (Figure 3.4) was cleaned
to completely remove the bone marrow using a detergent solution and a brush. The
bone marrow was removed in order to have equal interfaces (bone-water) in the outer
and inner bone cortex. The bone was immersed in pure water and scanned with the
microCT (microCT 40, Scanco medical, Switzerland) scanner using the scanning
protocol shown in Table 3.2. Segmentation of the image data was conducted using
Chapter 3: Image processing and surface reconstruction
40
the Canny edge detection method described in Section 3.8.1, generating 3D models
of the outer and inner surfaces. Before the segmentation with the Canny filter, a 20 ×
20 median filter was applied to reduce the salt and pepper noise contained in the
images (Figure 3.11).
Table 3.2 Scanner parameters used for microCT scanning
Parameter Value
Resolution 0.03 mm × 0.03 mm
Slice spacing 0.03 mm
kVp 140
Figure 3.11: a - The original microCT image (a cross section from the diaphysis);
and b - the image after applying a 20 × 20 median filter
Both microCT based outer and inner 3D models were subjected to 90% decimation
to reduce the number of triangles that were contained within the 3D models to 900
000. As a result of 0.03 mm voxel size used for microCT scanning, the final 3D
models contained about 9 000 000 triangles; this made the models difficult to handle
in the software systems used for the study. The number of triangles contained in the
microCT based models after the decimation was still higher than the number of
triangles contained within the contact scanner generated reference models (~350
000); this indicated that the microCT based model was accurate enough to use as a
reference standard.
The outer 3D models generated from microCT images were validated with the
contact scanner generated 3D models, resulting in a nearly uniform error of 0.12 mm,
where the microCT model underestimated the reference model. As there was no
Chapter 3: Image processing and surface reconstruction
41
significant deviation of the outer 3D models from the reference standard, it was
assumed that the microCT generated inner models could be used to validate the inner
models generated from the CT and MRI data.
3.8.5 Basic 3D modelling techniques using Rapidform 2006
Throughout this study, the Rapidform 2006 (INUS Technology inc. Korea) reverse
engineering software package was used for the reconstruction and manipulation of
3D models. Registration of a model of interest to the reference standard was
conducted using a built in function that is based on the ICP algorithm. The
comparison of the model of interest to the reference model was conducted using a
point to point comparison method available in Rapidform 2006 software system.
3.8.5.1 Registration of 3D surfaces using Rapidform 2006
Aligning of surfaces was basically used on three occasions in this study: first, in the
aligning of the contact scanner generated surfaces of the denuded bone in order to
generate the reference model; second, in the aligning of 3D models prior to the
quantification of the geometric deviation between a model of interest and the
reference model; and, third, in the correction of the lateral shift artefact of 3D models
of long bones based on MRI. The registration process of the surfaces or models was
carried out in two steps: gross alignment and fine alignment of the surfaces.
The gross alignment was accomplished using the „Shell Trackball‟ function built into
Rapidform 2006. The reference model was locked in 3D space to prevent
accidentally moving the model. The model of interest was then connected with the
trackball and moved until the model was roughly in alignment with the reference
model (Figure 3.12). The trackball tool allows moving a 3D model in x, y and z
directions and rotating around those three axes using the mouse. The fine alignment
of the models roughly aligned with the trackball was carried out using the built in
function „Fine Registration‟ which is based on the ICP algorithm.
Chapter 3: Image processing and surface reconstruction
42
Figure 3.12: The initial aligning of the CT based 3D model to the reference model
using Trackball prior to the application of fine registration function
Figure 3.13: A CT based model (red) is aligned to the reference model (blue) in
Rapidform 2006 using the fine registration function
Chapter 3: Image processing and surface reconstruction
43
3.8.5.2 Comparison of the aligned 3D models
A method of calculating average displacement between two surfaces was used for the
comparison of a model of interest to the reference model [116]. This method
calculates the average of the deviations of the points in the model of interest to the
corresponding points in the reference model. The method is built into the Rapidform
2006 software package and was used on the surfaces that had been aligned using the
method described in the previous section. A graphical representation of the
distribution of the point to point deviations was also generated by the software
package (Figure 3.14). The 3D models were compared as complete models;
however, in some of the investigations (Chapters 3 and 5), the different anatomical
regions of the models were also compared in order to quantify the errors associated
with each anatomical region (Figure 3.15).
Figure 3.14: Comparison of the aligned CT model to the reference model in
Rapidform 2006
Chapter 3: Image processing and surface reconstruction
44
Figure 3.15: Five anatomical regions used for the comparison: 1 - femoral head, 2 -
proximal region, 3 - diaphysis, 4 - distal region, 5 - distal articular region
3.8.5.3 Dividing the 3D models of bones into different anatomical regions
Where the comparison of different anatomical regions was required, the 3D models
were divided into five anatomical regions (Figure 3.15) according to the guidelines
given in „AO principles of fracture management‟ [117]. The bone was divided using
two reference planes and two curves created in 3D space of Rapidform 2006 (Figure
3.16). The same reference planes and curves were used to divide all the models of
one sample.
Figure 3.16: Reference planes and curves used for the splitting of the model into five
anatomical regions
3.9 Results
Comparison of the complete outer bone models based on three segmentation methods
to the reference model generated average deviations of 0.24 mm, 0.18 mm and 0.20
mm for single threshold, multi-threshold and edge detector methods respectively.
Comparison of inner medullary canal models generated average deviations of 0.43
mm for the single threshold method, 0.17 mm for the multi-threshold method and
Chapter 3: Image processing and surface reconstruction
45
0.27 mm for the edge detector method. Detailed results are available in the paper
presented at the end of this chapter.
3.10 Summary, discussion and conclusion
Increased utilisation of virtual 3D models of long bones for various practices in the
clinic and in research has made the 3D reconstruction of long bones a research
interest. The process of reconstructing a 3D model involves several steps and each
step has factors that determine the geometric accuracy of these models. Among these
steps, the data acquisition, segmentation and surface reconstruction are equally
important; however, image segmentation is the mostly user intervened process and
has been discussed widely in the literature.
While a large number of methods are available for segmentation of bone data from
CT or MRI data, intensity thresholding is the most commonly used method due to its
ease of use. The unavailability of a method to select the appropriate threshold level
means that this method relies mainly on visual selection of a threshold level. In
addition, a single threshold level does not accurately select the ROI from all the
anatomical regions of a long bone, as different regions require different threshold
levels. The Canny edge detector is another segmentation method that can be easily
implemented as it is already incorporated in many of the image processing software
packages (e.g. Matlab). However, in the relevant literature, there is no reported use
of the Canny edge detector for segmentation of long bones.
In the present study, intensity thresholding and the Canny edge detector were
investigated for their accuracy and repeatability in segmenting the CT data of long
bone from ovine hind limbs. These two methods were selected as they do not involve
complex programming and can be administered by researchers with a limited
knowledge of programming and mathematics. A threshold selection method based on
the Canny edge detector was introduced for intensity thresholding to minimise the
user dependent errors of selecting a threshold level. In addition, a multilevel
thresholding approach was used instead of a single threshold level for segmenting the
complete long bone. Intensity thresholding with a visually selected single threshold
was also carried out for comparison purposes.
Chapter 3: Image processing and surface reconstruction
46
The results indicate that the multilevel intensity thresholding approach with the
threshold selecting method can produce 3D models with a relatively higher accuracy
(average deviation = 0.18 mm), in comparison to edge detection (average deviation =
0.20 mm) and the single threshold method (average deviation = 0.24 mm). However,
the overall accuracy obtained from all three methods was within acceptable range
(0.18 – 0.24 mm) for reconstruction of accurate 3D models, depending on the
accuracy required by the specific application. When different anatomical regions are
considered, the multi-threshold method was able to generate accurate models for
most of the regions, while single threshold generated the least accurate models for
most of the regions. Compared to the single threshold method, the other two
segmentation methods had a relatively higher repeatability.
The study utilised 3D surfaces derived by mechanically digitising the denuded bone
surfaces for an accurate validation of CT based models. This method is also used for
the validation of MRI based 3D models in next part of the research. A limitation of
this method was that no measures were considered for preventing the dehydration of
the bones during the digitisation. Practically this was difficult to achieve as the
bone‟s surfaces could not be covered during the scanning. There is no evidence to
suggest that dehydration has an effect on the cortical bone geometry; however, this
shrinks the cartilages which might be a reason for higher error occurred in this
region. The number of samples used was also limited to five due to longer processing
times even though the calculated sample size was 28 to detect the obtained
difference. With the sample size of five the detectable difference is 0.108 mm. The
accuracy required for designing orthopaedic implants are in the order of few
millimetres and thus, a difference of 0.108 mm would not affect the accuracy of the
reconstructed models.
This study demonstrated that by using relatively simple segmentation methods, 3D
models with sub-voxel accuracy can be generated. This allows the general research
community to use relatively simple methods without having to involve complex
programming and mathematics. The segmentation methods investigated in this part
of the research project will be used to segment the CT and MRI bone data throughout
the project.
Chapter 3: Image processing and surface reconstruction
47
The next chapter demonstrates an application of 3D models generated from CT data
where a validation of two intramedullary nail designs was conducted in a 3D
environment using 3D models of nails and the intramedullary canal of the tibia. The
study utilised 3D models based on CT scans of cadaver bones, however, if MRI is
used for scanning, this method can be used to assess the fit of intramedullary nails to
patient‟s bones.
Chapter 3: Image processing and surface reconstruction
48
3.11 Paper 1: Effect of CT image segmentation methods on the
accuracy of long bone 3D reconstructions (published)
Chapter 3: Image processing and surface reconstruction
49
226-233
Contents lists available at ScienceDirect
Medical Engineering & Physics
ELSEVJER journal homepage: www.elsevier.com/locate/ medengphy
Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions
Kanchana Rathnayaka a, Tony Sahamaa, Michael A. Schuetza.b, Beat Schmutza .•
• Institute of Health and Biomedical Innovation. Queensland University of Technology, Brisbane, Australia b Department of Orthopaedics. Princess Alexandra Hospital. Brisbane. Australia
ARTICLE INF O ABSTRACT
Article history: Received 6 May 2010 Received in revised form 20 August 2010 Accepted 4 October 201 0
Keywords: Computed tomography Image segmentation Canny edge detection Thresholding Bone models MicroCT Femur Mechanical digitiser
An accurate and accessible image segmentation method is in high demand for generating 3D bone models from CT scan data, as such models are required in many areas of medical research. Even though numerous sophisticated segmentation methods have been published over the years, most of them are not readily available to the ge neral research community. Therefore. this study aimed to quantify the accuracy of three popular image segmentation methods, two implementations of intensity thresholding and Canny edge detection, for generating 3D models of long bones. In order to reduce user dependent errors a ssociated with visually se lecting a threshold value, we present a new approach of selecting an appropriate threshold value based on the Canny filter. A mechanical contact scanner in conjunction w ith a microCT scanner was utilised to generate the reference models for validating the 3D bone models generated from CT data of five intact ovine hind limbs. When the overall accuracy of the bone m odel is considered, t he three investigated segmentation me thods generated comparable results w ith mean errors in the range of 0.18- 0.24mm. However, for t he bone diaphysis, Canny edge detection and Canny filter based thresholding generated 3D models w ith a significantly higher accuracy compared to those generated through visually selected thresholds. This study demonstrates that 3D models with sub-voxel accuracy can be generated utilising relatively simple segmentation methods that are available to the general research community.
1. Introduction
Accurate three-dimensional (3D) models of long bones are required in appl ications, such as implant design [ 1-5). finite element analysis (FEA) [6-11) and computer-aided surgical planning [ 12-15]. Computed tomography (CT) is curren tly the gold standard for the acquisition of data from which the 3D models of long bones are generated. Two main steps are involved in generating a 3D model from aCT data set: image segmentation; and 3D reconstruction of the segmented bone contours. In commercial image data processing and 3D reconstruction packages the latter is performed automatically with the user being limited to choose the level of surface smoothing to be applied. While surface smoothing can influence the accuracy of the reconstructed bone model (16,17]. for the purpose of this study smoothing was treated as a fixed entity (default setting of the commercial software package). Therefore, it is the former that was investigated, as an accurate and reproducible image segmentation method is a necessity for gener-
* Corresponding author at: Institute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin Grove. QLD 4059, Australia. Tel.: +61 7 3138 6238: fax: +61 7 3138 6030.
£-mail addresses: [email protected], [email protected] (B. Schmutz).
© 2010 IPEM. Published by Elsevier Ltd. All rights reserved.
ating 3D models that are accurate geometric representations of the actual bones.
Segmentation techniques are used to separate the region of interest (ROI) from the remainder of the image. The segmentation is critical as it is the major step demarcating between ROl and the background and thus. has a major effect on the geometric accuracy of the 3D model. Therefore, studies have been carried out to develop segmentation techniques that can produce 3D models with a high geometric accuracy. As a result, many image segmentation methods are available ranging from manual segmentation to semi and fully automated techniques ( 17-24].
Manual segmentation/tracing of the ROI by humans has long been practiced (17] and is so far the simplest method available for medical image segmentation. The major disadvantages of the manual segmentation are intra- and inter-personal variability which makes it a less repeatable method. This method is also more labour intensive and time consuming than the other segmentation techniques available.
Intensity thresholding is a popular segmentation method, which is implemented in commercial medical Image 3D reconstruction packages. In its basic form, this technique relies on visual selection of the threshold level by the user which has an effect on the accuracy and the repeatability of this method. In the absence of a standard method of selecting an appropriate threshold level, various
1350-4533/$ - see front matter © 2010 IPEM. Published by Elsevier Ltd. All rights reserved. doi:1 0.1 016/j.medengphy.201 0.10.002
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
57
Chapter 4 Application of 3D modelling
techniques for orthopaedic implant
design and validation
4.1 Introduction
Three dimensional models (3D) with accurate geometric representation of long bones
are increasingly being used for various aspects of clinical practice and research. They
provide a useful platform for the design and validation of implants, avoiding the
necessity to use cadaver bones. They also provide researchers with an opportunity to
design and validate implants for younger age groups who are more prone to injuries
and for whom there are only a few cadavers available. Designing implants that fit the
anatomy of young age and ethnic groups is also of particular important as age and
ethnicity are two of the factors that determine the geometric and mechanical
properties of long bones. Even though 3D models have been used for implant design,
their use for validation of the anatomical fit of the implants has seldom been
reported. Therefore, to address this need, this study investigates an in-silico
validation process of two intramedullary nail designs using triangular meshed 3D
surfaces generated from CT data of cadaver bones.
The reconstruction process of accurate 3D models from CT data is discussed in detail
in Chapter 3 of this thesis. This chapter now focuses on the application of these
models in the validation process of already designed implants. Section 4.2 discusses
the relevant literature and Section 4.4 briefly introduces the methodology used. A
detailed methods section is available in the journal paper presented at the end of this
chapter.
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
58
4.2 3D models for implant design and validation
The conventional use of cadaver bones for implant design and validation has a
number of challenges that researchers have to face. Most of the available cadaver
bones are basically obtained from older (>60 years old) donors. Thus, these cadaver
bones do not represent the young patient population who make up about half of the
patients who require implants. As most of the implant validation studies are carried
out in regions where fewer cadavers of Asian origin are available, the use of such
cadavers for implant design and validation is limited. Anecdotal clinical evidence
also suggests that the currently used trauma fixation plates do not optimally fit the
bones of patients from the Asia-Pacific region, as they have been designed mainly
for the Caucasian population.[45] Therefore, the implant design and validation
process needs to be extended to both the young and Asian-pacific population.
Accurate 3D models of small or long bones provide a better platform for design and
validation of implants for different age and ethnic groups. Using MR imaging, 3D
models of long bones can be reconstructed from almost all age groups, as MRI is a
potential alternative to CT for generating 3D models of bones (See Chapter 3). The
limitation of not having enough Asian-Pacific cadavers can also be overcome by
using 3D models generated from such populations using MRI. This use of MRI also
enables researchers to repeatedly use the same specimen for validation studies in a
simulated environment without having to damage the already available, valuable
cadaver specimens.
There are only a few studies that have been conducted to quantify the anatomical fit
of an implant. The first reported is the study conducted by Goyal et al. [46] using 101
tibiae and medial and lateral proximal periarticular plates. In this study, the
quantification of the fit was conducted by digitising the position of the plate and the
bone, using a mechanical digitising arm (Microscribe). Haraguchi et al. [118] used
CT scans of 50 patients and a ORTHODOC workstation to compare the fit and fill
between anatomic stem and straight tapered stem. This was performed using 3D
surfaces extracted using the software system and placing the implants virtually in the
3D surface models. A study quantifying the plate fit using 3D models has been
reported by Schmutz et al. [45]. Twenty one 3D models of tibiae from a database at
AO Development Institute, Davos, Switzerland and a 3D model of the distal
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
59
periarticular tibia plate were used. Using 3D modelling techniques, the distance
between plate and the bone and the angle between plate and the bone were measured
to assess the fit of the plate to the bone.
Even though these few studies on validation of plates have been reported, no studies
to validate intramedullary nails using 3D modelling techniques have been conducted
to date. The insertion force and insertion distance of the nail is often used as an
indicator for anatomical fitting of a nail; however, fit of an intra-medullary nail in the
final position cannot be quantified using these methods. The validation using
cadavers is also limited by the small number of available cadavers, and those that are
available might not be representative of the target population‟s age and ethnicity.
4.3 Aims of the study
This study aims to develop a non-invasive method to quantitatively assess the
anatomical fitting between an intramedullary nail in the final position and the bone,
using 3D models of long bones.
4.4 Methods
The study used two designs of the expert tibial nail (ETN): ETN and ETN with bend
(Synthes, Bettlach, Switzerland) and 20 CT based 3D cortex models of Japanese
cadaver tibiae. 3D models of the ETN and ETN with bend were virtually positioned
in the 3D model of tibiae using the Rapidform 2006 software system to meet the set
criteria to obtain the optimal fit. The maximum distance and the area of the part of
the nail protrusion were measured using 3D modelling techniques. A detailed
methodology used for the study is available in the paper presented at the end of this
chapter.
4.5 Results
The total area of the nail protruding from the medullary canal was 540 mm2
for the
ETN with bend, and 1044 mm2 for the ETN. The maximum distance of the nail
protruding from the medullary canal was 1.2 mm for the ETN with bend, and 2.7 mm
for the ETN.
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
60
4.6 Summary, discussion and conclusion
The limited accessibility to cadaver bones of younger age groups and different
ethnicity makes the design and validation of the implants specific for them a difficult
process. The use of cadaver bones allows assessment of nail insertion force;
however, available cadavers are limited in number and do not represent the target
population. Validation using plain x-ray is also limited to 2D. The use of 3D models
provides the opportunity to access bone geometric data from younger age groups of
different ethnicity, using non-invasive MRI scanning. This also allows for repeated
use of the same sample for implant validation, and provides an accurate method to
quantify the anatomical fit of implants.
This study quantified the anatomical fit of two nail designs (ETN and ETN with
bend) to the 3D models of tibiae reconstructed from CT data of Japanese cadavers.
Based on the results, the total area and the maximum distance of the nail protruding
from the medullary canal were smaller for the ETN with bend compared to the ETN.
Both protruding area and the distance showed statistically significant differences
between ETN with bend and ETN. Therefore, compared to the original ETN, the
modified nail design (ETN with bend) had a better fit. This will provide a better
alignment of the fractured bone segments, resulting in a better fracture healing
outcome.
The method presented in the study using 3D models of the nails and tibiae was non-
invasive. This is also radiation hazard free when MRI is used to scan the bones.
Thus, this method has the potential of validating the nails or plate designs for healthy
human volunteers who represent the target patient populations of young age and
different ethnicity.
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
61
4.7 Paper 2: Quantitative fit assessment of tibial nail designs using
3D computer modelling (published)
Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation
62
Injury, lnt. j . Care Injured 41 (2010) 216-219
Contents lists available at ScienceDirect
Injury
ELSEVIER journal h omepage: www.elsevier.com/ lo catelin ju ry
Quantitative fit assessment of tibial nail designs using 30 computer modelling
B. Schmutz a.*. K. Rathnayaka a. M.E. Wullschleger a.b,]. Meek c. M.A. Schuetz a,b
• Institute of Health and Biomedical Innovation, Queensland University of Technology, 60 Musk Avenue, Kelvin Grove, Brisbane, QLD 4059, Australia b Trauma Services, Princess Alexandra Hospital, Brisbane, Australia c Synthes GmbH, Oberdorf, Switzerland
ART I CLE I NFO
Article history: Accepted 5 October 2009
Keywords: 3D model Tibia Intramedullary nail Nail fit Fracture fixation
Introduction
ABST R ACT
Intramedullary nailing is the standard fixation method for displaced diaphyseal fractures of the tibia in adults. The bends in modern tibial nails allow for an easier insertion, enhance the 'bone-nail construct' stability, and reduce axial malalignments of the main fragments. Anecdotal clinical evidence indicates that current nail designs do not fit optimally for patients of Asian origin. The aim of this study was to develop a method to quantitatively assess the anatomical fitting of two different nail designs for Asian tibiae by utilising 3D computer modelling.
We used 3D models of two diffe rent tibial nail designs (ETN (Expert Tibia Nail) and ErN-ProximalBend, Synthes), and 20 er-based 3D cortex models of Japanese cadaver tibiae. With the aid of computer graphical methods, the 3D nail models were positioned inside the medullary cavity of the intact 3D tibia models. The anatomical fitting between nail and bone was assessed by the extent of the nail protrusion from the medullary cavity into the cortical bone, in a real bone this might lead to axial malalignments of the main fragments. The fi tting was quantified in terms of the total surface area, and the maximum distance by which the nail was protruding into the cortex of the virtual bone model.
In all 20 bone models, the total area of the nail protruding from the medullary cavity was smaller for the ETN-Proximai-Bend (average 540 mm2
) compared to the ETN (average 1044 mm2) . Also. the
maximum distance of the nail protruding from the medullary cavity was smaller for the ErN-ProximalBend (average 1.2 mm) compared to the ETN (average 2.7 mm). The di fferences were statistically significant (p < 0.05) for both the total surface area and the maximum distance measurements.
By utilising computer graphical methods it was possible to conduct a quantitative fit assessment of different nail designs. The ETN-Proximai-Bend shows a statistical significantly better intramedullary fi t with less cortical protrusion than the original ETN. In addition to the application in implant design, the developed method could potentially be suitable for pre-operative planning enabling the surgeon to choose the most appropriate nail design for a particular patient.
© 2009 Elsevier Ltd. All rights reserved.
Intramedullary nailing is the standard fixation method for displaced dia physeal fractures of the tibia in adults.6
·10 The bends
in mode rn tibial nails a llow for a n easie r insertion, e nhance the 'bone-nail construct' stability, and reduce axial malalignments of the main fragme nts.3 - 5 ·
9 Typically, the nails a re designed with the view to fit the 50th percentile of a Caucasian/Weste rn population. Clinical tria ls of the nail designs a re then conducted in hospitals were the majority of the patients are of Caucasian orig in. Such was the case for the Expert Tibial Nail (ETN), one of the nail designs used in th is study. The results of a clinical study13 from one of the
hospitals involved in the multi-centre clinical tria l o f this nail confirmed the improvements5 of the nail design as appropriate fo r their patient collective. exclusively of Caucasian origin (Striegel A, personal communication, July 19, 2009). Despite this. anecdotal clinical feedback is emerging, indicating that the curre nt nail des ign does not fit optimally in the proximal dorsal region for the tibial geometry of Asian patients.
One important aspect of designing a new or improved implant shape is validation, which is often conducted in the form of cadaver trials. In the case of precontoured plates, the anatomical fitting can be visually assessed or quantified by fitting plates to bones.2
However. for nails, one aspect of validation pertains to the ease w ith w hich the nail can be inserted into the bone, and the other to the anatomical fitting between the nail and bone geometry in the nail's final position. Neither of these can be achieved in form of a visual assessment. The force required for inserting the nail into the bone is
• Corresponding author. Tel.: +61 7 3138 6238; fax: +61 7 3138 6030. £-mail address: [email protected] (B. Schmutz).
0020- 1383/$ - see front matter ~ 2009 Elsevier Ltd. All rights reserved. doi:1 0.1 016fj.injury.2009.1 0.012
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
67
Chapter 5 Magnetic resonance imaging for 3D
reconstruction of long bones
5.1 Introduction
The widely accepted standard for generating 3D models of bones for implant design
and related research is CT as this offers a better contrast at the bone–soft tissue
interface that greatly facilitates the segmentation process. However, due to the
involvement of a high dose of ionising radiation, CT cannot be used for scanning of
healthy volunteers for research purposes. Therefore, an alternative radiation free
imaging method such as MRI is necessary for the generation of 3D models of long
bones from healthy human volunteers. While bones do not generate a useful signal in
clinical MRI due to extremely short transverse relaxation times, the bone geometry
can be delineated using the signal generated from the surrounding soft tissue. Even
though this has been demonstrated in some studies, the accuracy of these models has
to be quantified using in vitro and in vivo studies before such models can be used for
implant design.
This chapter discusses the investigation carried out to quantify the accuracy of 3D
models reconstructed using a currently available clinical MRI scanner. Section 5.2
discusses the relevant literature, and Section 5.4 briefly introduces the methods used
for the study. More details of the study are available in the published journal article
that is presented at the end of this chapter. Basic principles of the MRI scanner and
their relevance to bone imaging have been discussed in Chapter 2. The segmentation
techniques and 3D modelling methods used in this study have been previously
investigated and validated as a part of the PhD project and the details are available in
Chapter 3.
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
68
5.2 Imaging of skeletal system with MRI
MRI is designed to scan soft tissues utilising 1H nuclei as the source of signal. The
signal intensity generated from a particular tissue type in MRI is determined by the
longitudinal relaxation time (T1), transverse relaxation time (T2), and proton density
of the tissue type. Different soft tissues of human or other mammalian bodies have
different T1 and T2 values. Hence, by using different TR and TE values, a better
contrast between two different soft tissue types can be obtained. Due to the superior
contrast obtained, MRI has become the method of choice for quantitative studies of
cartilage, muscles and other soft tissues [1, 13-17].
In contrast to the soft tissues, cortical bone (including ligaments and menisci) has
extremely short transverse relaxation times (T2) and does not produce an adequate
signal in clinically used pulse sequences [10-12]. Hence, the visualisation of the bone
structure is not possible with clinically used pulse sequences. This might be achieved
with special pulse sequences that have ultra-short TE values, in which TE is reduced
to 0.07-0.20 ms from usual values of 4-10 ms [10, 12]. Human or other mammalian
long bones are surrounded by a good bulk of muscles and other soft tissues. These
soft tissues are capable of producing MR signals with high intensity when clinically
used pulse sequences are employed and, hence, produce a high contrast between the
cortical bone and surrounding soft tissues. Using this high contrast, it is possible to
identify the cortical bone geometry in MRI images acquired with the clinically used
protocols (Figure 5.1).
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
69
Figure 5.1: Cross sections of CT (left) and MRI (right) from the same anatomical
location of a sample (In the CT image, the cortical bone appears in high intensity and
can be clearly identified from the surrounding soft tissue. In the MRI image, cortical
bone appears in black as it does not generate a significantly high signal; however, the
outer cortex can be identified due to the signal generated by surrounding soft tissues)
MRI has long been used for bone imaging, mainly for diagnosing metastatic disease,
for computer assisted surgery (CAS) and for bone motion kinematic studies. The
skeletal system is one of the main targets for cancer metastases and MRI has been a
superior imaging method to detect these metastases over the other imaging
techniques (CT and plain x-ray) [18]. MRI has also been used for quantification of
the trabecular bone structure in several studies [119-121]. The next main use of MRI
related to bone imaging is for CAS of the spine [19, 20, 24]. The usual practice for
CAS of the spine is to generate 3D models of the spine using CT to help the accurate
placement of pedicle screws. Due to the radiation exposure of CT, Hoad et al. [24]
have developed an MRI imaging sequence that can be used for generating 3D models
of the spine. A double echo sequence was used and a 3D model of the vertebrae was
generated by manually segmenting the image data. The model was compared with a
similarly generated model using CT. The results show that the accuracy of the MRI
based model is 90% compared to the 100% accuracy of the CT based model.
Bone kinematic studies have also been a major research area that has used MRI in
place of gold standard CT due to the high radiation exposure [21, 25, 122-124]. CT
has been the typical image acquisition method for quantification of position of bones
during various movements due to the very short image acquisition time and the high
soft tissue bone contrast. Wolf et al. [25] imaged the feet of five volunteers with a 3T
MRI system (resolution 0.39 mm × 0.39 mm × 0.7 mm) using a 3D T1 weighted
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
70
gradient echo sequence. The metatarsal bones were segmented using an intensity
thresholding segmentation method. Comparison of the MRI based 3D models with a
reference standard was not conducted; however, Wolf et al. recommends use of MRI
for foot bone motion quantification. Fassbind et al. [21] used a 1.5T MRI scanner to
quantify foot bone motion and obtained kinematic characteristics similar to those
cited in other published studies that used CT. Pillai et al. [123] studied the wrist bone
motion using 3D models of radius, scaphoid and lunate generated from 1.5T MRI
scanner for which a low resolution (0.31 mm × 0.31 mm × 2 mm ) 3D FLASH
sequence was used. Manual segmentation was performed excluding the cartilage
from bone. The kinematic analysis showed results similar to those published.
In addition to those mentioned above, a number of studies have been conducted to
investigate the kinematics of the tibio-femoral joint using MRI. DeFrate et al. [125]
and Chen et al. [126] studied the knee kinematics generating 3D models of the distal
part of the femur and the proximal part of the tibia using MRI, while Hao et al. [127]
used MRI to generate a finite element model of a knee joint. All three studies
obtained accurate results for kinematic studies. Even though all of the studies
described above have reconstructed the 3D models of various parts of long bones and
small bones using MRI, a validation with a proper reference standard was not
employed to quantify the accuracy of those models.
Musculoskeletal models that represent bone including cartilage, ligaments and
muscles have been successfully generated using a combination of MRI and CT. Lee
et al. [1] conducted a study using five porcine femora to generate a combined MRI
and CT model in which MRI was used to reproduce soft tissues, while CT was used
for bones. The in plane resolution of CT and MRI data was 0.4 mm × 0.4 mm and
0.3 mm × 0.3 mm respectively. CT had slice reconstruction interval of 0.625 mm,
while the slice thickness of MRI was 1.2 mm. The 3D models were reconstructed by
manually segmenting the image data. The 3D models derived from MRI were
registered to the CT models with a surface matching accuracy of 0.7 ± 0.1 mm.
Moro-oka et al. [124] conducted a study to compare three-dimensional kinematic
measurements from single plane radiographic projections. Three knee joints of
human volunteers were scanned using CT and MRI scanners with the resolution of
0.35 mm ×0.35 mm × 1.00 mm and 0.39 mm ×0.39 mm ×1.00 mm respectively. 3D
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
71
models of the knee joint were reconstructed using a commercial software package;
however, the method used for the segmentation has not been mentioned. The surface
matching of the MRI models to CT models presented differences of -0.11 ± 0.81
mm, -0.23 ± 0.48 mm and -0.12 ± 0.60 mm for femora of three subjects, and -0.14 ±
0.67 mm, -0.13 ± 0.48 mm and -0.15 ± 0.77 mm for tibiae.
The studies mentioned above have shown that a sub voxel level accuracy can be
obtained for 3D models of bones using MRI. The main drawback of the studies is
that MRI models have not been validated using a proper reference standard. Most of
the studies have not used MRI to generate the complete 3D models of long bones
which are necessary for the design of trauma fixation plates and nails. The studies
have also not focused on quantifying the accuracy of the medullary canal of a long
bone that is important for designing intramedullary nails. Therefore, studies that
focus on quantifying the geometric accuracy of complete 3D models of long bones
and the medullary canal are required to inform the improved design of orthopaedic
implants using 3D models based on MRI.
5.3 Aims of the study
The aim of the study was comprehensive quantification of the accuracy of 3D models
based on MRI compared to the 3D models based on CT, and their formal validation
using a reference standard based on the contact surface scanner.
5.4 Methods
MRI and CT scans of five intact femora were obtained by scanning five intact ovine
hind limbs. Ovine femora were used, as human volunteers cannot be CT imaged due
to radiation exposure, and contact mechanical scanning cannot be used for validation.
The sample size calculation showed that the required sample size to detect a
difference of 0.08 mm with standard deviation of 0.02 is four. A 1.5T MRI scanner
(Siemens Magnetom Avanto) and a 64 slice CT scanner (Phillips Brilliance 64) were
employed for scanning of the limbs. The MRI scanner was used with a 3D flash
sequence, TR = 11 ms, TE = 4.94 ms, FA = 15º and 0.45 mm ×0.45 mm resolution
with 1 mm slice thickness. These parameters were chosen as they produced the best
results for different parameter combinations used in the pilot study. The CT scanner
was used with a 0.4 mm ×0.4 mm in plane resolution and 0.5 mm slice spacing, kVp
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
72
= 140 and mAs = 231. The segmentation of MRI and CT data was performed using
the multi-threshold segmentation method combined with the threshold selection
method developed as a part of this research project (See Chapter 3).
The triangular meshed contact scanner and microCT generated surfaces of the soft
tissue free bones were used as reference standards for outer and inner surfaces
respectively. The method of generating the reference standard has been described in
Section 3.8.3. Using the point to point comparison method described in Section 4.3,
comparisons were carried out between the MRI based models and the reference
models, the CT based models and the reference models, and the MRI based models
and the CT based models. A detailed description of the methods is presented in the
journal article presented at the end of this chapter.
5.5 Results
Comparison of the MRI based and CT based 3D models to the reference models
showed average errors of 0.23 mm and 0.15 mm respectively. Statistically, there was
no significant difference between the 3D models based on two methods (p = 0.067).
A detailed results section is available in the publication presented at the end of this
chapter.
5.6 Summary, discussion and conclusion
MRI has shown to be an ionising radiation free potential alternative for CT. A
number of kinematic studies, finite element studies and studies of the diagnosis of
metastatic disease have successfully used MRI as an alternative to CT for scanning
the skeletal system, even though these studies have not validated MRI based 3D
models with a proper reference standard. Few studies [1, 124] surface matched MRI
based 3D models to CT based models and reported sub voxel level accuracies;
however, a reference standard such as a laser or contact scanner has not been used
for the validation.
The present study aimed at quantifying the accuracy of the surface geometry of MRI
based 3D models and the CT based 3D models, using state of the art dense triangular
meshed surface scans of the outer and inner surfaces of femora as the reference
standard. The study acquired MRI and CT data from five sheep femora with intact
soft tissues and intact joints. 3D models were generated using a multilevel
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
73
thresholding method in combination with a method to select the threshold level for
the particular region.
The resulting accuracy of the MRI based 3D models (average deviation = 0.23 mm)
was comparable to that of CT based models (average deviation = 0.15 mm). There
was no statistically significant difference between the two methods. This indicates
that the 3D models based on MRI can be used as an alternative to CT for 3D
reconstruction of long bones. The statistical analysis also shows that to detect this
difference (0.08 mm) a sample size of four is sufficient. The diaphyseal region of the
femora presented an accuracy of 0.15 mm, while the proximal and distal regions
which have complex geometric shapes gave a relatively lower accuracy. The poor
contrast obtained from the 1.5T MRI scanner for these articular regions forced to
manually segment most of these regions, potentially introducing errors to the final
3D surfaces.
The long scanning time of the MRI compared to the CT scanning time poses a
number of additional limitations when MRI is used for scanning of human
volunteers. The motion artefact from random movements is one of the important
limitations and this is addressed in Chapter 7 of this thesis. The longer segmentation
time of the MRI images compared to the segmentation time of the CT images also
limits the use of MRI for imaging of long bones. This study used ovine femora as the
study sample. In order to apply these methods to the much larger human long bones,
additional studies using human long bones are desirable before applying these
methods on humans. One limitation of the study is that only two of five bones were
used for the reconstruction of the inner medullary canal due to the inadequate
contrast in other bones. The fat/water only imaging would be advantageous here and
would be suggested as possibility in future.
The study showed that MRI can generate 3D models of long bones with accuracy
comparable to that of CT models. Using a higher field strength scanner, typically 3T,
the current drawbacks of poor contrast in certain anatomical regions and the longer
scanning times can be potentially overcome. The next chapter will investigate the use
of 3T MRI to overcome these drawbacks using human volunteers as the study
samples.
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
74
5.7 Paper 3: Quantification of the accuracy of MRI generated 3D
models of long bones compared to CT generated 3D models (in
press)
Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones
75
Medical Engineering & Physics xxx (2011) xxx-xxx
Contents lists available at Scie nceDirect
Medical Engineering & Physics
ELSEVIER journal homepage: www . elsev ie r .co m /locat e/medengp h y
Quantification of the accuracy of MRI generated 30 models of long bones compared to CT generated 30 models
Kanchana Rathnayaka a. Konstantin I. Momot b, Hansrudi Noserc, AndrewVoJp d, Michael A. Schuetza,d, Tony Sahama b, Beat Schmutza,. • lnstirute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin Grove, QLD 4059, Australia b Queensland University of Technology, 2 George Street, Brisbane, QLD 4000, Australia ' AO Research Institute Davos. Clavadelerstrasse 8. 7270 Davos. Switzerland d Princess Alexandra Hospital. 1 99 1pswich Road. Woolloongabba. Brisbane. QLD 4102, Australia
ARTICLE INFO ABSTRACT
Article history: Received 18 February 2011 Received in revised form 25 July 2011 Accepted 27 July 2011
Keywords: MRI CT 3D models Femur
Orthopaedic fracture fixation implants are increasingly being designed using accurate 3D models of long bones based on computer tomography (Cf). Unlike cr. magnetic resonance imaging (MRI) does not involve ionising radiation and is therefore a desirable alternative to cr. This study aims to quantify the accuracy of MRI-based 3D models compared to er-based 3D models of long bones. The femora of five intact cadaver ovine limbs were scanned using a 1.5 T MRI and a er scanner. Image segmentation of er and MRI data was performed using a multi-threshold segmentation method. Reference models were generated by digitising the bone surfaces free of soft tissue with a mechanical contact scanner. The MRIand er -derived models were validated against the reference models. The results demonstrated that the er-based models contained an average error of0.15 mm w hile the MRI-based models contained an average error of 0.23 mm. Statistical validation shows that there are no significant differences between 3D models based on er and MRI data. These results indicate that the geometric accuracy of MRI based 3D models was comparable to that ofCf-based models and therefore MRI is a potential alternative to er for generation of 3D models with high geometric accuracy.
1. Introduction
Three-dimensional (3D) models of long bones with a high geometric accuracy are widely utilised by medical engineering research and in clinical practice; the design of orthopaedic fracture fixation implants [1 ,2]. computer aided surgery simulations [3,4] and fracture healing models [5,6] are just a few examples. Computed tomography (CT) has become the gold standard for scanning of bones to produce 3D models with high geometric accuracy. Due to high radiation exposure, CT cannot be used to scan healthy human volunteers. Therefore, an alternative method for the scanning of long bones of the healthy human population needs to be investigated.
Among various uses of 3D models, orthopaedic implant design particularly requires 3D models with high geometric accuracy to produce implants with a better fit to the patients' anatomy [ 1 ,3 ]. Furthermore, the anatomically pre-shaped implants are often designed based on the Caucasian population and thus the size and
*Corresponding author. Tel.: +61 7 3138 6238; Fax: +61 7 3138 6030. f -mail address: [email protected] (B. Schmutz).
© 2011 IPEM. Published by Elsevier Ltd. All rights reserved.
shape do not accurately match the Asian population. Therefore, those pre-shaped implants still need some optimisation for a better anatomical fit to people of different ethnic origins and age groups [2]. Ethnicity and age are two important factors which determine the shape and size of bones [7,8]. Thus, a database with accurate bone data from different ethnic and age groups is essential for this purpose.
Some institutions have already started developing such databases using CT imaging of cadaver bones [9] but the majority of these bones are from older donors (>60 years) therefore do not represent the young patient population. Furthermore. cadaver bones can seldom be chosen according to the researchers' need (e.g. gender or specific subject height) due to the limited availability. Therefore, there is a need to collect bone data from healthy human volunteers who represent that part of the patient population for whom no CT data exists or can be acquired. This would facilitate researchers' access to specific population groups for the purpose of obtaining high-quality anatomical image data.
CT scanning of healthy human volunteers is not ethically justifiable due to the high radiation exposure [10,11 ]. Studies investigating CT imaging protocols that use low radiation doses, while keeping the original image quality, have become an important part
1350-4533/$- see front matter © 2011 IPEM. Published by Elsevier Ltd. All rights reserved. doi: 10.1 016/j.medengphy.2011.07.027
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
83
Chapter 6 Higher field strength MRI scanning of
long bones for generation of 3D models
6.1 Introduction
As discussed in Chapter 5, 1.5T MRI offered acceptable accuracy for reconstruction
of 3D models from long bones. However, the images of long bones acquired at 1.5T
MRI need to be further improved to overcome limitations such as poor contrast in
articular regions and long scanning times. The high field strength scanners are
promising to offer higher signal to noise ratio (SNR) levels [22] which can
potentially be used to overcome these limitations of 1.5T scanners. The higher SNR
obtained at higher field strengths could, in principle, be used either for higher
resolutions or for higher contrast levels. In this study, the signal gain was
investigated in the form of contrast to noise ratio (CNR).
As the intrinsic SNR of an MRI system is approximately proportional to the main
magnetic field (B0), if all the parameters, subjects and radio frequency (RF) coils are
equivalent, scanners with 3T magnets should theoretically yield approximately
double the SNR at 1.5T [128]. However, the increased main magnetic field affects
the tissue parameters such as the longitudinal relaxation time (T1) and transverse
relaxation time (T2). Therefore, before the use of 3T scanners for musculoskeletal
imaging, it is necessary to investigate the effect of the higher field strength on the
tissue parameters (e.g. T1 or T2) and imaging artefacts, and to optimise the imaging
protocols accordingly.
Section 6.2 of this chapter discusses the theoretical increase of SNR at 3T. Relevant
literature on the use of the 3T MRI system for scanning of the musculoskeletal
system is discussed in Section 6.3 . Section 6.5.1 discusses the basic principles of the
methods used for quantification of image quality of an MRI system. While a detailed
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
84
description of the methodology used in this study is available in the publication, a
brief introduction is included in Sections 6.5.2 and 6.5.3.
6.2 Theoretical consideration of increased SNR at 3T
The theoretical signal gain in an MRI system is proportional to the square of the
main magnetic field; thus, the signal gain at 3T should be 4 times that at 1.5T, as
given in the equation below:
2
0BS 6.1
Where S = signal, B0 = main magnetic field, = gyromagnetic ratio.
The noise level (N) of a MRI system is proportional to the Larmor frequency and,
hence, to the main magnetic field. Thus, the noise level at 1.5T becomes two fold at
3T (6.2). Therefore, the actual SNR gain at 3T is two times that of 1.5T.
00 BNvN 6.2
22
4
N
SSNR 6.3
Where N = Noise level, 0v = Larmor frequency, B0 = main magnetic field, SNR =
signal to noise ratio and S = signal [23].
6.3 3T MRI for musculoskeletal system imaging
Scanners with higher field strengths, typically 3T, became available for clinical
scanning in the 1990s. Since then, a number of quantitative and qualitative
comparisons between 1.5T and 3T have been carried out, mainly to compare various
soft tissue compartments [43, 129-134]. In comparison, fewer articles have been
published comparing 3T imaging of the musculoskeletal system. Of these, some are
related to quantifying the cartilage morphology [135, 136] and the spin relaxation
times or anatomical structure demonstrations [128, 137-139]. A relatively large
number of review articles have been published by MRI experts regarding various
aspects of 3T or high field MRI [23, 140-146].
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
85
Among the articles published regarding musculoskeletal imaging with 3T, Stehling et
al. [137] assessed a multi-contrast high resolution imaging protocol for imaging of
the wrist of 10 volunteers using 3T and 1.5T scanners. The imaging protocol at 3T
had half the in plane resolution and half the slice thickness of those at 1.5T imaging
protocol (0.5 mm × 0.5 mm and 3.0 mm respectively). The idea was to use the SNR
gain at 3T for better image resolution. The qualitative assessment showed that the
structure and overall image quality was significantly higher in 3T (p<0.01).
Gold et al. [147] conducted a study to calculate the spin relaxation times of 3T and
1.5T of the musculoskeletal system using five human volunteers. The T1 was
increased in the 3T MRI compared to the 1.5T MRI, while T2 was slightly decreased.
In the same study, SNR and CNR of muscles were compared using one volunteer
whose knee was scanned in both 1.5T and 3T. A sagittal, proton density weighted
fast spin-echo sequence was used with a TR and TE of 4000 ms and 14 ms
respectively. A coronal T1-weighted spin-echo sequence was used with TR and TE of
800 ms and 14 ms respectively. The parameters were identical at both field strengths.
SNR and CNR were calculated for cartilage, muscle, fat and synovial fluid. SNR of
muscle and fat were more than twice at 3T compared with 1.5T. CNR between
synovial fluid and cartilage at both long TR (4000 ms) and short TR (800 ms) had
increased at 3T compared with 1.5T. All the SNR measurements were higher at 3T
and the values were statistically significant. A pictorial comparison conducted by
Gold et al. has also shown that 3T MRI gives significantly better results for
musculoskeletal tissues [128].
In addition to the studies mentioned above, Lambert et al. showed that 3T MRI can
be used to detect rotator cuff tears; this was a follow up qualitative study and no
comparison was performed with 1.5T MRI images [138]. No studies that compare
1.5T and 3T bone–muscle interface in relation to generation of 3D models have been
reported.
6.3.1 Spin relaxation times and flip angle
Longitudinal relaxation time (T1) is highly dependent on the strength of the main
magnetic field and increases with the increase of the main magnetic field [23, 147].
The reported increase of T1 relaxation times at 3T are: 20 - 22% for fat [131, 147],
20% for skeletal muscle, 62% for gray matter, 42% for white matter in the brain and
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
86
41% for liver [147, 148]. Previous measurements of relaxation times have shown 70-
90% increase of T1 at 4T than at 1.5T [139]. This increased longitudinal relaxation
time at 3T requires the TR to be increased to obtain the same imaging contrast as at
1.5T, especially for T2 and proton density weighted images; this, in turn, increases
the imaging time at 3T. This also decreases the SAR values and allows more slices
per scan.
Transverse relaxation time (T2) is relatively less dependent on the main magnetic
field. T2 slightly decreases with the increase of the main magnetic field; thus, similar
contrast levels may be obtained at slightly shorter TEs at 3T compared to 1.5T [141].
Gold et al. have reported that the decrease of T2 is about 10% for muscle and 19%
for fat (marrow and subcutaneous) [147].
For a given TR at a certain T1, the optimum flip angle should be used to obtain the
maximum signal. This is called the „Ernst angle‟, and is given by the following
equation:
1cosTTR
e 6.4
Where is the flip angle, TR = repetition time, and T1 = longitudinal relaxation time
of the tissue.
6.3.2 Fat suppression
Fat suppression is a technique used in MRI to suppress the signal from normal
adipose tissue to reduce the artefacts or to characterise the tissue [149]. This is
usually achieved by sending pulse to suppress the resonance frequency of the fat
tissue. Since the difference of the resonance frequencies of water and fat is 220Hz at
1.5T, a pulse used to suppress the signal from fat will partially or completely
suppress the signal from water molecules. As this difference becomes 440Hz at 3T,
fat suppression can be achieved easily without suppressing the signal from water
molecules [141].
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
87
6.3.3 Magnetic susceptibility at 3T MRI
Magnetic susceptibility occurs due to the change of the static magnetic field in the
presence of materials such as bones, metals, certain blood products and air. Due to
this variability of the static magnetic field strength, spins precess at different
frequencies. High magnetic fields are highly sensitive and prone to magnetic
susceptibility more than 1.5T fields [150]. Magnetic susceptibility causes a signal
loss, leading to geometric distortions of the region. This will affect the localisation of
certain anatomical structures, not only in clinical practice but also in the 3D models
generated using such a data set.
The degree of magnetic susceptibility can be minimised with various practices.
Increasing bandwidth decreases susceptibility artefact; however, this is at the
expense of SNR (Equations 2.4 & 2.5 P15). By doubling the bandwidth, SNR is
decreased by about 40% [23]. This can also be achieved by decreasing echo time
(TE). Decreasing the voxel volume is another method which can be used to reduce
the susceptibility artefact. This is usually achieved by using higher special resolution
with thinner slices.
6.3.4 Chemical shift at 3T
Chemical shift is the difference in precessional frequency conferred by magnetic
shielding effect of the electron clouds that surround protons within tissues relative to
that of a standard reference compound (in the case of protons, tetramethylsilane, or
TMS). The chemical shift is present in any tissue, but the electron cloud in the fat has
a major effect as fat is an abundant homogenous tissue in the body [141]. Due to this
frequency difference between lipid/fat and water, protons from the same tissue
location map to different positions in the reconstructed image.
The difference of the precessional frequencies between water and fat is
approximately 220Hz at 1.5T. As this frequency is directly proportional to the main
magnetic field, it becomes approximately 440 Hz at 3T [128, 141, 151]. This doubles
the chemical shift between water and fat, doubling the number of misregistered
pixels and worsening the artefact. The chemical shift can be minimised by increasing
the receiver bandwidth of the scanner; for example, doubling the BW will decrease
the chemical shift to the same as at 1.5T. However, doubling the receiver BW
decreases the SNR by a factor of square root of 2.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
88
6.3.5 MRI safety at 3T
As discussed in Chapter 2, MRI does not have known direct hazards from the main
magnetic field. However, the RF energy absorbed is converted to heat inside the
body. To prevent the excessive heat generation inside the patient, specific absorption
rates (SAR) are monitored and maintained within certain limits. For low field
strength scanners, SAR levels do not offer any limitations; however, high field
strength scanning has certain effects from increased SAR levels. SAR is proportional
to the square of the strength of the main magnetic field, as given below (6.5).
Therefore, at 3T, the SAR level is four times that at 1.5T:
DBSAR 2
0 )( 6.5
Where SAR = Specific absorption rate, B0 = strength of the main magnetic field, =
flip angle and D = the duty cycle.
6.4 Aims of the study
The study aimed to quantitatively compare the quality of the images obtained at 1.5T
to those acquired at 3T. The study also quantified the change of spin relaxation times
at 3T compared to 1.5T MRI.
6.5 Methods
6.5.1 Samples
Right legs of five human volunteers were scanned in 1.5T and 3T MRI scanners. The
calculation of sample size showed that for 80% of power a total of 20 samples need
to be used to detect a difference of 1.76 with standard deviation of 1.00. However,
this study had a sample size of five due to the very long processing time. Using the
sample number of five the difference that can be detected is 3.55 with the same
standard deviation.
6.5.2 Measuring the quality of MR images
Generally, the quality of an image is assessed using signal to noise ratio (SNR).
Signal to noise ratio is the proportion between the signal and the noise of the image;
thus, the higher the SNR, the better the image quality. SNR is one of the important
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
89
measures of the performance of a MRI system, as discussed in Section 6.2 [152]. The
equation below (Equation 6.6) gives the signal to noise ratio of any signal:
Noise
SignalSNR 6.6
In images, average pixel intensity of a sample from the ROI is taken as the signal and
the noise is calculated from the background of the image. Noise can either be the
average of the background intensity or the distribution of signal (Standard deviation
of the background intensity). In both cases, equal results can be obtained when
proper conversion factors from noise statistics are used [153]. With regards to SNR
of MRI, this method can be applied to calculate the SNR from a single MR image.
However, methods which utilise two or a series of images are also available [152,
154].
When SNR for a single MR image is calculated, a statistics derived factor
)4/(2 is introduced to the background noise [155]; thus, the equation can be
modified as in Equation 6.7:
noise
tissue
STD
MSNR
4
2
6.7
Where SNR = Signal to noise ratio, Mtissue = Average intensity of desired tissue type,
and STDnoise = Standard deviation of background noise.
Contrast to noise ratio (CNR) can be used to assess the contrast (intensity difference)
between two ROIs of image. This is an important measure used to assess or compare
the quality of MR images where higher contrast is required to identify certain tissue
types from the others; as an example, to identify the cortical bone from the
surrounding muscle tissue, the contrast between bone and muscle should be
considerably higher. In the basic form, CNR is the ratio between intensity difference
of two signals and the noise level (Equation 6.8):
Noise
SignalSignalCNR 21 6.8
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
90
As per SNR, the noise level can be calculated using the SD of the background; thus,
with addition of the noise statistics derived conversion factor, the CNR can be
calculated as in Equation 6.9:
noise
tissuetissue
STD
MMCNR
4
2
21 6.9
Where Mtissue1 = Average intensity of tissue type-1, Mtissue2 = average intensity of
tissue type-2, and STDnoise = Standard deviation of the average intensity of muscle.
6.5.3 Quantification of spin relaxation times
As per the discussion earlier in this chapter, T1 is highly field-dependent and, thus,
there is a need to use higher TR values in order to obtain comparable SNR values as
at 1.5T. Higher TR values, however, result in increased imaging time and, therefore,
there is a need to optimise the imaging protocol of 3T to obtain a better contrast
without losing the SNR. In order to calculate the T1 and T2 values of the muscle and
bone marrow, a series of images of the centre of the femur was obtained with varying
TR and TE values (Table 6.1) using 3T and 1.5T MRI scanners, while keeping the
other parameters constant.
Table 6.1 TR and TE values used for the MRI scanning at 1.5T and 3T for
calculation of T1 and T2
TR (ms) TE (ms)
1.5T 10 12 15 20 30 50 1.5T 4 5 6 8 10 12
3T 10 14 20 30 50 100 3T 4 5 7 9 12 14.8
The effect of different FAs on SNR was investigated by scanning the mid femoral
region of a volunteer with varying FAs (Table 6.2), while keeping all the other
parameters constant.
Table 6.2 Different flip angles used for scanning
Flip angle
1.5T 7 9 12 15
3T 7 9 12 15
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
91
T1 and T2 spin relaxation times were calculated for muscle and bone marrow
compartments from the images obtained, as described above. T1 and T2 were
determined, as indicated by the equations below:
s
S
NRTR
RF
NRTR
e
eSTRS
1
1
cos1
1)( 0 6.10
2
0)(RTE
eSTES 6.11
Where NS = 20 was the number of slices and R1 and S0 were the adjustable fit
parameters. The fitted value of R1 was taken as the longitudinal relaxation rate, 1/T1,
in the respective voxel. The fitted value of R2 was taken as the transverse relaxation
rate, 1/T2, in the respective voxel. SNR was also calculated for muscle, cortical bone
and bone marrow, while CNR was calculated at the muscle–cortical bone and bone
marrow–cortical bone interfaces. The results are included in the paper presented at
the end of this chapter.
6.5.4 Comparison of 1.5T and 3T imaging of musculoskeletal system
This section of the chapter describes the methods used for quantitative comparison of
the image quality of MRI at 1.5T to 3T, with special consideration to generation of
3D models of long bones. Comparison of the image quality was carried out using the
measures described in the previous section of this chapter dealing with SNR and
CNR.
The right legs of five healthy human volunteers were scanned using a 1.5T and 3T
MRI scanner from the same manufacturer. Identical protocols (Table 6.3) were used
for the scanning. Two RF coils (PA and Body matrix) were used to cover the lower
limbs completely (Figure 6.1).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
92
Table 6.3 The protocols used for MRI scanning
Parameter 1.5T 3T
In plane resolution 0.45 mm ×0.45 mm 0.45 mm ×0.45 mm Slice thickness 1 mm 1 mm
TR 11 ms 11 ms
TE 4.66 ms 4.66 ms
Flip angle 7° 7° Number of averages 1 1
FOV 176×256 176×256
Matrix size 352×512 352×512 Image sequence 3D FLASH 3D FLASH
Manufacturer Siemens Siemens
Model Magnetum Avanto Trio Tim
RF Coils PA & Body Matrix PA & Body Matrix
Figure 6.1: Positioning of the volunteer in the MRI scanner and the position of the
matrix coils that cover the lower limbs and the pelvis
In order to cover the lower limb completely, five scanning stages (Figure 6.2), each
containing 256 slices, were required, depending on the height of the subject. These
were positioned so that there were 66 slices of overlap between two successive
scanning stages. These overlapping regions are later used for the alignment of the 3D
models.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
93
Figure 6.2: Positioning of the field of view (FOV) on volunteer‟s leg
SNR and CNR were calculated for the proximal, diaphyseal and distal regions of
both tibia and femur. Image slices were obtained from nearly identical locations of
both 1.5T and 3T data sets. Five samples from each of muscle, cortical bone and
bone marrow were obtained using a customised Matlab script. The calculated SNR
and CNR values were used to compare 1.5T and 3T MR images. A detailed methods
section is available in the paper presented at the end of this chapter.
6.6 Results
The comparison between 1.5T and 3T images of the femora produced the following
results. In the mid diaphyseal region, CNR and SNR of muscles were higher for 3T
compared to 1.5T. In the proximal diaphyseal region, CNR and SNR of muscle at 3T
were slightly higher than 1.5T. In the distal diaphyseal region, CNR and SNR of the
other soft tissues were slightly higher at 3T than 1.5T; however; CNR and SNR of
muscles were slightly higher at 1.5T compared to 3T. For all regions, CNR and SNR
of medulla were higher in 1.5T compared to 3T.
In the tibia, the mid diaphyseal region had higher CNR and SNR for muscles at 1.5T
than at 3T. The distal diaphyseal region had higher CNR and SNR for muscle at 3T
compared to 1.5T. In both the regions, CNR and SNR of medulla were higher at
1.5T. All of the measurement sites, with the exception of two at articular regions of
both femur and tibia, had higher CNR and SNR for soft tissue at 3T than 1.5T. A
detailed results section is available in the paper presented at the end of this chapter.
6.7 Summary, discussion and conclusion
Since the introduction of MRI scanners to the clinical setting in the 1990s, a number
of studies and reviews were conducted that focused on the useability of 3T MRI
systems for scanning of various anatomical compartments, and on the various
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
94
advantages and disadvantages of the systems. Whilst most of these studies discussed
3T MRI scanning of the soft tissues, few studies have been conducted regarding
scanning of the skeletal system. As a radiation hazards free alternative to CT, 3T
MRI scanning for reconstruction of 3D models has become a research interest
because the higher magnetic field of 3T MRI can be used to potentially overcome
some of the limitations of the 1.5T MRI.
The present study was conducted to compare the images obtained at 1.5T and 3T,
using the instruments from the same manufacturer and identical imaging protocols.
The study resulted in higher SNR and CNR at 3T for most of the anatomical regions
compared to the 1.5T MRI. This will potentially improve the accuracy of the
articular regions of the segmented 3D models. According to the author‟s experience
of image segmentation, there is a potential for slight improvement of segmentation
time. The study had a sample size of five even though statistical power analysis
showed that a sample size of 20 is required to detect 1.76 difference between two
groups with standard deviation of 1.00. Sample size of five can detect a difference of
3.55 with the same standard deviation. Despite the small sample size, the study has
shown an overall improvement of image quality for 3T MRI compared to 1.5T MRI.
A samples size of twenty is not expected yield very different results considering that
the variability between the measurements are slightly high.
The spin relaxation times (T1 and T2) change at higher fields and the present study
showed that T1 of muscles is highly dependent on the main magnetic field and has
increased values at higher fields. T2 is relatively less dependent and takes slightly
less value at higher fields. Due to this increased T1 of muscles at 3T, theoretically,
longer TR values need to be used to obtain similar contrast to that at 1.5T and this
will, in turn, increase the scanning time. Due to this reason short TR values were
used for scanning of the human volunteers. However, for the purpose of comparison,
identical TR values were used in both 1.5T and 3T scanning protocols.
Higher magnetic fields also worsen most of the artefacts that occur in MRI scanning
of tissue compartments. The magnetic susceptibility artefact increases as higher
magnetic fields are more sensitive and this will affect the localisation of certain
anatomical structures of the images acquired. The chemical shift artefact also
becomes greatly pronounced as the frequency difference between water and fat
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
95
becomes double at 3T (220Hz at 1.5T and 440Hz at 3T). This increased frequency
difference, however, helps to achieve a better fat suppression of images. Increased
SAR values also should be considered with 3T MRI whereas, at 1.5T, SAR values
rarely increase beyond the limitations.
The next chapter discusses the investigation carried out to correct the step artefact
caused by volunteers moving their leg during the MRI of long bones.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
96
6.8 Paper 4: 3T MRI improves bone-soft tissue image contrast
compared with 1.5T MRI (Submitted – under review)
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
97
3T MRI improves bone-soft tissue image contrast compared
with 1.5T MRI
Kanchana Rathnayaka1, Konstantin I Momot
2, Alan Coulthard
3, Andrew Volp
4, Tony
Sahama2, Michael A. Schütz
1,4, Beat Schmutz
1
1. Institute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin
Grove, QLD 4059, Australia
2. Queensland University of Technology, Brisbane, Australia
3. Royal Brisbane and Women‟s Hospital, Brisbane, Australia
4. Princes Alexandra Hospital, Brisbane, Australia
Submitted to Journal: Magnetic Resonance Imaging
Manuscript ID: MRI-D-11-00307
Corresponding Author:
Dr. Beat Schmutz
60 Musk Avenue
Kelvin Grove
QLD 4059, Australia
Email: [email protected]
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
98
Abstract
Orthopaedic implants are designed using 3D models of long bones based on accurate
computed tomography (CT), which is the gold standard for scanning of bones.
However, CT exposes a healthy human volunteer to a high dose of ionising radiation;
thus, CT is generally limited to scanning of clinical cases and cadaver specimens.
Magnetic resonance imaging (MRI), on the other hand does not involve ionising
radiation and is therefore more appropriate for scanning of healthy human volunteers
for research purposes. Current limitations of MRI include poor contrast in certain
anatomical regions and long scanning times; these limitations can potentially be
overcome by using scanners with higher field strength. This study quantitatively
compares 1.5T MRI to 3T MRI and optimises the scanning protocol of 3T MRI for a
better outcome.
Protocol optimisation was carried out by scanning the right leg of one volunteer in
three sets of images, each with varying repetition times (TR), echo times (TE) and
flip angles (FA), while keeping the other parameters constant. Longitudinal
relaxation time (T1) of muscle and bone marrow and transverse relaxation time (T2)
of muscle were calculated for both 1.5T and 3T field strengths. To compare the
images acquired at 1.5T to 3T, the right legs of five human volunteers were scanned
with 1.5T and 3T scanners from the same manufacturer (Siemens), using identical
protocols. Signal to noise ratio (SNR) and contrast to noise ratio (CNR) were
calculated for different anatomical locations of femora and tibiae.
The results show that T1 of muscle is extremely dependent on the main magnetic
field (0.9 ± 0.14 s at 1.5T and 1.5 ± 0.15 s at 3T), yielding a higher value at 3T while
T1 of bone marrow was weakly field dependent (0.25 ± 0.03 s at 1.5T and 0.30 ±0.07
s at 3T). T2 of muscle was not field dependent and was measured as 0.029 ± 0.007 s
at both 1.5T and 3T. CNR and SNR comparison of 1.5T and 3T showed a high CNR
and SNR for most regions of the femur and tibia at 3T, with the exception of the
distal diaphyseal region of the femur and the mid diaphyseal region of the tibia. The
results show that 3T MRI is expected to reduce segmentation time and potentially
will improve the accuracy of 3D models generated from such data sets compared to a
3D model generated from a 1.5T data set.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
99
Introduction
Design and validation of orthopaedic implants increasingly utilises 3D models that
characterise the outer and inner geometry of long bones based on computed
tomography (CT) [1-5]. CT has become the gold standard for this purposes due to the
higher image contrast offered for the bone-soft tissue interface. CT, however,
exposes a subject to a high dose of ionising radiation thus limiting its use to scanning
of clinical cases and cadaver specimens. Due to this, imaging techniques such as
magnetic resonance imaging (MRI) which does not involve ionising radiation are
becoming more popular for scanning of long bones of healthy human volunteers for
research purposes. Some of the current limitations of using MRI for long-bone
imaging include extended scanning times and difficulty of image segmentation in
certain anatomical regions caused by the poor contrast at bone-soft tissue interfaces
of those regions [6, 7]. Higher field strength MRI scanners could potentially
overcome these limitations offering faster imaging times or better contrast levels [8].
In the present study, the gain obtained from the 3T system was solely invested in
improving the contrast at the bone-soft tissue interfaces.
Higher field strength MRI scanners (typically 3T) have been used clinically since the
1990s [9]. Since then, 3T MRI scanners have been validated for various soft tissue
compartments of the human body [10-15]. Because MRI utilises 1H nuclei as the
source of signal, it is rational to use it for studies involving soft tissues. While most
of these studies were qualitative, a few quantitative comparisons of the image quality
between 1.5T and 3T has also been reported [16]. However, to date, there are no
studies which have quantitatively compared 3T MRI with 1.5T MRI with regards to
generating 3D models of long bones.
The intrinsic signal to noise ratio (SNR) of a clinical MR system is approximately
proportional to the strength of the main magnetic field (B0) [17]. Thus, in principle, a
3T MRI system should offer twice the SNR that a 1.5T system offers if used with
equivalent parameters, receiver coils and subjects. However, the actual SNR of
acquired images is dependent on various other factors than on the doubled main
magnetic field: hardware design, change of tissue characteristics at higher fields (T1
and T2), increased sensitivity to magnetic susceptibility and the increased precession
frequency difference between fat and water. Therefore, optimisations of protocols
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
100
(TR, TE and FA) to meet the changed tissue characteristics and minimisation of
artefacts are important for a higher signal gain at 3T.
Longitudinal relaxation time (T1) and transverse relaxation time (T2) of a particular
tissue are two of the parameters that determine the contrast of the acquired MR
images. T1 is highly dependent on the main magnetic field (B0) and becomes longer
on increasing B0. The reported rise of T1 is about 20% for fatty tissue [10] and 40%
for muscle tissue at B0 = 3T compared to B0 = 1.5T [18]. This elongation of T1
reduces the signal intensity at shorter TRs and if similar values are used, the SNR at
3T is only slightly higher than at 1.5T. T2 is relatively less dependent on the external
magnetic field; however, about 10% reductions in T2 in certain tissue types have
been reported at 3T compared to 1.5T [18, 19]. Due to the different behaviour of spin
relaxation times at 3T, the repetition time (TR) and echo time (TE) should be
changed accordingly to obtain the maximum contrast levels. In general, a relatively
longer TR value and slightly shorter TE values should yield the optimal SNR in 3T
MR system.
SNR and contrast to noise ratio (CNR) are two most commonly used comparison
characteristics of MR images. SNR has long been used for evaluation of MR
systems, measurement of contrast enhancement, pulse sequences and RF coil
comparison [20]. CNR offers a meaningful way of comparing the contrast of the
bone-soft tissue interface of images, which is the most important feature responsible
for an accurate image segmentation, from two different field strength of MRI [21].
This study quantitatively compares the image quality at 1.5T to that at 3T using SNR
and CNR to compare the image quality. In addition, an investigation was carried out
to determine the optimum parameters to use with a FLASH (Fast Low Angle Shot)
sequence in the scanning of human volunteers.
Methods
MRI data acquisition
The MRI data for the first part of the investigation (SNR and CNR optimisation) was
acquired using 1.5T (Siemens Magnetom Avanto) and 3T (Siemens Trio Tim)
scanners. The optimisation was achieved by varying the values of one of TR, TE and
flip angle (FA), while keeping the values of the other parameters constant (Table 1).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
101
The right mid femoral region of one human volunteer was scanned using a 3D
FLASH sequence. The peripheral angiography (PA) matrix coil was used to cover
the lower limb.
Table 1: The imaging protocols used to scan the human volunteer with varying TE,
TR and FA values
Parameter Varying TR Varying TE Varying FA
TR Varying 16 ms 11 ms TE 5 ms Varying 4.66 ms FA 15° 15° Varying Pixel size 0.5 mm
2 0.5 mm2 0.5 mm
2 Slice thickness 5 mm 5 mm 1 mm
Data acquisition for the second part of the investigation was carried out by scanning
the right leg of five healthy male volunteers (age range: 30-54 years) with 1.5T and
3T clinical MRI systems. A customised imaging protocol (Table 2) that was
determined in accordance with the results obtained in the first part was used for the
scanning. The legs were scanned in five segments (Figure 1), moving the table so
that the centre of each segment was positioned in the centre of the magnet. One
scanning segment contained 256 image slices and a 66 slice overlap was maintained
between two successive scanning stages.
Table 2: The MRI imaging protocols for 1.5T and 3T scanners
Parameter 1.5T 3T
In plane resolution 0.45 mm ×0.45 mm 0.45 mm ×0.45 mm Slice thickness 1 mm 1 mm TR 11 ms 11 ms TE 4.66 ms 4.66 ms Flip angle 7° 7° Number of averages 1 1 Image sequence 3D FLASH 3D FLASH Manufacturer Siemens Siemens Model Magnetom Avanto Trio Tim RF Coils PA & Body Matrix PA & Body Matrix
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
102
Figure 1. Five imaging segments were used to scan the lower limb completely in 3T
and 1.5T MRI scanners.
Measurement of spin relaxation times
T1 and T2* spin relaxation times of the „muscle‟ and „bone marrow‟ compartments
were measured from the appropriate series of FLASH images [22]. For the
measurement of T1, the RF excitation pulse was set to RF = 15o, the gradient echo
time to 5 ms, the number of averages to 1, and a series of TR were used. At 1.5T, the
TR values used were 10, 12, 15, 20, 30 and 50 ms. At 3T, the TR values used were
10, 14, 20, 30, 50 and 100 ms. For each voxel within the image, the T1 value was
determined from a two-parameter nonlinear least-squares fit of the intensity of the
steady-state FLASH signal as a function of TR:
(1)
Where NS = 20 was the number of slices and R1 and S0 were the adjustable fit
parameters. The fitted value of R1 was taken as the longitudinal relaxation rate, 1/T1,
in the respective voxel. The standard errors of the fitted R1 and image amplitude ( R1
and S0) were also determined for each voxel. The voxels where any of the following
conditions were observed were rejected: R1 > 0.5 R1; S0 > 0.5 S0; R1 < 0; S0 < 0.
The average R1 values in the „muscle‟ and „bone marrow‟ compartments were then
determined by averaging the fitted R1 values over the „non-rejected‟ voxels within
the appropriate ROI (~1000 voxels for the muscle and ~300 voxels for the marrow).
For the measurement of T2*, the RF excitation pulse was set to RF = 15
o, the
repetition time to 16 ms, the number of averages to 1, and a series of gradient TE
values were used. At 1.5T, the TE values used were 4, 5, 6, 8, 10 and 12 ms. At 3T,
s
S
NRTR
RF
NRTR
e
eSTRS
1
1
cos1
1)( 0
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
103
the TE values used were 4, 5, 7, 9, 12 and 14.8 ms. For each voxel within the image,
the T2 value was determined from a two-parameter nonlinear least-squares fit of the
intensity as a function of TE:
(2)
The fitted value of R2* was taken as the apparent transverse relaxation rate, 1/T2
*, in
the respective voxel. Fit quality control and T2* averaging over muscle and marrow
were performed as described above for the T1. T1 and T2* processing was performed
using custom-written Mathematica (Wolfram Research, Champaign, IL, USA) code
running on a desktop PC.
SNR was calculated for the „muscle‟ and „bone marrow‟ tissue types of image series.
ROIs were selected at five sites of an image slice of each of the image stack, as
indicated in Figure 4-b. The SNR was calculated using the method described in the
next section.
SNR and CNR for comparison of MR images
SNR and CNR were used to compare image quality between images obtained from
1.5T and 3T scanners. CNR is one of the most important parameters as the contrast
between bone and the soft tissue is the key feature that is responsible for an accurate
segmentation of the bone. In the basic form, SNR is the ratio of a signal to the
background noise, while CNR is the ratio of contrast to the background noise
(Equations 3 & 4).
(3)
(4)
When the equations were applied to the MR images, the mean intensity of the
specific tissue type was considered as the signal and the standard deviation of the
background was considered as the noise level [19]. The noise statistics derived
correction factor [20] was introduced to standardise the SNR and CNR
*2
0)(RTE
eSTES
Noise
SignalSNR
Noise
SignalSignalCNR 21
)4/(2
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
104
values derived. In this study, the background noise level was not measured due to the
unevenly distributed noise in background and thus the noise level of the cortical bone
was used to calculate the SNR and CNR [20]. Thus, with the noise statistics derived
factors, the equations used to calculate the SNR and CNR of MR images were as
follows:
(5)
(6)
Where, SNR = signal to noise ratio, CNR = contrast to noise ratio, Mtissue = Mean
intensity of the tissue and STDbone = standard deviation of mean intensity of cortical
bone.
Comparison of the images obtained from 1.5T with 3T MRI
SNR and CNR measurements were taken in the proximal articular, proximal
diaphyseal, mid diaphyseal, distal diaphyseal and distal articular regions of the femur
and of the tibia. In diaphyseal regions (Figure 2), SNR and CNR were measured
(Figure 3) at five sites around the bone (Figure 4-b) in axial image slices. In articular
regions, a varying number of sites were used (Figure 4-a, c, d, e & f) and coronal
sections were used, with the exception of the distal articular region of the femur for
which axial images were used. The measurements were taken in three consecutive
image slices at any given site.
bone
tissue
STD
MSNR
4
2
bone
tissuetissue
STD
MMCNR
4
2
21
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
105
Figure 2. The diaphyseal regions of femur (top) and tibia (bottom) where the axial
image slices were obtained for the calculation of SNR and CNR.
Figure 3: In each site of the diaphyseal regions, pixel samples were obtained from
bone marrow, cortical bone, and Muscle.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
106
Figure 4. ROIs selected at four/five positions in each tissue type in: a-femoral head,
b- mid femoral diaphysis, c- distal femoral diaphysis, d- distal femoral articular, e-
proximal tibial articular, and f- distal tibial articular regions as shown in the figure
(left and right images are from two different planes).
SNR and CNR were measured in muscle, bone and bone marrow tissue types at
diaphyseal regions of both femora and tibiae, with the exception of the distal
diaphyseal region of the femora where the bone was surrounded by other soft tissues
in addition to the muscle tissue (mainly fat and fibrous tissue). In this region the
measurements were taken in these soft tissues in addition to muscle. In the articular
regions, the bone does not come into contact with the muscle tissue but with various
other tissues such as fat, tendons, fibrous capsules and synovial fluid. Moreover, the
articular regions no longer contain bone marrow, and the medulla is basically
composed of a mixture of trabecular bone and bone marrow. Thus, the measurements
were taken in soft tissues, bone and the medulla.
Statistical differences of SNR and CNR values between the 1.5T and 3T images were
calculated using one way ANOVA. The level of statistical significance was set to p ≤
0.05. The validation was performed using PASW Statistics 18 software package.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
107
Results
The measurement of the longitudinal relaxation time (T1) and the apparent transverse
relaxation time (T2*) was carried out using a series of images acquired with varying
TR and varying TE values, respectively. The measured T1 value of the muscle was
1.5 ± 0.2 s at 3T and 0.9 ± 0.1 s at 1.5T. The measured T1 values of the voxels in the
bone marrow compartment were 0.25 ± 0.03 s at 1.5T and 0.30 ± 0.07 s at 3T. The
apparent transverse relaxation time, T2*, of the muscle was measured as 0.029 ±
0.007 s at both 1.5T and 3T. The T2* of the bone marrow could not be measured
reliably.
The SNR calculation of the images obtained with varying TR, TE and FA values
(Figure 5) showed the following trends. SNR of muscle and bone marrow increased
with the TR while SNR of muscle and bone marrow declined with the TE in both
1.5T and 3T filed strengths; and SNR of muscle had downward trend with FA, while
SNR of bone marrow had upward trend with the FA in both 1.5T and 3T field
strengths.
Figure 5. Change of SNR with varying TR, TE and FA at 1.5T and 3T.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
108
The comparison between 1.5T and 3T images of the femora produced the following
results (Figure 6). In the mid diaphyseal region 3T had the highest CNR and SNR for
muscles (CNR = 4.49, 7.29 and SNR = 7.50, 10.00 for 1.5T and 3T respectively) and
1.5T had the highest CNR and SNR for bone marrow (CNR = 6.49, 5.70 and SNR =
9.66, 8.67 respectively for 1.5T and 3T). In the proximal diaphyseal region, CNR and
SNR of muscle at 3T were slightly higher than 1.5T and CNR and SNR of bone
marrow was higher at 1.5T. In the distal diaphyseal region, CNR and SNR of the
other soft tissues were slightly higher at 3T (CNR = 4.74, SNR = 6.97) than 1.5T
(CNR = 4.54, SNR = 6.96); however, CNR and SNR for muscles were slightly
higher at 1.5T compared to 3T. For the same region, CNR and SNR of medulla were
higher in 1.5T compared to 3T.
Figure 6: CNR and SNR of diaphyseal regions of femur (TR = 11 ms and TE = 4.66
ms at both 1.5T and 3T, * = statistically significant).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
109
SNR and CNR measurements of four sites at the proximal articular region and five
sites at the distal articular region show that 3T MRI gives higher SNR and CNR for
all the regions with the exception of region -4 of the distal articular region that has
higher SNR and CNR for 1.5T (Figure 7). Images illustrating the improvement in
image contrast are shown in Figure 8.
Figure 7. Proximal and distal articular regions of the femur (TR = 11 ms and TE =
4.66 ms at both 1.5T and 3T, * = statistically significant).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
110
Figure 8. Comparison of 1.5T images to 3T images of the proximal region (top) and
the mid shaft (bottom) of the femur (TR = 11 ms and TE = 4.66 ms at both 1.5T and
3T).
In tibia, the proximal diaphyseal region, muscles presented higher SNR and CNR
values for 3T MR images while medulla showed similar SNR and CNR values for
both 1.5T and 3T. For the mid diaphyseal region; however, 1.5T showed higher SNR
and CNR than 3T (SNR = 15.4 and 14.5, CNR = 13.3 and 12.4 respectively for 1.5T
and 3T). For the distal diaphyseal region higher CNR and SNR was reported for 3T
(Figure 9).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
111
Figure 9. CNR and SNR values of diaphyseal regions of tibia femur (TR = 11 ms and
TE = 4.66 ms at both 1.5T and 3T, * = statistically significant).
CNR and SNR measured at four sites in both the proximal articular region and the
distal articular region of tibiae showed higher CNR and SNR for 3T images with the
exception of the region -3 of the distal articular region (Figure 10).
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
112
Figure 10. CNR and SNR of articular regions of tibia (TR = 11 ms and TE = 4.66 ms
at both 1.5T and 3T, * = statistically significant).
Discussion
The study aimed to quantitatively compare the MR image quality at two applied
magnetic field strengths, 1.5T and 3T, using the femora and tibiae of five healthy
volunteers as the study sample and SNR and CNR as the comparison parameters. An
investigation was also carried out to optimise the imaging protocol at 3T by
identifying the optimum TR, TE and FA values at that field strength. The effect of
the magnetic field on the T1 and T2 of the tissues imaged (muscle and bone marrow)
was also investigated.
The T1 of the muscle was strongly dependent on the applied magnetic field strength
(B0): The T1 at 3T (1.5 s) was more than 50% longer than that at 1.5T (0.9 s). The
apparent T1 values of the bone marrow exhibited significantly weaker field
dependence, with the apparent T1 values at the two fields differing by ~15% (0.25 s
at 1.5T and 0.30 s at 3T). (We use the term „apparent T1‟ for the bone marrow
because no fat suppression was used, and the measured T1 in this tissue can therefore
include contributions from both lipid and water.) The lengthening of the T1 values of
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
113
both tissues with the increasing B0 is consistent with the well-established body of
knowledge concerning the relaxometry of biological tissues [18, 23, 24]. It is also
consistent with the fact that longitudinal relaxation is controlled by fast molecular
motions[25]; that is, motions whose time scale is comparable to the Larmor
precession frequency of the MRI systems used in this study (~10 ns). The relatively
small increase of the apparent T1 of bone marrow can be attributed to the relatively
low mobilities of lipid molecules and water molecules in a lipid-rich environment.
This observation is consistent with the field dependence of the T1s of lipid and water
protons previously observed in a model lipid/water system [26].
The apparent transverse relaxation time, T2*, of the muscle exhibited no discernible
dependence on the applied magnetic field. This observation can be rationalised as
follows. T2* is a complicated function dependent on the local inhomogeneities of the
static magnetic field, slow molecular motions, fast molecular motions, and chemical
exchange between „free‟ and „bound‟ states of water molecules. (The last three
factors determine the true transverse relaxation time, T2.) The four factors listed
serve to shorten, shorten, lengthen, and shorten T2* with the increasing B0,
respectively [27]. The true T2 in muscle has been reported variously to become
slightly shorter [19] or slightly longer [18] with the increasing B0. Under the
conditions of the present study, the effects of the four factors listed evidently nearly
cancel each other out, resulting in the absence of a significant field dependence of
T2*.
When a 3D FLASH sequence was used, the SNR of both muscle and bone marrow
increased upon increasing TR from 10 ms to 50 ms. Beyond 50 ms, SNR at 3T
started to decline and TR > 50 ms was not used with 1.5T imaging due to practical
difficulties of setting up the scanner with TR = 100 ms. Even though a higher SNR
can be obtained at higher TR, doubling TR in turn doubles the scanning time.
Compared to the scanning time of 65 minutes to scan the complete lower limb with
TR = 11, TR beyond this would result in extremely long scanning times that are
impracticable in the clinical environment, but also due to the increased risk of motion
artefacts resulting from long scanning times.
With increased TE, SNR dropped in both muscle and bone marrow tissues at 1.5T
and in muscle tissue at 3T; however, SNR increased in bone marrow at 3T (Figure
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
114
5). FA presented converse SNR results for muscle and bone marrow. At both 1.5T
and 3T, SNR of muscles declined on increasing FA whereas SNR of bone-marrow
increased. Based on the results obtained, the protocol used to scan human volunteers
was determined to have TR = 11, TE = 4.66 and FA = 7 for both 1.5T and 3T
scanners.
Comparison of images obtained at 1.5T to 3T showed that, in general, 3T MRI
generates images with a high contrast between bone-muscle and bone-soft tissue
interfaces. The mid diaphyseal region of the femur, and the proximal and distal
diaphyseal regions of the tibia, presented a greater increase in CNR and SNR in the
bone-muscle interface, while the proximal diaphyseal region of the femur showed
slight increase. Among them, the mid diaphyseal region of the femur showed
statistically significant increase in SNR and CNR at 3T. The distal diaphyseal region
of the femur and the mid diaphyseal region of the tibia did not show any increase in
CNR or SNR for muscle at 3T, the reason for this could not be determined. However,
there was a slight increase in CNR and SNR at soft tissue-bone interface of the distal
diaphyseal region of the femur. The reason why CNR and SNR were lower in these
regions could not be determined.
CNR at the bone marrow-bone interface was higher at 1.5T than 3T in all the cases
and this was statistically significant in mid diaphyseal region of tibia. As mentioned
at the beginning of the discussion, T1 of bone marrow (0.25 ± 0.03 s at 1.5T and 0.30
±0.07 s at 3T) is comparatively shorter than T1 of muscle tissue (0.86±0.14 s and
1.5±0.15 s at 3T). As the extremely short TR value (11 ms) have been used for both
1.5T and 3T scanning, tissues with longer T1 (muscle in this case) produce a low
signal due to inadequate recovery of the transverse component of the net
magnetisation vector. This is the main reason why bone marrow produced a higher
signal compared to the muscle. The low CNR of bone-bone marrow interface at 3T is
unlikely to affect the segmentation process as the obtained CNR is sufficient for an
accurate segmentation of the medullary canal. Compared to the outer cortex, the
inner cortex has a relatively simple, bone-bone marrow interface in the medullary
canal.
Articular regions of both the femur and tibia showed increased CNR and SNR for
3T, with the exception of one site in the distal articular region of the femur and the
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
115
second site at the distal articular region of the tibia. These differences were
statistically significant in two regions in each of the proximal articular region of
femur, proximal articular region of tibia and distal articular region of tibia for CNR.
In distal articular region tibia, the differences were also statistically significant for
SNR at two sites. The reason for this difference in CNR and SNR could be due to the
number of different interfaces present at the articular regions (bone-ligament, bone-
tendon, bone-synovial fluid, bone-synovial membrane and bone-cartilage). These
different tissue types exhibit different MRI properties (T1, T2* and proton density)
that result in various contrast levels at the articular regions. This increase the partial
volume effect at articular regions and therefore only the average CNR and SNR can
be measured in these regions. However, increased CNR at most of the sites of the
articular regions will potentially facilitate the segmentation process by improving the
accuracy, which was a problem in 1.5T MR imaging of those regions.
Overall, 3T MRI generated images with higher quality for most of the anatomical
regions of the femur and tibia. Even though the theoretical doubling of SNR gain is
not achievable due to the practical reasons, the articular regions had impressively
higher CNR and SNR values. These are the regions where segmentation at 1.5T was
difficult and this increased CNR is expected to significantly facilitate the
segmentation process of the articular regions [6, 7]. CNR and SNR of distal femur
and mid diaphysis of tibia were not improved; however, these regions could be
segmented accurately with 1.5T images [6]. At the same time, the obtained higher
contrast levels at bone-muscle and bone–bone marrow interfaces will potentially
improve the accuracy of segmentation and in addition decrease the time required for
the segmentation. This study investigated the improvement of the image contrast by
using higher field strength MRI. Another important aspect that needs to be improved
through future studies is the scanning time, which is considerably longer compared to
CT at present.
Acknowledgement
This research was supported in part by Synthes GmbH. The last author has received
an industrial scholarship from Synthes GmbH.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
116
References
[1]. Schmutz B, Rathnayaka K, Wullschleger ME, Meek J, Schuetz MA.
Quantitative fit assessment of tibial nail designs using 3D computer
modelling. Injury 2010;41(2):216-219.
[2]. Schmutz B, Wullschleger ME, Kim H, Noser H, Schuetz MA. Fit Assessment
of Anatomic Plates for the Distal Medial Tibia. Journal of Orthopaedic
Trauma 2008;22(4):258-263.
[3]. Schmutz B, Wullschleger ME, Noser H, Barry M, Meek J, Schuetz MA. Fit
optimisation of a distal medial tibia plate. Computer Methods in
Biomechanics & Biomedical Engineering 2011;14(4):359-364.
[4]. Park A, DiStefano J, Nguyen T, Congruency of Scapula Locking Plates
Implications for Implant Design, in 56th Annual Meeting of the Orthopaedic
Research Society 2010: New Orleans, LA. p. Poster No. 1746.
[5]. Kozic N, Weber S, Büchler P, Lutz C, Reimers N, González Ballester MÁ,
Reyes M. Optimisation of orthopaedic implant design using statistical shape
space analysis based on level sets. Medical Image Analysis 2010;14(3):265-
275.
[6]. Rathnayaka K, Momot KI, Noser H, Volp A, Schuetz M, Sahama T, Schmutz
B. Quantification of the accuracy of MRI generated 3D models of long bones
compared to CT generated 3D models. Medical Engineering & Physics
2011(in press DOI: 10.1016/j.medengphy.2011.07.027).
[7]. Schmutz B, Volp A, Momot K, Pearcy M, Schuetz M. Using MRI for the
imaging of long bones: First Experience. Journal of Biomechanics
2008;41(Supplement1):S188.
[8]. Edelstein WA, Glover GH, Hardy CJ, Redington RW. The intrinsic signal-to-
noise ratio in NMR imaging. Magnetic Resonance in Medicine
1986;3(4):604-618.
[9]. Soher BJ, Dale BM, Merkle EM. A Review of MR Physics: 3T versus 1.5T.
Magnetic Resonance Imaging Clinics of North America 2007;15(3):277-290.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
117
[10]. Rakow-Penner R, Daniel B, Yu H, Sawyer-Glover A, Glover GH. Relaxation
times of breast tissue at 1.5T and 3T measured using IDEAL. Journal of
Magnetic Resonance Imaging 2006;23(1):87-91.
[11]. Tsurusaki M, Semelka RC, Zapparoli M, Elias Jr J, Altun E, Pamuklar E,
Sugimura K. Quantitative and qualitative comparison of 3.0 T and 1.5 T MR
imaging of the liver in patients with diffuse parenchymal liver disease.
European Journal of Radiology 2009;72(2):314-320.
[12]. Ramalho M, Altun E, Herédia V, Zapparoli M, Semelka R. Liver MR
Imaging: 1.5T versus 3T. Magnetic Resonance Imaging Clinics of North
America 2007;15(3):321-347.
[13]. Frayne R, Goodyear BG, Dickhoff P, Lauzon ML, Sevick RJ. Magnetic
resonance imaging at 3.0 Tesla: Challenges and advantages in clinical
neurological imaging. Investigative Radiology 2003;38(7):385-402.
[14]. Beyersdorff D, Taymoorian K, Knosel T, Schnorr D, Felix R, Hamm B,
Bruhn H. MRI of Prostate Cancer at 1.5 and 3.0 T: Comparison of Image
Quality in Tumor Detection and Staging. American Journal of Roentgenology
2005;185(5):1214-1220.
[15]. Stehling C, Langer M, Bachmann R, Kraemer S, Kooijman H, Heindel W,
Vieth V. Three-Tesla Magnetic Resonance Imaging of the Wrist: Diagnostic
Performance Compared to 1.5-T. Journal of Computer Assisted Tomography
2009;33(6):934-939.
[16]. Duewell SH, Ceckler TL, Ong K, Wen H, Jaffer FA, Chesnick SA, Balaban
RS. Musculoskeletal MR imaging at 4 T and at 1.5 T: comparison of
relaxation times and image contrast. Radiology 1995;196(2):551-555.
[17]. Gold GE, Suh B, Sawyer-Glover A, Beaulieu C. Musculoskeletal MRI at 3.0
T: Initial Clinical Experience. American Journal of Roentgenology
2004;183(5):1479-1486.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
118
[18]. Stanisz GJ, Odrobina EE, Pun J, Escaravage M, Graham SJ, Bronskill MJ,
Henkelman RM. T1, T2 relaxation and magnetization transfer in tissue at 3T.
Magnetic Resonance in Medicine 2005;54(3):507-512.
[19]. Kuo R, Panchal M, Tanenbaum L, Crues JV. 3.0 Tesla imaging of the
musculoskeletal system. Journal of Magnetic Resonance Imaging
2007;25(2):245-261.
[20]. Dietrich O, Raya JG, Reeder SB, Reiser MF, Schoenberg SO. Measurement
of signal-to-noise ratios in MR images: Influence of multichannel coils,
parallel imaging, and reconstruction filters. Journal of Magnetic Resonance
Imaging 2007;26(2):375-385.
[21]. Chang G, Friedrich KM, Wang L, Vieira RLR, Schweitzer ME, Recht MP,
Wiggins GC, Regatte RR. MRI of the wrist at 7 tesla using an eight-channel
array coil combined with parallel imaging: Preliminary results. Journal of
Magnetic Resonance Imaging 2010;31(3):740-746.
[22]. Haase A, Frahm J, Matthaei D, Hanicke W, Merboldt K. FLASH imaging.
Rapid NMR imaging using low flip-angle pulses. Journal of Magnetic
Resonance 1986(67):258-66.
[23]. Koenig SH, Brown RD, Relaxometry of Tissue, in Encyclopedia of Magnetic
Resonance. 1996, John Wiley & Sons, Ltd.
[24]. Koenig SH, Brown RD. Field-cycling relaxometry of protein solutions and
tissue: Implications for MRI. Progress in Nuclear Magnetic Resonance
Spectroscopy 1990;22(6):487-567.
[25]. Abragam A. The Principles of Nuclear Magnetism. 1961: Oxford University
Press.
[26]. Momot KI, Takegoshi K, Kuchel PW, Larkin TJ. Inhomogeneous NMR Line
Shape as a Probe of Microscopic Organization of Bicontinuous Cubic Phases.
The Journal of Physical Chemistry B 2008;112(21):6636-6645.
Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models
119
[27]. Cavanagh J, Fairbrother WJ, Palmer AG, Rance M, Skelton NJ. Protein NMR
Spectroscopy: Principles and Practice. 2nd ed. 2007, Burlington: Elsevier
Academic Press.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
121
Chapter 7 Step artefact caused by Magnetic
Resonance Imaging of long bone
7.1 Introduction
MR imaging of the musculoskeletal system is affected by various artefacts such as
motion artefacts, chemical shift artefact, and magnetic susceptibility artefact (Some
of these important artefacts have been discussed in Section 2.3.7). The motion
artefacts (also referred to as the „movement artefact‟) occur due to the random or
periodic movements of anatomical structures, resulting in blurred images and
inaccuracies to the 3D models reconstructed from such image data. In an initial
study, the supervisory team observed a step in the 3D model reconstructed from a
data set obtained from the lower limb of a human volunteer that might have resulted
from the volunteer moving the leg between two successive scanning stages [26]. In
orthopaedic implant design, these artefacts can affect the design of anatomically
well-fitting implants or their accurate validation.
Figure 7.1: The step artefact caused by volunteer moving the leg between two
successive scanning stages
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
122
Motion artefacts due to periodic movements can be eliminated by synchronising the
data acquisition with the movement, or by using post processing techniques.
However, artefacts due to random movements cannot be eliminated with such
techniques though radial K-space techniques are now available on clinical scanners
to combat such motion artefacts to some degree. Since the step artefact has been
observed in the reconstructed 3D models, the artefact can be eliminated using a 3D
model aligning technique such as the iterative closest point (ICP) algorithm. This
was successfully used in this study.
This chapter is focused on correction of the step artefact of 3D models based on
MRI. Section 7.2 will discuss the literature relevant to motion artefacts. Section 7.4
briefly introduces the methods used in the study and section 7.4 presents a summary
of the study.
7.2 Motion artefact of MRI
Motion artefacts are one of the challenges that researchers have faced when MRI is
used for 3D reconstruction of long bones. This manifests as signal misregistration
along the phase encoding direction, and the appearance may vary with the type and
rate of the movement [40, 44]. The artefacts are caused by tissue excited at one
location producing signals that are mapped to a different location during the data
acquisition [40]. Motion artefacts in MR imaging are basically of two categories. The
first category is the motion artefacts that occur due to periodic movements such as
respiration, heart beat or flow of blood and cerebrospinal fluid. The second category
is due to random movements such as the movements occur by the person‟s inability
to keep the limbs still for long scanning duration or muscle contraction due to nerve
stimulation from rapid change of the imaging gradients.
The motion artefacts due to periodic movements have minimum or no effect for
scanning of long bones of lower limbs, although scanning of the upper limbs might
be affected by respiratory movements. The artefacts due to random movements,
however, affect the MR imaging of the long bones of lower limbs. The lateral shift of
the 3D models is one of the artefacts resulting from random movements of the lower
limb. Due to the limitation of scanning length caused by the non-uniform magnetic
field, the scanning of a long bone (e.g. femur or tibia) is conducted in several
segments (Figure 7.2).
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
123
Figure 7.2: MRI scanning of human lower limb with five scanning segments to scan
the complete limb
Correction of the motion artefacts is generally achieved by synchronising the data
acquisition with the movement, or by post processing the data; however, this is
feasible only in the case of periodic movements [156, 157]. Artefacts due to random
movements are hard to correct and different techniques have to be used, depending
on the type of artefact. Immobilisation of the limb is one of the practices that can be
used in the clinic; however, the muscle contractions due to the nerve excitations from
RF waves cannot be prevented. Since this study is focused on correcting the lateral
shift of 3D models, the use of 3D modelling technique is possible. The ICP algorithm
is a robust method used to align 3D surfaces utilising the geometric features [158].
The ICP algorithm and the 3D -3D aligning process are described in Section 3.5.
7.3 Aims of the study
This study aimed at correcting the step artefact that occurs due to the random
movement of the lower limb, using the robust ICP algorithm based 3D modelling
techniques.
7.4 Methods
Five ovine hind limbs amputated from the pelvic and the ankle joint were used with
intact soft tissue. The statistical sample size analyse shows that five samples would
detect a difference of 0.07 mm with standard deviation of 0.02 for 80% power. The
femora of the sheep hind limbs were scanned using a 3T MRI scanner with a
customised protocol. Scanning was basically conducted to simulate the lateral shift
artefact incurred by the random movements, which were achieved by shifting the
bone laterally after scanning the first half of the bone. The artefact was corrected
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
124
using 3D modelling techniques to align the 3D models reconstructed from the two
halves of the scanned bone. In addition, the errors resulting from the table movement
were also quantified. A detailed description of the methods is available in the paper
presented at the end of this chapter.
7.5 Results
When the models with the corrected shift artefact were compared to the reference
models, an average error of 0.32 ± 0.02 mm was generated. The 3D models
reconstructed from the single MRI scan generated an error of 0.25 ± 0.02 mm. A
detailed results section is available in the paper presented at the end of the chapter.
7.6 Summary, discussion and conclusion
The motion artefacts occurring as a result of random movements play an important
part when MRI data is acquired from long bones (mainly) for 3D reconstruction of
long bones. Such an artefact causes the 3D models to have a step between two
successive scanning segments. Unlike the movement artefacts due to periodic
movements, the artefacts due to random movements cannot be eliminated by
synchronising the data acquisition or by post processing techniques.
Since the artefact is observed once the 3D model is reconstructed, a 3D surface
aligning method is feasible to correct the artefact. The ICP algorithm is a robust and
widely used method for 3D-3D alignment and was successfully implemented in this
study for the correction of the step artefact. The results show that the geometric
deviation of the corrected model is within the accepted accuracy levels for implant
design. This error was slightly higher than the error obtained for the MRI based
model reconstructed from the single scan. This residual error might have resulted
from the slight mal-alignment between proximal and distal halves models. Statistical
analysis of the sample size showed that this residual error (0.07 mm) with standard
deviation of 0.02 could be detected statistically with the sample size of five.
The study showed that by using the ICP algorithm, the step artefact observed in the
3D models of long bones can be corrected with sufficient accuracy to allow
researchers to design orthopaedic implants using the 3D models generated from
MRI. The present study utilised one simulated lateral shift; however, the human long
bones have to be scanned in at least three stages, resulting in two lateral shift
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
125
artefacts. This might introduce higher errors to the corrected 3D models. Therefore,
further validation of this method with human long bones has to be conducted before
using it for correction of artefacts in human bone models. In this study the correction
of the artefact was performed manually, using commercially available software and
this is a labour intensive process. Hence, automatic processing to correct the artefact
will be desirable in future.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
126
7.7 Paper 5: Correction of step artefact associated with MRI
scanning of long bones (Submitted – under review)
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
127
Correction of step artefact associated with MRI scanning of
long bones
1Kanchana Rathnayaka,
2Gary Cowin,
1,3Michael A Schuetz,
4Tony Sahama,
1Beat
Schmutz
1Institute of Health and Biomedical Innovation, Brisbane, QLD, Australia
2University of Queensland, St Lucia, QLD, Australia
3Princess Alexandra Hospital Brisbane, QLD, Australia
4Queensland University of Technology Brisbane, QLD, Australia
Submitted to Journal: Medical Engineering and Physics
Manuscript ID: MEP-D-11-00529
Corresponding Author:
Dr. Beat Schmutz
60 Musk Avenue
Kelvin Grove
QLD 4059, Australia
Email: [email protected]
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
128
Abstract
Magnetic resonance imaging (MRI) has been shown to be a potential alternative to
computed tomography (CT) for scanning of volunteers for 3D reconstruction of long
bones, essentially avoiding the high radiation dose from CT. In MRI imaging of long
bones, the artefacts due to random movements of the skeletal system create
challenges for researchers as they generate inaccuracies in the 3D models generated
by using data sets containing such artefacts.
One of the defects that have been observed during an initial study is the lateral shift
artefact occurring in the reconstructed 3D models. This artefact is believed to result
from volunteers moving the leg during two successive scanning stages (The lower
limb has to be scanned in at least five stages due to the limited scanning length of the
scanner). As this artefact creates inaccuracies in the implants designed using these
models, it needs to be corrected before the application of 3D models to implant
design. Therefore, this study aimed to correct the lateral shift artefact using 3D
modelling techniques.
The femora of five ovine hind limbs were scanned with a 3T MRI scanner using a
3D VIBE based protocol. The scanning was conducted in two halves, while
maintaining a good overlap between them. A lateral shift was generated by moving
the limb several millimetres between two scanning stages. The 3D models were
reconstructed using a multi threshold segmentation method. The correction of the
artefact was achieved by aligning the two halves using the robust iterative closest
point (ICP) algorithm, with the help of the overlapping region between the two. The
models with the corrected artefact were compared with the reference model
generated by CT scanning of the same sample.
The results indicate that the correction of the artefact was achieved with an average
deviation of 0.32 ± 0.02 mm between the corrected model and the reference model.
In comparison, the model obtained from a single MRI scan generated an average
error of 0.25 ± 0.02 mm when compared with the reference model. An average
deviation of 0.34 ± 0.04 mm was seen when the models generated after the table was
moved were compared to the reference models; thus, the movement of the table is
also a contributing factor to the motion artefacts.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
129
Introduction
Magnetic resonance imaging (MRI) is theoretically designed to scan the soft tissues
utilising the hydrogen nuclei as the source of signal. In recent studies, it has been
shown to be a possible alternative to computed tomography (CT) for scanning of
long bones [1, 2]. This alternative provides researchers designing orthopaedic
implants with an opportunity to acquire long bone image data from the young healthy
human population, who represent nearly half of all trauma patients, without having to
expose them to the ionising radiation of CT [3]. However, MRI still suffers from
some limitations such as very long scanning times, motion artefact and poor contrast
in certain anatomical regions. Of these limitations, the motion artefact is crucial as it
reduces the accuracy of the 3D models reconstructed from such image data [2]. A
lateral shift has been observed in the 3D models reconstructed from data sets; this is
believed to occur as a result of random patient movements [4].
The design of an orthopaedic implant needs accurate 3D representations of the
relevant bone geometry. The current gold standard for acquisition of data for this
purpose, CT, exposes a person to a high dose of ionising radiation. This exposure
limits CT to the scanning of cadaver bones which are, in most cases, more than 60
year old. Since most of the patients who have been implanted with a plate or
intramedullary nail are from the younger population, the implants need to be
designed to suit this age group. For this purpose, there is an urgent need for the
acquisition of data from this younger population. MRI is a versatile alternative for
this purpose as there are no radiation hazards involved in MRI scanning. The poor
contrast of certain anatomical regions in the MRI scanning of long bones can be
overcome to some extent by using a higher field MRI scanner [2]. However, artefacts
due to random movements of a subject remain a problem which needs to be
addressed in order to utilise the models for the intended application.
The motion artefacts occur when the protons of the tissue sample being scanned
excited at one site are misregistered to another region of the image during the data
acquisition [5]. This results in repeated reconstruction of the moving structures along
the phase encoding direction [6]. The motion artefacts are of two types: the artefacts
due to, periodic movements and the artefacts due to random movements [6]. The
motion artefacts due to regular, periodic movements occurs (mainly) as a result of
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
130
respiratory movements, flow of blood in vessels or peristalsis. These artefacts mainly
affect the MRI scanning of the relevant anatomical regions. Although respiratory
movements might have an impact on the scanning of long bones of the upper limb,
they have minimal or no effect on the scanning of the long bones of the lower limbs
(e.g. the femur and tibia). The artefacts resulting from random movements may be
due to nerve excitation during the scanning, or the patient randomly moving the limb.
Due to the limited linearity of the gradients and B0 field of the MRI scanner, the
lower limb of a subject has to be scanned in several stages (usually four to five). A
preliminary investigation conducted by Schmutz et al. [4] using a MRI scanner has
shown that the movement of the subject between these scanning stages produces a
lateral displacement/shift in the final 3D model.
The artefacts that result from periodic movements can be minimised by using various
scanning protocols that synchronise the movement with the data acquisition or by
using post processing/filtering techniques [7-10]. The artefacts due to random
movements, on the other hand, cannot be eliminated easily by synchronising or post
processing techniques of the image data. This can be achieved to some extent by
immobilising the subject; however, immobilising a limb for ~60 minutes is not easily
achievable. With regards to 3D model reconstruction, it can be achieved by 3D
modelling techniques in which 3D models from the consecutive scanning stages can
be aligned by using an iterative closest point (ICP) algorithm based technique [11].
The ICP algorithm is a widely used 3D-3D registration technique and has shown a
high accuracy for translational as well as rotational alignments of 3D models. Lee et
al. [1] conducted a preliminary registration test using the ICP algorithm in which a
part of the bone model separated from the original model was matched perfectly to
its original full model. In another study, the ICP algorithm was able to register a CT
derived model to a real patient‟s model with an average error of 0.079 ± 0.068° for
rotation and 0.12±0.09 mm on translations [12]. The ICP algorithm guarantees
convergence to a local minimum from any given transformation of the data point set
[11]. However, the obtained local minimum may not be the desired global minimum,
as it depends on the initial registration. While the ICP algorithm has been used in
numerous studies for aligning bone models, the effect of its initial position on the
optimal global alignment has not yet been reported.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
131
This study aimed to correct the lateral shift artefact that is associated with MRI
scanning of long bones using the ICP algorithm for aligning the models. The
dependency of the ICP algorithm on the initial position of the 3D surfaces to register
them was also investigated.
Methods
MRI scans of five ovine femora (Average age = 7 years and average weight = 49 kg)
obtained by scanning five intact sheep hind limbs were used for the study. A 3T MRI
scanner with the following imaging protocol (Table 1) and the body matrix coil was
used.
Table 1 MRI Protocol used for scanning of ovine femora
Parameter Value
Instrument Siemens Trio tim
Field Strength 3T
In plane resolution 0.47 mm × 0.47 mm
Slice thickness 1 mm
TE 1.83 ms
TR 11 ms
FA 10°
Image sequence 3D VIBE
Number of Averages 1
Scanning of the femora was conducted using the setting described below. With the
exception of the first step, the sample was scanned in two halves (proximal and
distal) in all steps, while maintaining an approximately 7 cm overlap between the
proximal and distal halves (Figure 1).
1. The femur was positioned in the centre of the magnet and a complete scan
was obtained using a single field of view (FOV) (Figure 1a).
2. The femur was positioned in the centre of the magnet and the scanning was
conducted in two halves using two FOVs without moving the table (Figure
1b).
3. The scanning was conducted in two halves using the same FOV used in the
previous step; however, the table was moved such that the centre of the lower
or upper half of the sample moved to the centre of the magnet.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
132
4. The sample was scanned in two halves using the same FOVs as used
previously and moving the table, as described in Stage 3. After the distal half
was scanned, the proximal end of the specimen was shifted laterally to
simulate the lateral shift caused by a volunteer moving their leg. Then, the
proximal half was scanned (Figure 1c).
Figure 1: a - samples scanned with a single scanning segment; b – samples scanned
in two segments without moving the table; c – samples scanned in two segments with
a translated proximal segment (right) caused by the lateral shift of the specimen.
3D models of bones were reconstructed from all MRI data sets using the multi-
threshold segmentation method previously developed by the author [13]. This
method combines a multilevel threshold approach with a method of selecting an
appropriate threshold level for a particular anatomical region of a long bone. Two
threshold levels were used for the two anatomical regions: the distal/proximal region
and the diaphyseal region. Most of the articular regions were segmented manually
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
133
due to the presence of a number of different soft tissue types in the bone–soft tissue
interface at those regions.
The five femora were also scanned with a CT scanner as the reference standard
against which to compare the MRI based models. A Toshiba 4 slice helical CT
scanner was used with kVp = 120, mAs = 50, in plane resolution = 0.35 × 0.35 mm
and slice spacing 0.5 mm. The CT data was segmented using the Canny edge
detection method previously investigated by the author.
The pair of 3D models reconstructed from the scans obtained without moving the
table (Step 1) was used to quantify any displacement that might have occurred from
the data acquisition process. The pair of 3D models reconstructed from the scans
obtained after moving the table (Step 2) was used to quantify any displacement that
might have resulted from movement of the table.
The correction of the lateral shift that had been simulated during the scanning
process was conducted using the ICP algorithm built into Rapidform 2006. The two
3D models reconstructed from scans of two halves of the bone were roughly aligned
using the „Shell Trackball‟ tool in Rapidform 2006. The „Shell Trackball‟ tool allows
translation of the model in any of the x, y and z directions and rotation around x, y
and z axes. Only the distal half model was moved, while the proximal half model
was kept locked in the 3D space of Rapidform 2006. After the rough alignment was
carried out, the fine registration function that is based on the ICP algorithm was used
for the final alignment of the models. For its operation, the ICP algorithm requires an
overlapping region between the corresponding halves of the 3D models to be aligned
(Figure 2 a & b).
After alignment, the geometric deviation between two overlapping regions of the 3D
models of two halves was measured using a point to point comparison method built
into Rapidform 2006. Then the models of the two halves were merged using
functions built into Rapidform 2006 to obtain the complete 3D model of the bone
(Figure 2 a & b). The complete 3D models obtained without the table movement,
with the table movement, and with corrected shift artefact were compared with the
CT based reference model using the point to point comparison method built into
Rapidform 2006 (Figure 2 c & d). Before the comparison, the 3D models were
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
134
aligned to the reference model using the fine registration (ICP based) function of
Rapidform 2006.
Figure 2: a - 3D models of distal and proximal halves before the correction of the
artefact, b – the artefact has been corrected by aligning the two halves, c - the
corrected model (brown) is aligned with the reference model (blue), d - the colour
map showing the differences between the corrected model and the reference model.
The minimum overlap length that was required to accurately align the 3D models of
the distal and proximal halves were determined prior to the correction of the artefact
through the following procedure. A femur was MRI scanned two times using the
same imaging protocol, shifting the sample from one end in the second scan to
simulate a lateral shift. Two 3D models of the distal and proximal halves of the
femur were reconstructed so that there was an approximately 7 cm overlap between
the two models (Figure 3). This pair of models was copied 13 times. The models
were then split so that each pair of models had varying overlap starting from 1cm,
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
135
and increasing in 0.5cm increment. In each pair of models, lateral shift was corrected
using the same procedure used to align models based on the ICP algorithm, as
described above. For each set, the corresponding two halves were then merged and
compared against the reference model that was generated from a CT scan of the
bone.
Figure 3: Overlapping region with reference planes created to divide the models.
According to the results obtained (Figure 4), it can be deduced that the overlap of
more than 4.5 cm produces acceptable alignment. Therefore, in this study, a 4.5 cm
overlap was maintained between proximal and distal halves of all the reconstructed
3D models.
Figure 4: Average deviations obtained for different overlapping regions of the 3D
models.
The dependency of the ICP algorithm on the initial positions of the 3D models for its
alignment was investigated using two MRI based 3D models (proximal and distal
halves) and their reference model. The proximal half of the MRI based 3D models
0.00
0.02
0.04
0.06
0.08
0.10
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
6.5
7.0
Aver
age
dev
iati
on
(m
m)
Overlap (cm)
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
136
was placed in three positions (5mm, 10 mm and 15 mm apart from the distal half of
the models) in each of the x, -x, y and -y directions (Figure 5), generating a total of
twelve positions around the distal half for the MRI based models. The models were
not positioned more than 15 mm apart, as the software‟s ICP based function was not
able to align the models with more than a 15 mm distance between the two models.
Then the function based on the ICP algorithm was used to align the proximal and
distal halves of the MRI models. The two models were then merged and compared
with the reference model for geometric deviations.
Figure 5: Proximal half of the MRI based model (blue) positioned in X and Y axes
around the distal half of the MRI based model (red).
Statistical differences between the average deviations of the models obtained from
various scanning methods and the single scan model were calculated using one way
ANOVA. The level of statistical significance was set to p ≤ 0.05. The validation was
performed using PASW Statistics 18 software package.
Results
When the geometric deviations between the overlapping regions of two halves were
measured, the 3D models obtained without any table movements showed 0.18±0.11
mm average deviation. When it was measured in the models obtained after the table
had moved, an average deviation of 0.49 ± 0.10 mm was obtained (Figures 6 & 7).
After correcting the lateral shift artefact, the average deviation between the two
overlapping regions was 0.05±0.01 mm.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
137
Figure 6: Average deviations between overlapping regions of the models.
Figure 7: The lateral shift between the two 3D models obtained after the table was
moved [A part of the distal model (pink) has been removed to show the
displacement].
After merging the two halves, the obtained complete 3D models were compared to
the reference models. The models obtained from two scans but without any table
movements generated an average deviation of 0.26 ± 0.02 mm (Figure 8). The
models obtained after the table had been moved presented an average deviation of
0.34 ± 0.04 mm (Figure 8), and the models with corrected lateral shift artefact
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
No table movement With table movement
With corrected lateral shift artefact
De
via
tio
n ±
SD
(m
m)
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
138
presented an average deviation of 0.32 ± 0.02 mm when compared to the reference
models. The 3D models reconstructed from the MRI data that was obtained in a
single scan showed an average deviation of 0.25 ± 0.02 mm when compared with the
reference models.
Figure 8: Average deviations between the complete models and the CT based
reference standards (* = statistically significant).
The results obtained for the investigation carried out to determine the dependency of
the ICP algorithm on initial alignment of the models presented similar average
deviations for the 12 positions (Table 2).
Table 2 The accuracy of the ICP algorithm in aligning the 3D surfaces which have
different initial alignments
Axis X -X
Initial deviation (mm) 5 10 15 5 10 15
Maximum (mm) 2.54831 2.55382 2.55255 2.54727 2.55275 2.51663 Average (mm) 0.31716 0.31729 0.31725 0.31713 0.31722 0.31662
SD 0.25116 0.25187 0.25164 0.25099 0.25172 0.24708
Axis Y -Y
Initial deviation (mm) 5 10 15 5 10 15
Maximum (mm) 2.54655 2.54958 2.55511 2.54944 2.54584 2.54726
Average (mm) 0.31708 0.31719 0.31733 0.31715 0.31710 0.31714 SD 0.25094 0.25134 0.25201 0.25135 0.25081 0.25101
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
No table movement
With table movement
With corrected lateral shift
artefact
Single scan MRI
De
via
tio
n ±
SD (
mm
)
* *
*
*
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
139
Discussion
This study investigated a method of correcting the lateral shift artefact that occurred
as a result of the random movements of the subject between two successive scanning
stages. This random movement is considered as one of the motion artefacts that
occur in MRI imaging of long bones, in which the scanning is performed in a number
of stages. The correction of the artefact is important before the models are used in
various applications. In this study, a method of correcting this artefact was proposed
and validated using the robust ICP algorithm to align the overlapping regions of two
models with the simulated lateral shift artefact.
It is known that the accuracy of the final optimal alignment performed by using the
ICP algorithm is dependent on the initial position of the 3D surfaces. The
investigation performed in this study, utilising two halves of a long bone, showed
that the ICP algorithm based aligning method does not depend on the initial
alignment of up to 15 mm for its registration process. The average errors obtained
from this investigation were in the range of 0.31662 – 0.31733 mm with a standard
deviation of 0.00018 mm between twelve measurements performed. Therefore, any
effects on the alignment that might have been caused by the initial position of the
models can be excluded. The minimum overlap required for the alignment of the two
halves of the models can be as low as 4.5 cm, as determined by the investigation
carried out in this study.
The models obtained after the table was moved but without moving the specimen
presented a higher error compared to the error obtained for the 3D models based on a
single MRI scan. A lateral displacement of ~0.5 mm was visible in the 3D models
reconstructed from two halves that were obtained after the table was moved. This
lateral displacement is most likely to be caused by the mechanical instability of the
moving table and/or by the slight movement of the sample resulting from the
momentum contained in it. Generally, the scanning of a human long bone has to be
performed by moving the table to cover the complete length of the bone and thus,
any error generated due to the table movement is inevitable.
The proposed method was able to correct the generated lateral shift artefact with an
average error of 0.32 ± 0.02 mm between the model with corrected shift artefact and
the reference model (CT based model). The error was within sub-voxel levels and
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
140
was slightly higher than the average error obtained for the models (scanned with two
FOVs) reconstructed without moving the table (0.26 ± 0.02 mm) and the model
obtained by using a single scan (0.25 ± 0.02 mm). The small residual error of the
model with the corrected lateral shift artefact, compared to the model obtained with
the single scan, is most likely to be the result of a slight mal-alignment between the
proximal and distal halves. The average deviations between the model with the
corrected shift artefact and the single scan model were significantly different
statistically (p = 0.001); however, the difference between the model with the
corrected shift artefact and the model obtained after moving the table was not
statistically significant. Thus, using the proposed method, the lateral shift artefact can
be corrected to an accuracy that is expected from clinical scanning where the table is
moved. Generally the clinically acceptable tolerances for anatomically fitting
fracture fixation plates are in the order of millimetres [14, 15]. Thus, the accuracy
obtained, after correcting the shift artefact, is within the acceptable range for
designing fracture fixation implants.
The errors obtained for the single scan based MRI models (0.25 ± 0.02 mm) could
have been the result of the manual segmentation that was performed in the articular
regions of the MRI based models, and the larger slice spacing (1 mm) used in MR
imaging compared to the 0.5 mm used in CT imaging. This error is consistent with
the average error obtained for the comparison of MRI based models with the CT
based models (0.23 mm) in a previous study conducted by the authors [2]. The
articular regions are covered with a number of different types of soft tissue that
exhibit different MRI properties. The contrast between those certain soft tissue types
and the bone is generally not high enough for an accurate thresholding of the bone.
Thus these regions were segmented manually, potentially introducing errors to those
regions of the 3D models.
The comparison of the models generated without table movements to the reference
model presented an average error of 0.26 ± 0.02 mm. This error might have occurred
mainly as a result of the segmentation process and the large slice spacing as
mentioned in the previous paragraph [2]. However, the average deviation of 0.18 ±
0.11 mm that was measured between the overlapping regions suggests that there is a
slight lateral deviation between two halves of the models that resulted from the
scanning process. The exact reason for this deviation could not be determined.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
141
The present study utilised only two scanning segments, resulting in one lateral shift
artefact; however, when a human femur is being scanned, at least three segments
have to be used; this results in two lateral shift artefacts. Thus, the error generated
might be higher with a greater number of segments, compared to the present study.
In addition, the correction of the artefact was performed manually and this is a labour
intensive process. Therefore, automatic processing of the correction of artefact is
desirable in future.
The method proposed in this study was able to correct the lateral shift artefact of the
3D models based on MRI with acceptable accuracy for implant design. This was
achieved using the robust ICP algorithm to align the 3D models using an overlapping
region. The study also demonstrated that the ICP algorithm based function used in
this study does not depend on the initial position of up to 15 mm for its alignment
process. This allows medical engineering researchers to reconstruct accurate 3D
models of long bones using MRI with minimum effect from the lateral shift artefact.
Acknowledgement
This research was supported in part by Synthes GmbH. The last author has received
an industrial scholarship from Synthes GmbH. The authors acknowledge the
National Imaging Facility for providing 100% subsidised access to the 3T MRI
scanner.
References
[1]. Lee Y, Seon J, Shin V, Kim G-H, Jeon M. Anatomical evaluation of CT-MRI
combined femoral model. Biomedical Engineering Online 2008;7(1):6.
[2]. Rathnayaka K, Momot KI, Noser H, Volp A, Schuetz M, Sahama T, Schmutz
B. Quantification of the accuracy of MRI generated 3D models of long bones
compared to CT generated 3D models. Medical Engineering & Physics
2011(in press DOI: 10.1016/j.medengphy.2011.07.027).
[3]. Henley G, Harrison JE, Serious injury due to land transport accidents,
Australia 2006-07, in Injury research and statistics series. 2009, Australian
Institute of Health and Welfare: quCanberra. p. 4-6.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
142
[4]. Schmutz B, Volp A, Momot K, Pearcy M, Schuetz M. Using MRI for the
imaging of long bones: First Experience. Journal of Biomechanics
2008;41(Supplement1):S188.
[5]. Brown MA, Semelka RC. MRI Basic principles and applications. 4th ed.
2010, New Jersey: John Wiley & Sons.
[6]. Peh WCG, Chan JHM. Artifacts in musculoskeletal magnetic resonance
imaging: identification and correction. Skeletal Radiology 2001;30(4):179-
191.
[7]. Stadler A, Schima W, Ba-Ssalamah A, Kettenbach J, Eisenhuber E. Artifacts
in body MR imaging: their appearance and how to eliminate them. European
Radiology 2007;17(5):1242-1255.
[8]. Wood ML, Henkelman RM. MR image artifacts from periodic motion.
Medical Physics 1985;12(2):143-151.
[9]. Cîndea N, Odille F, Bosser G, Felblinger J, Vuissoz P-A. Reconstruction
from free-breathing cardiac MRI data using reproducing kernel Hilbert
spaces. Magnetic Resonance in Medicine 2010;63(1):59-67.
[10]. Odille F, Cîndea N, Mandry D, Pasquier C, Vuissoz P-A, Felblinger J.
Generalized MRI reconstruction including elastic physiological motion and
coil sensitivity encoding. Magnetic Resonance in Medicine 2008;59(6):1401-
1411.
[11]. Besl PJ, McKay ND. A Method for Registration of 3-D Shapes. IEEE
Transaction on Pattern Analysis and Machine Intelligence 1992;14(2):239-
256.
[12]. Popescu F, Viceconti M, Grazi E, Cappello A. A new method to compare
planned and achieved position of an orthopaedic implant. Computer Methods
and Programs in Biomedicine 2003;71(2):117-127.
[13]. Rathnayaka K, Sahama T, Schuetz MA, Schmutz B. Effects of CT image
segmentation methods on the accuracy of long bone 3D reconstructions.
Medical Engineering & Physics 2010;33(2):226-233.
[14]. Schmutz B, Wullschleger ME, Noser H, Barry M, Meek J, Schuetz MA. Fit
optimisation of a distal medial tibia plate. Computer Methods in
Biomechanics & Biomedical Engineering 2011;14(4):359-364.
Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone
143
[15]. Schmutz B, Wullschleger ME, Kim H, Noser H, Schuetz MA. Fit Assessment
of Anatomic Plates for the Distal Medial Tibia. Journal of Orthopaedic
Trauma 2008;22(4):258-263.
Chapter 8: Summary, conclusion and future directions
145
Chapter 8 Summary, conclusion and future
directions
8.1 Summary and conclusion
The overall objective of the research was to investigate the use of magnetic
resonance imaging (MRI) to replace the current gold standard–computed tomography
(CT)–so as to acquire long bone geometric data from healthy human volunteers. This
data is required to design pre-contoured fracture fixation implants (plates and nails)
to fit the anatomy of young patient age groups and patients from different ethnic
groups. CT cannot be used for this purpose due to the involvement of high amounts
of ionising radiation. With this overall objective, the study specifically aimed to:
develop a simple and accurate segmentation method for segmentation of MRI and
CT data of long bones; formally validate the geometric accuracy of the MRI and CT
based 3D models of long bones with an appropriate reference standard; use higher
field 3T MRI to improve the poor contrast of certain anatomical regions (which is a
limitation of current 1.5T MRI scanners); and correct the step artefact in the 3D
models caused by the movement of volunteers during the MRI scan.
The reconstruction of 3D models of bones with accurate representation of the surface
geometry requires using an accurate segmentation method. Currently available
sophisticated segmentation methods are capable of segmenting relatively short bones
with minimum user intervention; however, the accessibility of these methods by the
general research community is limited due to the complex mathematics and
programming involved. This study proposed and validated two relatively simple but
accurate segmentation methods: multi-threshold segmentation and Canny edge
detector based segmentation, which can be used to accurately segment the CT and
MRI images of long bones. The former uses the popular intensity thresholding with
Chapter 8: Summary, conclusion and future directions
146
multiple threshold levels for regions of the bone that have different intensity levels.
The threshold levels were calculated using the developed threshold selection method
to minimise user dependent errors of selecting a threshold level. The latter uses the
Canny edge detector which is already built into common image processing platforms
(e. g. Matlab and IDL). Both segmentation methods were capable of segmenting
outer and inner surfaces of ovine femora from CT images with high accuracy when
compared with the reference standards.
MRI, as an ionising radiation free imaging method, has shown potential for scanning
of bones for reconstructing 3D models. This was formally validated for
reconstructing 3D models of long bones with accurate surface geometry, using 1.5T
MRI and CT scans of ovine femora. The state of the art dense triangular meshed
surfaces generated from a contact mechanical scanner were used as the reference
standard. Image segmentation of both CT and MRI data was conducted using the
multi-threshold segmentation method developed in this study.
Results showed that there was no statistically significant difference between the
obtained MRI based 3D models and the CT based models. Compared to the
diaphyseal regions, the articular regions of the MRI based 3D models presented
lower accuracy. This is due to the poor contrast in those regions resulting from a
number of different types of soft tissue with different MRI properties that surround
the bone. Segmentation of MRI images takes longer than segmentation of the CT
images, especially in articular regions; this is also labour intensive compared to CT
images. In addition, MRI‟s very long scanning times make the images vulnerable to
the artefacts caused by random movements of the subject. These factors might limit
the use of MRI for reconstruction of 3D models of long bones.
There are some promising approaches to addressing these current limitations of using
MRI for scanning of long bones. Higher field strength MRI scanners promisingly
offer higher signal to noise ratio (SNR) levels that can be used either to reduce the
scanning time or improve the poor contrast in articular regions. Since the commonly
used higher field strength MRI scanner in the clinical setting is 3T, in the present
study, a comparison between 1.5T and 3T was conducted to quantify the improved
image quality at 3T. The comparison using signal to noise ratio (SNR) of soft tissues
and bone marrow, and contrast to noise ratio (CNR) of bone–muscle and bone–bone
Chapter 8: Summary, conclusion and future directions
147
marrow interfaces resulted in comparatively higher SNR and CNR levels for most of
the regions of the femur and tibia.
The increased contrast at 3T might improve the segmentation accuracy of the
articular regions; however, according to the author‟s experience of segmentation, this
only marginally reduces the segmentation time in comparison to the 1.5T images.
Whilst SNR and CNR are increased at 3T, some of the artefacts may also be
exaggerated. The magnetic susceptibility becomes more apparent at 3T and the
chemical shift artefact is doubled due to the increased difference of the resonance
frequency between water and fat molecules. Since the strength of the magnets is
being increased over time, scanners with higher magnetic field (e.g. 7T) than 3T will
potentially increase the image quality. However, increased SAR levels at higher
magnetic fields will potentially limit their use for human imaging.
The investigation showed that the longitudinal relaxation time (T1) of the muscle was
highly field dependent, while T1 of bone marrow was weakly field dependent. In
both muscle and bone marrow, T1 increased at 3T. In contrast, the transverse
relaxation time (T2) of muscle was not field dependent; however, the literature
reports that T2 takes slightly lower values at higher magnetic field strengths.
Increased T1 at 3T requires relatively higher TR values to be used to get the
maximum intensity levels and this, in turn, increases the scanning time. In general,
this can be compensated for by using fewer averages when data is acquired;
however, in the present study, it is not possible as the number of averages used is
one.
In MRI imaging of long bones of lower limbs, the artefacts due to the random
movements of the subject are relatively more prominent and important than those
due to the periodic movements such as respiration and blood flow. As observed by
the supervisory team, the random movements between two successive scanning
stages causes a step in the 3D models reconstructed from such data sets of lower
limbs. The artefacts due to the random movements cannot be eliminated by post
processing or by synchronising the data acquisition. Immobilisation of the limb for
about 60 minutes is also not achievable unless an invasive anaesthetic method is
used. However, as the lateral shift artefact appears in the reconstructed 3D models,
the robust ICP algorithm that has been widely used for 3D-3D registration of
Chapter 8: Summary, conclusion and future directions
148
surfaces was implemented for correcting this artefact when simulated in ovine
femora. The resulting 3D models had sub voxel-level accuracy (voxel size = 0.35
mm2) when the surface geometry was compared to the reference standard. The study
also indicated that the movement of the table makes a displacement in the data sets;
however, this may not be important in clinical applications. Nevertheless, to
minimise the geometric errors, the data acquisition of long bones for 3D
reconstruction should consider this displacement caused by the table movement.
In conclusion, magnetic resonance imaging, together with simple multi-level
thresholding segmentation, is able to produce 3D models of long bones with accurate
geometric representations. It is, thereby, a potential alternative scanning method
where the current gold standard CT imaging cannot be used. However, there are a
number of limitations such as long scanning times, long segmentation time, and
movement artefacts that have to be resolved before employing MRI for this purpose.
8.2 Future directions
This study successfully validated the accuracy of MRI to reconstruct 3D models
from long bones using simple but accurate segmentation methods. The usability of
3T MRI scanners was also investigated, while 3D modelling techniques were used to
correct the shift artefacts. However, there are a number of limitations or challenges
that should be addressed in the future, before using MRI as an alternative to CT for
imaging of long bones for 3D reconstruction.
The segmentation methods described in this research may also be used in fields other
than 3D reconstruction of long bones. Cardiac MRI and CT image segmentation is
one such area where accurate segmentation is required for volumetric measurements.
MR only radiotherapy planning is another aspect in the clinic that requires accurate
segmentation of bone and soft tissue and these methods can potentially be employed
for these purposes.
With regards to image segmentation, segmentation of MRI images takes a
considerably longer time compared to the segmentation of CT images. Even though
3T scanners are able to improve the contrast levels in articular regions, according to
the author‟s experience, this only marginally reduces the segmentation time. Future
studies might focus on automating the segmentation process to reduce the
Chapter 8: Summary, conclusion and future directions
149
segmentation time. In addition, the use of magnetic fields stronger than 3T may also
improve the contrast levels, thus allowing faster segmentation, especially of the
articular regions.
In addition to the use of higher field scanners to improve the image quality, this may
also be achieved with specially designed RF coils or imaging protocols. In the
present study, the peripheral angiography (PA matrix) coil was used for imaging of
the lower limbs; however; there are no RF coils currently available for scanning of
upper limbs. Therefore, designing RF coils especially for scanning of long bones of
the upper and lower limbs in a future study will improve the quality of MRI images
of bones and, hence, the segmentation accuracy and time. Using imaging protocols
such as the protocols designed for fat and water only imaging or protocols with ultra
short TE (UTE) will potentially improve the CNR between bone and soft tissues.
Furthermore, currently available imaging sequences are also mainly designed to scan
soft tissues. Collaboration with manufacturers to design protocols for scanning of
bones could also have an influence on improving image contrast of MRI of bone.
The present study validated the correction of step artefacts in MR imaging of long
bones using ovine femora which is relatively smaller than human femora. Therefore,
this method has still to be validated using human long bones before using it to
successfully generate 3D models of human long bones. Future studies can be
conducted using fresh human cadaver bones and CT as the reference standard.
According to the studies conducted using 1.5T and 3T MRI scanners, MRI of human
long bones results in very long scanning times. This can potentially be shortened in
the future by using higher magnetic fields (e.g. 7T). In addition, optimising the
scanning protocol for different regions of the bone (e.g. use of larger slice spacing
and low resolutions for diaphyseal region where the geometry is relatively simple)
may also reduce the scanning time.
Even though artefacts due to periodic movements are not prominent in the MRI
scanning of long bones of lower limbs, scanning of upper limbs are affected by
respiratory movements. Therefore, minimising periodic motion artefacts is also
important in the long bone MRI of upper limbs. In this study, the lateral shift artefact
was corrected using 3D modelling techniques by manually positioning the bone
models. However, in future, automatic processing is desirable in order to reduce the
Chapter 8: Summary, conclusion and future directions
150
time taken for manual processing. In addition to motion artefacts, minimising the
artefacts produced by magnetic susceptibility and the chemical shift may also be
important in MRI scanning of bones, especially when high magnetic fields are used
as these exaggerate the artefacts. Minimising or elimination of these artefacts is
important for improving the accuracy of the implants designed using the MRI based
3D models.
Appendix 1
151
Appendix 1 Ethical approval for the study in Chapter 6
••
-,
~
Royttl Brishant: o.mU \Voml!n·s Hospital Mctm Nonh Health Service District
Queensland Government
Office of the Human Research Ethics Committees Queensland Health
Or Kanchana Rathnayaka Queensland University ofTechnology Institute of Health & Biomedical Innovation 60 Musk A venue Kelvin Grove Q 4059
Dear Dr Rathnayaka,
Em1uioics 10: Odcnc Pctco,;cn Coordinator
Phone: 07 3636 5490 Fax: 07 3636 5849 Our Rcf: HREC/1 0/QRBW/1 41 E-mail IW\1'11-I'oluc''" healt h qld go,·.au
Re: Ref N!l: HREC/10/QRBW/141: Comparative study of 3T MRI vs I.ST for the acquisition of 3D morphological bone data of the lower extremity
Thank you for submitting the above project for ethical and scientific review. This project was considered at the Royal Brisbane & Women's Hospital Human Research Ethics Committee (HREC) meeting held on I 9 April, 2010.
I am pleased to advise that the Human Research Ethics Committee has granted approval of this research project on 13 May, 2010. HREC approval is valid for three (3) years from the date of this letter.
This HREC is constituted and operates in accordance with the National Health and Medical Research Council 's (NHMRC) National Statement 011 Ethical Co11duct in Human Research (2007). NHMRC and Universities Australia Australian Code.for tlze Responsible Conduct of Research (2007) and the CPMPIICH Note.for Guida11ce on Good Clinical Practice. Attached is the HREC Composition with specialty and affiliation with the Hospital (A twchmc!Tt lj.
You are reminded that this letter constitutes ethical approval only. You must not commence this research proJect at a site until separate authorisation from the District CEO or Delegate oftlwt site ftas been obtained.
A copy t?fthis approval will also be sent to the Di.vtrict Research Governance Office (RGO). Please ensure you submit a completed Site Specific A!!·sessmeut (SSA) Form to the RGO for authorisation from tfze CEO or Delegate to conduct this research at the Royal Brisbane & Women's Ho.\pital ML'fm North District.
The documents reviewed and approved include:
!!to' /lo_•·t~f llri.</1tmc & Women·_. /ln;pital /lunwn Rcst•llrcil Ethic.< Commillct• is cvmlilulcd ami operate.< according lo tile NNMRC's Nmioual Swu·mcnl ou l:'thical Couducl ill /fumau Rc.<carch (2 1!117).
Office
Bullcrlidd St reet l·krstun Q ~02'1
Postal
Pnst Ollicc Hcrstun Queensland 4029 Australia
Phone Fnx
07 3636 541)0 07 3636 5849 lSD + 61 7 3636 541JO
Appendix 1
152
ii•J,ro/ /lri.rholl<' & Woml'll '.< llmpilol /111/o'C' 1/cf N": 1//II:"Ci/ INJIIIJII'I / .f I
Document
Covering Letter
Application: NEAF
Protocol: MR protocol for scanning tibia & femur of a volunteer
Curriculum Vitae of Kanchana Ratl~nay:~~a
Version .,l
I 2.0 (2008)
/3.115.111/11
Date -·-_I
29 March 20 I 0 I i ................. .. I I I March 2010 I
-' --- -~ --- _.. -· ___ _j . --··· -· ···- _,_, _________ - _j
Curriculum Vitae of Beat Schmutz . i __ --· --·. _! ------·-------·---: ' Letter of Support from Professor Alan Coulthard, Dept of I '1 23 March 2010 !
M d. 11 . I I e 1ca magmg 1 ... J. . .......... ___ __I
Research Funding Schedule (reviewed in accordance with I l _j Section 3.3. /§_o[!lze _Nc:J..~ona~ Sta_f_!!!J.c:!!JL .. --·---.. _ ______ ]_______ ·--
Emai I coJTespo_n~en<:e ~~ Q!:!I l_n~~m~~ie~. ~EJ!:l~.':l-~nc~---·_j ___ __] 1 0 Ma~J_O __ j Response to R_~q!:l~tf'?!. _F!.!l'!!~e~In_fo~nation ____________ .. ___ __j ______ j _j_!.Jv.1~Y- 20lQ__ j
Participant lnfOTJTI~~ion Sl~e~-~-~O':JS_el_l! ..f.<?!.!'!!_ ....... _____ j __ 2 ____ _ __1 _I_!_~~ :X: 20 1_0 __ j
Please note the following conditions of approval:
I. The P1incipal Investigator will immediately rep011 anything which might wrunnt review of ethical approval of the project in the specified fon'nat, including:
• Unforeseen events that might affect continued ethical acceptability of the project. Serious Adverse Events must be notified to the Committee as soon as possible. In addition, the Investigator must provide a summary of the adverse events, in the specified fmmat, including a comment as to suspected causality and whetl1er changes are required to the Patient lnfonnation and Consent Form. In the case of SeJious Adverse Events occuJTing at the local site, a full report is required from the Principal Investigator, including duration of treatment and outcome of event.
2. Amendments which do not affect either the ethical acceptability or site acceptability of the project (e.g. typographical eJTors) should be submitted in hard copy to the HREC Coordinator. These should include a covering letter from the Principal Investigator providing a brief descliption of the changes and the rationale for the changes, and accompanied by all relevant updated documents with tracked changes.
3. Proposed amendments to the research project which may affect both the ethical acceptability and site suitability of the project must be submitted firstly to the HREC for review and, once HREC approval has been granted, then submitted to the Research Governance Office.
4. Amendments to the research project which only affect the ongoing site acceptability of the project are not required to be submitted to the HREC for review. These amendment requests should be submitted directly to the Research Governance Office
Appendix 1
153
lio!l a/ 1/ri ,·hall<' & ll'omwn 's 1/ospillll /III/X ' llr:{-No: 1/1/I:'C/ /11/Q/IIl ll'/1·11
(by-passing the HREC).
3 IJ.05.211/IJ
5. Amendments to the research project which may affect the ongoing ethical acceptability of a project must be submitted to the HREC for review. Major amendments should be reflected in a revised online NEAF (accompanied by all relevant updated documentation and a coveting letter fi·om the Principal Investigator, providing a brief desctiption of the changes, the rationale for the changes, and their implications for the ongoing conduct of the study). Hard copies of the revised NEAF, the cover letter and all relevant updated documents with tracked changes must also be submitted to the HREC Coordinator as per standard HREC SOP. Further advice on submitting amendments is available fi·om hllp: . \\' lV II' .hc:1l th .qld. !!.I >v .au/uhnl r/dlH.:umcnls/n:scarchcr uscr!!.uidc.pd f
6. The I-IREC will be notified, giving reasons, if the project is discontinued at a site before the expected date of completion.
7. The HREC will be notified, giving reasons, on any sponsor repmis or other infonnation which might affect the ongoing ethical acceptability in line with the requirements of the ICH GCP guidelines as annotated by the TGA: hll p :ti ii'WW .ll!a.l!OV .au/doc.:s/pd li'CUl!Uidc/ich/ich [J 51J5.pd f
8. The Ptincipal Investigator will provide an Annual Repoti to the HREC and at completion of the study in the specified fonnat.
9. The District Administration and the Human Research Ethics Committee may inquire into the conduct of any research or purpotied resem'ch, whether approved or not and regardless of the source of funding, being conducted on Hospital premises or claiming any association with the Hospital, or which the Committee has approved if conducted outside Royal Brisbane & Women's Hospital Metro Nmih Health Service District.
Should you have any queries about the HREC's consideration of your project please contact I lie Ill\ I:C ('pnrtfinalor on 07 363(> 5.:JlJO. The HREC terms of Reference, Standard Operating Procedures, membership and standard forms are available from li lip :i w 11 11· . l1 L'a it h. ql d .l!O v .J u.'nl11 nr/h t 1111 /n:l!ui rcgu hnmc.asp
Once authorisation to conduct the research has been granted, please complete the Commencement Fom1 ( : I twcluncnt 1/J and retum to the office ofthe Human Research Ethics Committee.
The HREC wishes you every success in your research.
Yours sincerely, -1 . --,J
' / ,, -'~r . --, -- ')
. ~
Or Conor Brophy Chair·pcrson RBWH Human Research Ethics Committee ,\ ·ki r•' N, >rl h District 13.05.2010
Appendix 2
154
Appendix 2 Participant information and Consent form
used in Chapter 6
~ ~n~u~fi Health and Biomedical Innovation
Queensland Government Queensland Health
PARTI:PANT lNFORMATDN for RESEARCH PROJECT
"Comparative study of 3T M RI vs l.ST for the acquisition of 3D mor holo ical bone data of the lower extremi "
Research Team Contacts Dr Kanchana Rathnayaka, Principal Investigator
Institute of Health and Biomedical Innovation F acuity of Built Environment and Engineering
(07) 3138 6234 k.rathnava ka@ aut.edu .au
Description
Dr. B. Schmutz, Senior Research Fellow Institute of Health and Biomedical Innovation Faculty of Built Environment and Engineering
(07) 3138 6238 b.schmutz@ aut.edu.au
This research is undertaken as a part of the PhD project of Kanchana Rathnayaka, Trauma Research Group (www.ihbi.gut.edu.au/qo/mdtrauma) at the Queensland University of Technology (QUT), with the image data being acquired at the Royal Brisbane and Women's Hospital.
The purpose of this project is to assess and quantify the improvement in image quality obtained with a high field 3 Tesla (3T) Magnetic Resonance Imaging (M RI) scanner compared to images acquired with a 1.5 Tesla (l.ST) conventional scanner. Some of the current limitations of MRI are long scanning times and low image contrast for certain anatomical regions. Higher field strength (3T and above) scanners offer improved signal which may be translated to faster imaging times or better image quality.
The research team requests your assistance because you have previously participated in a similar project (Human bone morphology database (MRI of volunteers) where l.ST M RI images of your leg were acquired. The images of your leg, acquired in the previous study, can now be compared with those acquired with a stronger 3T magnet in this study.
Participation Your participation in this project is voluntary. If you do agree to participate, you can withdraw from participation at any time before your data has been acquired without comment or penalty. Your decision to participate will in no way impact upon your current or future relationship with QUT, with the Royal Brisbane and Women's Hospital or with Queensland Health.
Your participation will involve the acquisition of anonymous (non-identifiable) Magnetic Resonance Image (MRI) data of your legs. You will also be asked to provide details of your age, gender, height and weight which will be stored with your image data.
It is expected that it will take approximately 45- 60 minutes to acquire the image data of your leg, during that time you will be asked not to move your leg which is being imaged. To reduce the level of noise generated during the operation of the M RI scanner you will be given earplugs or headphones to wear (you may bring along your choice of relaxing music on a CD). Your total stay in the Radiology Department could last up to 2 hours.
The imaging session will be conducted at the Royal Brisbane and Women's Hospital, Butterfield Street, Herston, QLD 4029, Australia. The data will be securely stored and analysed at QUT.
You will receive Coles-Myer's Gift Vouchers equal to an amount of $50.- as reimbursement for your travel expenses and for your time spent in association with this project.
Expected benefits It is expected that this project will not benefit you. The outcome if the project will facilitate the generation of 3D models from 3T M RI images using less segmentation time compared to the l.ST scanner images.
Volunteer info consentfonn Version 3. 8 June 2010 1 of 3
Appendix 2
155
• ~n~uefi Health and Biomedical Innovation
Risks
Queensland Government Queensland Health
There are no risks beyond normal day-to-day living associated with your participation in this project. Because M RI uses low-energy, non-ionising radio waves, there are no known risks or side effects except for the groups specified below.
While there are no known hazards, MRI is not proven to be safe during the first three months of pregnancy. Therefore if you are pregnant (or if you suspect that you are pregnant) we ask you not to participate in this study.
The magnet at the centre of the scanner may affect, or be affected by, any person fitted with a pacemaker, hearing aid, other electrical device, or metal implants. If you are fitted with any such devices or implants we ask you not to participate in this study.
You wil l be asked not to move your leg which is being imaged for approximately 45 - 60 minutes during the scanning. If you have any medical conditions which are affected by this, we ask you not to participate in this study.
If you have sustained a penetrating eye injury in the past, there might be harmful effects from M RI scanning and we ask you not to participate in this study.
For a part of the imaging session your whole body will be inside the confined space of the scanner's tunnel. Therefore, if you are claustrophobic we ask you not to participate in this study.
Confidentiality All comments and responses are anonymous and will be treated confidentially. The names of individual persons are not required in any of the responses. Your name will not be recorded or stored with your image data.
Your data will be stored on a secure server at QUT. Access to the database is not available to the public. Upon request, the Trauma Research Group atQUT reserves the right to share your data with third parties to be used in projects aiming to benefit human kind.
Thus, the data acquired from your leg might be used for projects which involve research, development, education and teaching.
Consent to Participate We would like to ask you to sign a written consent form (enclosed) to confirm your agreement to participate.
Questions /further information about the project Please contact the research team members named above to have any questions answered or if you require further information about the project.
Concerns /complaints regarding the conduct of the project QUT and the Royal Brisbane and Women's Hospital are committed to researcher integrity and the ethical conduct of research projects. However, if you do have any concerns or complaints about the ethical conduct of the project you may contact the Human Research Ethics Committee (HREC) Coordinator at the Royal Brisbane and Women's Hospital on (07) 3636 5490 and/or Research Ethics Coordinator at Queensland University of Technology on (07) 31382091. The HREC Coordinator is not connected with the research project and can facilitate a resolution to your concern in an impartial manner.
Volunteer info consent form Version 3, 8 June 2010 2 of 3
Appendix 2
156
• ~n~u~fi Health and Biomedical lnnoval ion
Queensland Government Queensland Health
CONSENT FORM for RESEARCH PROJECT
"Comparative study of 3T M RI vs l.ST for the acquisition of 30 mor holo ical bone data of the lower extremi "
Statement of consent
By signing below, you are indicating that you:
• have read and understood the information document regarding this project
• have had any questions answered to your satisfaction
• have been given the opportunity to have a friend or relative present when the study was explained
• understand that if you have any additional questions you can contact the research team
• understand that you are free to withdraw at any time before your data has been entered into the database, without comment or pena lty
• understand that you can contact the HREC Coordinator at the Royal Brisbane and Women's Hospital on (07) 3636 5490 if you have concerns about the ethic a I conduct of the project
agree to participate in the project which involves the acquisition and storage of anonymous (nonidentifiable) magnetic resonance image data of your leg
• agree that the data which has been acquired of your leg can be used for various projects in research, development, education and teaching
• agree that images of your bones might be used for publications, teaching and educationa l purposes
• are not pregnant to the best of your knowledge
• have not sustained any penetrating eye injuries in the past
Participant Name ____________________________ _
Signature ----·---------·--·--·----------------------·---------------------------·--·-------·--· -·-------·----·------·---------------
Date __ _ I I
I have explained the nature and purpose of this study to the above participant and have answered their questions.
~vestigarorName ___________________________ _
Signature __ _
Date I I
Volunteer info consent form Version 3, 8 June 2010 3 of 3
Appendix 3
157
Appendix 3 Animal tissue use notification
Dear Dr Kanchana Rathnayaka Mudiyanselage,
Re: TISSUE USE NOTIFICATION: Use of ovine limbs for a study entitled
"Correction of the step artefact associated with MRI of long bones" (Source
studies: 08-0848, Goss)
Thank you for your notification of animal tissue use which has been noted and
confirmed as falling outside the scope of requiring review by the UAEC.
Your confirmed application is attached and your approval number is:
1000000529. Please quote this number in all future correspondence.
SPECIFIC CONDITIONS OF APPROVAL:
a. the provision of specimen material is approved by the CI of the source study (i.e.
Dr Ben Goss) b. there are no changes to the source study's protocol to facilitate
provision of the specimens:
- the animal is only killed by the person authorised to do so
- the researcher has no input into the treatment and handling of the
animal prior to euthanasia
- the researcher has no input into the timing or manner of euthanasia
- the tissue/whole animal is collected by the researcher after death is
confirmed.
Note: Tissue use notifications are made available to the UAEC for noting at the next
available meeting. You will only be contacted again if the UAEC raise any questions
at that time.
Please do not hesitate to contact the Research Ethics Unit if you have any queries.
Best regards
Research Ethics Unit | Office of Research | Level 4 | 88 Musk Ave | Kelvin Grove
p: +61 7 3138 2340 | f: +61 7 3138 4543 | e: [email protected]
w: http://www.research.qut.edu.au/ethics/
References
159
References
1. Lee, Y., et al., Anatomical evaluation of CT-MRI combined femoral model.
Biomedical Engineering Online, 2008. 7(1): p. 6.
2. Gelaude, F., J.V. Sloten, and B. Lauwers, Semi-automated segmentation and
visualization of outer bone cortex from medical images. Computer Methods
in Biomechanics & Biomedical Engineering, 2006. 9(1): p. 65-77.
3. Gelaude, F., J.V. Sloten, and B. Lauwers, Accuracy assessment of CT-based
outer surface femur meshes. Computer Aided Surgery, 2008. 13(4): p. 188-
199.
4. Gassman, E.E., et al., Automated bony region identification using artificial
neural networks: reliability and validation measurements. Skeletal
Radiology, 2008. 37: p. 313-319.
5. Pilgram, R., et al., Knowledge-based femur detection in conventional
radiographs of the pelvis. Computers in Biology and Medicine, 2008. 38(5):
p. 535-44.
6. Revol-Muller, C., et al., Automated 3D region growing algorithm based on
an assessment function. Pattern Recognition Letters, 2002. 23(1-3): p. 137-
150.
7. Van Luong, H. and J. Kim, A New Parallel Approach to Fuzzy Clustering for
Medical Image Segmentation, in Advances in Visual Computing. 2008. p.
1092-1101.
8. Noble, P.C., et al., The effect of aging on the shape of the proximal femur.
Clin Orthop Relat Res, 1995(316): p. 31-44.
9. Henley, G. and J.E. Harrison, Serious injury due to land transport accidents,
Australia 2006-07, in Injury research and statistics series. 2009, Australian
Institute of Health and Welfare: quCanberra. p. 4-6.
References
160
10. Cao, H., et al., Quantitative 31P NMR spectroscopy and 1H MRI
measurements of bone mineral and matrix density differentiate metabolic
bone diseases in rat models. Bone, 2010. 46(6): p. 1582-1590.
11. Gold, G.E., et al., MR Spectroscopic imaging of collagen: Tendons and knee
menisci. Magnetic Resonance in Medicine, 1995. 34(5): p. 647-654.
12. Reichert, I.L.H., et al., Magnetic resonance imaging of cortical bone with
ultrashort TE pulse sequences. Magnetic Resonance Imaging, 2005. 23(5): p.
611-618.
13. Moore, S., et al., 3D models of blood flow in the cerebral vasculature. Journal
of Biomechanics, 2006. 39(8): p. 1454-1463.
14. Lösch, A., et al., A non-invasive technique for 3-dimensional assessment of
articular cartilage thickness based on MRI part 1: Development of a
computational method. Magnetic Resonance Imaging, 1997. 15(7): p. 795-
804.
15. Haubner, M., et al., A non-invasive technique for 3-dimensional assessment of
articular cartilage thickness based on MRI part 2: Validation using CT
arthrography. Magnetic Resonance Imaging, 1997. 15(7): p. 805-813.
16. Felix Eckstein, D.B.T.M.L., Quantitative MRI of cartilage and bone:
degenerative changes in osteoarthritis. NMR in Biomedicine, 2006. 19(7): p.
822-854.
17. Augat, P. and F. Eckstein, Quantitative Imaging of Musculoskeletal Tissue.
Annual Review of Biomedical Engineering, 2008. 10(1): p. 369-390.
18. Schmidt, G.P., et al., Whole-body MR imaging of bone marrow. European
Journal Of Radiology, 2005. 55(1): p. 33-40.
19. Hoad, C.L., et al. Development of a 3D MRI sequence for computer assisted
surgery of the lumbar spine. in ISMRM. 8th Scientific Meeting and Exhibition
(Denver, Colorado). 2000: Proc. Intl. Soc. Mag. Reson. Med.
References
161
20. Martel, A.L., et al., Assessment of 3-dimensional magnetic resonance
imaging fast low angle shot images for computer assisted spinal surgery.
Computer Aided Surgery, 1998. 3(1): p. 40-44.
21. Fassbind, M., et al. Quantifying foot bone motion via magnetic resonance
imaging. in 52nd Annual Meeting of the Orthopaedic Research Society. 2006.
Chicago, IL: The orthopaedic research society.
22. Edelstein, W.A., et al., The intrinsic signal-to-noise ratio in NMR imaging.
Magnetic Resonance in Medicine, 1986. 3(4): p. 604-618.
23. Ramnath, R.R., 3T MR imaging of the musculoskeletal system (Part I):
considerations, coils, and challenges. Magnetic Resonance Imaging Clinics
of North America, 2006. 14(1): p. 27-40.
24. Hoad, C.L., et al., A 3D MRI sequence for computer assisted surgery of the
lumbar spine. Physics in Medicine and Biology, 2001. 46: p. N213-N220.
25. Wolf, P., et al., Reliability of tarsal bone segmentation and its contribution to
MR kinematic analysis methods. Computerized Medical Imaging and
Graphics, 2007. 31(7): p. 523-530.
26. Schmutz, B., et al., Using MRI for the imaging of long bones: First
Experience. Journal of biomechanics, 2008. 41(Supplement1): p. S188.
27. Cogbill, T.H. and K.J. Zeigelbein, Computed Tomography, Magnetic
Resonance, and Ultrasound Imaging: Basic Principles, Glossary of terms,
and Patient Safety. Surgical Clinics of North America, 2011. 91(1): p. 1-14.
28. Novelline, R.A., et al., Helical CT in Emergency Radiology1. Radiology,
1999. 213(2): p. 321-339.
29. Blackmore, C.C., et al., Cervical Spine Screening with CT in Trauma
Patients: A Cost-effectiveness Analysis1. Radiology, 1999. 212(1): p. 117-
125.
30. HALL, E.J. and D.J. BRENNER, Cancer risks from diagnostic radiology.
British Journal of Radiology, 2008. 81(965): p. 362-378.
References
162
31. Smith-Bindman, R., et al., Radiation Dose Associated With Common
Computed Tomography Examinations and the Associated Lifetime
Attributable Risk of Cancer. Archives of Internal Medicine, 2009. 169(22): p.
2078-2086.
32. Berrington de Gonzalez, A., et al., Projected Cancer Risks From Computed
Tomographic Scans Performed in the United States in 2007. Archives of
Internal Medicine, 2009. 169(22): p. 2071-2077.
33. Griffey, R.T. and A. Sodickson, Cumulative Radiation Exposure and Cancer
Risk Estimates in Emergency Department Patients Undergoing Repeat or
Multiple CT. American Journal of Roentgenology, 2009. 192(4): p. 887-892.
34. Raeburn, S., Radiation doses in computed tomography: The increasing doses
of radiation need to be controlled. BMJ, 2000. 320: p. 593-594.
35. Wall, B. and D. Hart, Revised radiation doses for typical X-ray examinations.
Report on a recent review of doses to patients from medical X-ray
examinations in the UK by NRPB. National Radiological Protection Board.
British Journal of Radiology, 1997. 70(833): p. 437-439.
36. Radiologists, T.R.C.o., Making the best use of department of clinical
radiology: guidelines for doctors. 5th ed. 2003, London: Royal College of
Radiologists.
37. Semelka, R.C., et al., Imaging strategies to reduce the risk of radiation in CT
studies, including selective substitution with MRI. Journal of Magnetic
Resonance Imaging, 2007. 25(5): p. 900-909.
38. Zannoni, C., A. Cappello, and M. Viceconti, Optimal CT scanning plan for
long-bone 3-D reconstruction. Medical Imaging, IEEE Transactions on,
1998. 17(4): p. 663-666.
39. Sint Jan, S.V., et al., Low-dose computed tomography: A solution for in vivo
medical imaging and accurate patient-specific 3D bone modeling? Clinical
Biomechanics, 2006. 21(9): p. 992-998.
References
163
40. Brown, M.A. and R.C. Semelka, MRI Basic principles and applications. 4th
ed. 2010, New Jersey: John Wiley & Sons.
41. Lipton, M.L., Totally Accessible MRI. 2008, New York: Springer.
42. Kangarlu, A. and P.-M.L. Robitaille, Biological effects and health
implications in magnetic resonance imaging. Concepts in Magnetic
Resonance, 2000. 12(5): p. 321-359.
43. Merkle, E.M. and B.M. Dale, Abdominal MRI at 3.0 T: The Basics Revisited.
American Journal of Roentgenology, 2006. 186(6): p. 1524-1532.
44. Erasmus, L., et al., A short overview of MRI artefacts. South African Journal
of Radiology, 2004. 8(2): p. 13-17.
45. Schmutz, B., et al., Fit Assessment of Anatomic Plates for the Distal Medial
Tibia. Journal of Orthopaedic Trauma, 2008. 22(4): p. 258-263.
46. Goyal, K.S., et al., Analysis of anatomic periarticular tibial plate fit on
normal adults. Clinical Orthopaedics and Related Research, 2007. 461: p.
245-257.
47. Schmutz, B., et al., Fit optimisation of a distal medial tibia plate. Computer
Methods in Biomechanics & Biomedical Engineering, 2011. 14(4): p. 359-
364.
48. Park, A., J. DiStefano, and T. Nguyen, Congruency of Scapula Locking
Plates Implications for Implant Design, in 56th Annual Meeting of the
Orthopaedic Research Society 2010: New Orleans, LA. p. Poster No. 1746.
49. Schmutz, B., et al., Quantitative fit assessment of tibial nail designs using 3D
computer modelling. Injury, 2010. 41(2): p. 216-219.
50. Wehner, T., Simulation der Frakturheilung beim Fixateur Externe und bei
intramedullarem Nagel., in Medizinische Fakultat. 2008, Universitat Ulm.
References
164
51. Niu, Q., et al., Image processing, geometric modeling and data management
for development of virtual bone surgery system. Computer Aided Surgery,
2008. 13(1): p. 30-40.
52. Handels, H., et al., Virtual planning of hip operations and individual
adaption of endoprostheses in orthopaedic surgery. International Journal of
Medical Informatics, 2000. 58-59: p. 21-28.
53. Zachow, S., et al., Draw and cut: intuitive 3D osteotomy planning on
polygonal bone models. International Congress Series, 2003. 1256: p. 362-
369.
54. Xia, J., et al., Computer-assisted three-dimensional surgical planning and
simulation: 3D color facial model generation. International Journal of Oral
and Maxillofacial Surgery, 2000. 29(1): p. 2-10.
55. Gomez-Benito, M.J., et al., A 3D Computational Simulation of Fracture
Callus Formation: Influence of the Stiffness of the External Fixator. Journal
of Biomechanical Engineering, 2006. 128(3): p. 290-299.
56. Lacroix, D. and P.J. Prendergast, Three-dimensional simulation of fracture
repair in the human tibia. Computer Methods in Biomechanics and
Biomedical Engineering, 2002. 5(5): p. 369-76.
57. Anwar, A., A. Mekonnen, and P. Diamantopoulos, FEA modelling of
anatomical structures. Journal of Biomechanics, 2006. 39(Supplement 1): p.
S423-S423.
58. Trabelsi, N., Z. Yosibash, and C. Milgrom, Validation of subject-specific
automated p-FE analysis of the proximal femur. Journal of Biomechanics,
2009. 42(3): p. 234-241.
59. Keyak, J.H. and Y. Falkinstein, Comparison of in situ and in vitro CT scan-
based finite element model predictions of proximal femoral fracture load.
Medical Engineering & Physics, 2003. 25(9): p. 781-787.
References
165
60. Pérez del Palomar, A. and M. Doblaré, Dynamic 3D FE modelling of the
human temporomandibular joint during whiplash. Medical Engineering &
Physics, 2008. 30(6): p. 700-709.
61. MacNeil, J.A. and S.K. Boyd, Bone strength at the distal radius can be
estimated from high-resolution peripheral quantitative computed tomography
and the finite element method. Bone, 2008. 42(6): p. 1203-1213.
62. Viceconti, M., et al., CT data sets surface extraction for biomechanical
modeling of long bones. Computer Methods and Programs in Biomedicine,
1999. 59(3): p. 159-166.
63. Silva, M.J., et al., Direct and computed tomography thickness measurements
of the human, lumbar vertebral shell and endplate. Bone, 1994. 15(4): p. 409-
414.
64. Dougherty, G. and D. Newman, Measurement of thickness and density of thin
structures by computed tomography: A simulation study. Medical Physics,
1999. 26(7): p. 1341-1348.
65. Maloul, A., J. Fialkov, and C. Whyne, The Impact of Voxel Size-Based
Inaccuracies on the Mechanical Behavior of Thin Bone Structures. Annals of
Biomedical Engineering, 2011. 39(3): p. 1092-1100.
66. Prevrhal S, et al., Accuracy of CT-based thickness measurement of thin
structures: modeling of limited spatial resolution in all three dimensions.
Medical Physics, 2003. 30(1): p. 1-8.
67. Souza, A., J.K. Udupa, and P.K. Saha, Volume rendering in the presence of
partial volume effects. Medical Imaging, IEEE Transactions on, 2005. 24(2):
p. 223-235.
68. Heuscher, D.J. and M. Vembar, Reduced partial volume artifacts using spiral
computed tomography and an integrating interpolator. Medical Physics,
1999. 26(2): p. 276-286.
References
166
69. Schmutz, B., M.E. Wullschleger, and M.A. Schuetz. The effect of CT slice
spacing on the geometry of 3D models. in Proceedings 6th Australasian
Biomechanics Conference. 2007. Auckland, New Zealand: The University of
Auckland.
70. Drapikowski, P., Surface modeling-Uncertainty estimation and visualization.
Computerized Medical Imaging and Graphics, 2008. 32(2): p. 134-139.
71. Kurmis, A.P., J.P. Slavotinek, and K.J. Reynolds, The influence of slice
thickness on the volume measurement accuracy of 3-D MR reconstructions of
acrylic phantoms: a precursor to knee imaging. Radiography, 2004. 10(4): p.
277-285.
72. Pal, N.R. and S.K. Pal, A review on image segmentation techniques. Pattern
Recognition, 1993. 26(9): p. 1277-1294.
73. Olabarriaga, S.D. and A.W.M. Smeulders, Interaction in the segmentation of
medical images: A survey. Medical Image Analysis, 2001. 5(2): p. 127-142.
74. Pham, D.L., C. Xu, and J.L. Prince, Current methods in medical image
segmentation. Annual Review of Biomedical Engineering, 2000. 2(1): p. 315-
337.
75. Behiels, G., et al., Evaluation of image features and search strategies for
segmentation of bone structures in radiographs using Active Shape Models.
Medical Image Analysis, 2002. 6(1): p. 47-62.
76. Lee, H., et al., Left Ventricular segmentation using graph searching on
intensity and gradiant and a priori knowledge (1vGIGA) for short-axis
cardiac mgnetic resonance imaging. Journal of Magnetic Resonance
Imaging, 2008. 28: p. 1393-1401.
77. Santarelli, M.F., et al., Automated cardiac MR image segmentation: theory
and measurement evaluation. Medical Engineering & Physics, 2003. 25(2): p.
149-159.
References
167
78. Zhou, X., et al., Automated segmentation of mammary gland regions in non-
contrast X-ray CT images. Computerized Medical Imaging and Graphics,
2008. 32(8): p. 699-709.
79. Zoroofi, R.A., et al., Automated segmentation of necrotic femoral head from
3D MR data. Computerized Medical Imaging and Graphics, 2004. 28(5): p.
267-278.
80. Peter, Z., et al., A constrained region growing approach based on watershed
for the segmentation of low contrast structures in bone micro-CT images.
Pattern Recognition, 2008. 41(7): p. 2358-2368.
81. Hirano, S. and Y. Hata, Fuzzy expert system for foot CT image segmentation.
Image and Vision Computing, 2001. 19(4): p. 207-216.
82. Cristoforetti, A., et al., Isolation of the left atrial surface from cardiac multi-
detector CT images based on marker controlled watershed segmentation.
Medical Engineering & Physics, 2008. 30(1): p. 48-58.
83. DeVries, N., et al., Validation of phalanx bone three-dimensional surface
segmentation from computed tomography images using laser scanning.
Skeletal Radiology, 2008. 37(1): p. 35-42.
84. Haas, B., et al., Automatic segmentation of thoracic and pelvic CT images for
radiotherapy planning using implicit anatomic knowledge and organ-specific
segmentation strategies. Physics in Medicine and Biology, 2008. 53(6): p.
1751-1771.
85. de Bruin, P.W., et al., Interactive 3D segmentation using connected
orthogonal contours. Computers in Biology and Medicine, 2005. 35(4): p.
329-346.
86. Bokde, A.L.W., et al., Reliable manual segmentation of the frontal, parietal,
temporal, and occipital lobes on magnetic resonance images of healthy
subjects. Brain Research Protocols, 2005. 14(3): p. 135-145.
References
168
87. Stammberger, T., et al., Interobserver reproducibility of quantitative
cartilage measurements: comparison of B-spline snakes and manual
segmentation. Magnetic Resonance Imaging, 1999. 17(7): p. 1033-1042.
88. Bryce, C.D., et al., Validation of three-dimensional models of in situ
scapulae. Journal of Shoulder and Elbow Surgery, 2008. 17(5): p. 825-832.
89. Ward, K.A., J.E. Adams, and T.N. Hangartner, Recommendations for
Thresholds for Cortical Bone Geometry and Density Measurement by
Peripheral Quantitative Computed Tomography. Calcified Tissue
International, 2005. 77(5): p. 275-280.
90. Jong Bae, K., et al., A Real-Time Region-Based Motion Segmentation Using
Adaptive Thresholding and K-Means Clustering. Lecture notes in computer
science, 2001. 2256: p. 213-224.
91. Wiemker, R. and V. Pekar, Fast computation of isosurface contour spectra
for volume visualization. International Congress Series, 2001. 1230: p. 389-
394.
92. Batenburg, K.J. and J. Sijbers, Optimal Threshold Selection for Tomogram
Segmentation by Projection Distance Minimization. Medical Imaging, IEEE
Transactions on, 2009. 28(5): p. 676-686.
93. Guo, R. and S.M. Pandit, Automatic threshold selection based on histogram
modes and a discriminant criterion. Machine Vision and Applications, 1998.
10(5): p. 331-338.
94. Wang, L. and J. Bai, Threshold selection by clustering gray levels of
boundary. Pattern Recognition Letters, 2003. 24(12): p. 1983-1999.
95. Zhang, J., et al., Fast segmentation of bone in CT images using 3D adaptive
thresholding. Computers in Biology and Medicine, 2010. 40(2): p. 231-236.
96. Marr, D. and E. Hildreth, Theory of Edge Detection. Proceedings of the Royal
Society of London. Series B. Biological Sciences, 1980. 207(1167): p. 187-
217.
References
169
97. Testi, D., et al., Border-tracing algorithm implementation for the femoral
geometry reconstruction. Computer Methods and Programs in Biomedicine,
2001. 65(3): p. 175-182.
98. Canny, J., A computational approach to edge detection. IEEE Transaction on
Pattern Analysis and Machine Intelligence, 1986. 8(6): p. 679-698.
99. Demigny, D., A discrete expression of Canny's criteria for step edge detector
performances evaluation. 1997. p. 1199-1211.
100. Bourennane, E., et al., Generalization of Canny-Deriche filter for detection of
noisy exponential edge. Signal Processing, 2002. 82(10): p. 1317-1328.
101. Heath, M.D., et al., A robust visual method for assessing the relative
performance of edge-detection algorithms. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 1997. 19(12): p. 1338-1359.
102. Adams, R. and L. Bischof, Seeded region growing. Pattern Analysis and
Machine Intelligence, IEEE Transactions on, 1994. 16(6): p. 641-647.
103. Yan, K., K. Engelke, and W.A. Kalender, A new accurate and precise 3-D
segmentation method for skeletal structures in volumetric CT data. Medical
Imaging, IEEE Transactions on, 2003. 22(5): p. 586-598.
104. Syoji, K., K. Katsuya, and H. Yutaka, Fully Automated Segmentation of
Cerebral Ventricles from 3-D SPGR MR Images using Fuzzy Representative
Line. Soft computing, 2006. 10.
105. Ahmadian, A., et al., Correction of oral contrast artifacts in CT-based
attenuation correction of PET images using an automated segmentation
algorithm. European Journal of Nuclear Medicine & Molecular Imaging,
2008. 35(10): p. 1812-1823.
106. Middleton, I. and R.I. Damper, Segmentation of magnetic resonance images
using a combination of neural networks and active contour models. Medical
Engineering & Physics, 2004. 26(1): p. 71-86.
References
170
107. Ramme, A., et al., Semi-automated Phalanx Bone Segmentation Using the
Expectation Maximization Algorithm. Journal of Digital Imaging, 2009.
22(5): p. 483-491.
108. Ramme, A., et al., EM Segmentation of the Distal Femur and Proximal Tibia:
A High-Throughput Approach to Anatomic Surface Generation. Annals of
Biomedical Engineering, 2011. 39(5): p. 1555-1562.
109. Chen, T., L. Serra, and H. Ng, Surface extraction: dividing voxels.
International Congress Series, 2004. 1268: p. 225-230.
110. Newman, T.S. and H. Yi, A survey of the marching cubes algorithm.
Computers & Graphics, 2006. 30(5): p. 854-879.
111. Lucas, L., D. Gillard, and Y. Rémion, A new unsupervised cube-based
algorithm for iso-surface generation. Computer Networks and ISDN
Systems, 1997. 29(14): p. 1737-1744.
112. Lorensen, W.E. and H.E. Cline, Marching cubes: A high resolution 3D
surface construction algorithm. SIGGRAPH Comput. Graph., 1987. 21(4): p.
163-169.
113. Popescu, F., et al., A new method to compare planned and achieved position
of an orthopaedic implant. Computer Methods and Programs in Biomedicine,
2003. 71(2): p. 117-127.
114. Tocheri, M.W., et al., A 3D quantitative comparison of trapezium and
trapezoid relative articular and nonarticular surface areas in modern
humans and great apes. Journal of Human Evolution, 2005. 49(5): p. 570-
586.
115. Li, K., et al., Mapping ligament insertion sites onto bone surfaces in knee by
co-registration of CT and digitization data. Journal of Biomechanics, 2009.
42(15): p. 2624-2626.
References
171
116. Schmutz, B., K.J. Reynolds, and J.P. Slavotinek, Development and validation
of a generic 3D model of the distal femur. Computer Methods in
Biomechanics & Biomedical Engineering, 2006. 9(5): p. 305-312.
117. Ruedi, T.P. and W.M. Murphy, AO Principles of Fracture Management.
Thieme Stuttgart, New York, 2001: p. 50-51.
118. Haraguchi, K., et al., Comparison of Fit and Fill Between Anatomic Stem and
Straight Tapered Stem Using Virtual Implantation on the ORTHODOC
Workstation. Computer Aided Surgery, 2001. 6(5): p. 290-296.
119. Genant, H. and S. Majumdar, High-resolution magnetic resonance imaging of
trabecular bone structure. Osteoporosis International, 1997. 7(0): p. 135-139.
120. Majumdar, S., et al., Magnetic resonance imaging of trabecular bone
structure in the distal radius: Relationship with X-ray tomographic
microscopy and biomechanics. Osteoporosis International, 1996. 6(5): p. 376-
385.
121. Wehrli, F.W., et al., Role of Magnetic Resonance for Assessing Structure and
Function of Trabecular Bone. Topics in Magnetic Resonance Imaging, 2002.
13(5): p. 335-355.
122. Liu, J., et al., Rigid model-based 3D segmentation of the bones of joints in
MR and CT images for motion analysis. Medical Physics, 2008. 35(8): p.
3637-3649.
123. Pillai, R.R., et al., MRI-based modeling for evaluation of in vivo contact
mechanics in the human wrist during active light grasp. Journal of
Biomechanics, 2007. 40(12): p. 2781-2787.
124. Moro-oka, T.-a., et al., Can magnetic resonance imaging-derived bone
models be used for accurate motion measurement with single-plane three-
dimensional shape registration? Journal of Orthopaedic Research, 2007.
25(7): p. 867-872.
References
172
125. DeFrate, L.E., et al., In vivo tibiofemoral contact analysis using 3D MRI-
based knee models. Journal of Biomechanics, 2004. 37(10): p. 1499-1504.
126. Chen, B., et al., Combined MR imaging towards subject-specific knee contact
analysis. Visual Computer, 2011. 27(2): p. 121-128.
127. Hao, Z., et al., A finite element 3D model of in vivo human knee joint based
on MRI for the tibiofemoral joint contact analysis. 2007. p. 616-622.
128. Gold, G.E., et al., Musculoskeletal MRI at 3.0 T: Initial Clinical Experience.
American Journal of Roentgenology, 2004. 183(5): p. 1479-1486.
129. Tsurusaki, M., et al., Quantitative and qualitative comparison of 3.0 T and
1.5 T MR imaging of the liver in patients with diffuse parenchymal liver
disease. European Journal of Radiology, 2009. 72(2): p. 314-320.
130. Ramalho, M., et al., Liver MR Imaging: 1.5T versus 3T. Magnetic Resonance
Imaging Clinics of North America, 2007. 15(3): p. 321-347.
131. Rakow-Penner, R., et al., Relaxation times of breast tissue at 1.5T and 3T
measured using IDEAL. Journal of Magnetic Resonance Imaging, 2006.
23(1): p. 87-91.
132. Frayne, R., et al., Magnetic resonance imaging at 3.0 Tesla: Challenges and
advantages in clinical neurological imaging. Investigative Radiology, 2003.
38(7): p. 385-402.
133. Beyersdorff, D., et al., MRI of Prostate Cancer at 1.5 and 3.0 T: Comparison
of Image Quality in Tumor Detection and Staging. American Journal of
Roentgenology, 2005. 185(5): p. 1214-1220.
134. Schindera, S.T., et al., Abdominal Magnetic Resonance Imaging at 3.0 T:
What Is the Ultimate Gain in Signal-to-Noise Ratio? Academic Radiology,
2006. 13(10): p. 1236-1243.
135. Bowers, M.E., et al., Effects of ACL interference screws on articular
cartilage volume and thickness measurements with 1.5 T and 3 T MRI.
Osteoarthritis and Cartilage, 2008. 16(5): p. 572-578.
References
173
136. Chen, C.A., et al., Cartilage morphology at 3.0T: Assessment of three-
dimensional magnetic resonance imaging techniques. Journal of Magnetic
Resonance Imaging, 2010. 32(1): p. 173-183.
137. Stehling, C., et al., Three-Tesla Magnetic Resonance Imaging of the Wrist:
Diagnostic Performance Compared to 1.5-T. Journal of Computer Assisted
Tomography, 2009. 33(6): p. 934-939.
138. Lambert, A., et al., Rotator cuff tears: value of 3.0T MRI. Journal of
Radiology, 2009. 90: p. 583-588.
139. Duewell, S.H., et al., Musculoskeletal MR imaging at 4 T and at 1.5 T:
comparison of relaxation times and image contrast. Radiology, 1995. 196(2):
p. 551-555.
140. Bernstein, M.A., J. Huston, and H.A. Ward, Imaging artifacts at 3.0T.
Journal of Magnetic Resonance Imaging, 2006. 24(4): p. 735-746.
141. Kuo, R., et al., 3.0 Tesla imaging of the musculoskeletal system. Journal of
Magnetic Resonance Imaging, 2007. 25(2): p. 245-261.
142. Soher, B.J., B.M. Dale, and E.M. Merkle, A Review of MR Physics: 3T versus
1.5T. Magnetic Resonance Imaging Clinics of North America, 2007. 15(3): p.
277-290.
143. Ramnath, R.R., 3T MR imaging of the musculoskeletal system (Part II):
clinical applications. Magnetic Resonance Imaging Clinics of North
America, 2006. 14(1): p. 41-62.
144. Lawrence, N.T., 3T MRI in clinical practice. Applied Radiology, 2005. 34(1):
p. 8.
145. Mosher, T.J., Musculoskeletal Imaging at 3T: Current Techniques and Future
Applications. Magnetic Resonance Imaging Clinics of North America, 2006.
14(1): p. 63-76.
146. Norris, D.G., High field human imaging. Journal of Magnetic Resonance
Imaging, 2003. 18(5): p. 519-529.
References
174
147. Gold, G.E., et al., Musculoskeletal MRI at 3.0 T: Relaxation Times and Image
Contrast. American Journal of Roentgenology, 2004. 183(2): p. 343-351.
148. Stanisz, G.J., et al., T1, T2 relaxation and magnetization transfer in tissue at
3T. Magnetic Resonance in Medicine, 2005. 54(3): p. 507-512.
149. Delfaut, E.M., et al., Fat Suppression in MR Imaging: Techniques and
Pitfalls. Radiographics, 1999. 19(2): p. 373-382.
150. Abduljalil, A.M. and P.-M.L. Robitaille, Macroscopic Susceptibility in Ultra
High Field MRI. Journal of Computer Assisted Tomography, 1999. 23(6): p.
832-841.
151. Peh, W.C.G. and J.H.M. Chan, Artifacts in musculoskeletal magnetic
resonance imaging: identification and correction. Skeletal Radiology, 2001.
30(4): p. 179-191.
152. Firbank, M.J. and et al., A comparison of two methods for measuring the
signal to noise ratio on MR images. Physics in Medicine and Biology, 1999.
44(12): p. N261.
153. Dietrich, O., et al., Measurement of signal-to-noise ratios in MR images:
Influence of multichannel coils, parallel imaging, and reconstruction filters.
Journal of Magnetic Resonance Imaging, 2007. 26(2): p. 375-385.
154. Henkelman, R.M., Measurement of signal intensities in the presence of noise
in MR images. Medical Physics, 1985. 12(2): p. 232-233.
155. Edelstein, W.A., P.A. Bottomley, and L.M. Pfeifer, A signal-to-noise
calibration procedure for NMR imaging systems. Medical Physics, 1984.
11(2): p. 180-185.
156. Stadler, A., et al., Artifacts in body MR imaging: their appearance and how
to eliminate them. European Radiology, 2007. 17(5): p. 1242-1255.
157. Odille, F., et al., Generalized MRI reconstruction including elastic
physiological motion and coil sensitivity encoding. Magnetic Resonance in
Medicine, 2008. 59(6): p. 1401-1411.
References
175
158. Besl, P.J. and N.D. McKay, A Method for Registration of 3-D Shapes. IEEE
Transaction on Pattern Analysis and Machine Intelligence, 1992. 14(2): p.
239-256.