Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface...
Transcript of Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface...
![Page 1: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/1.jpg)
Bruce FischlMGH NMR Center
Computational Neuroanatomy.
![Page 2: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/2.jpg)
Collaborators
Anders M. Dale MGH
Andre van der Kouwe MGH
Marty Sereno UCSD
David Salat MGH
Christophe Destrieux Tours
Thanks also to Randy Buckner, Bruce Rosen, Eric Halgren, MarilynAlbert, Gina Kuperberg, Ron Killiany, Diana Rosas, David Kennedy,
Nikos Makris, Verne Caviness, Paul Raines, Chad Wissler, Roger Tootell,Doug Greve, Sean Marrett, Janine Mendola, Rahul Desikan, Kevin Teich,Chris Moore, Christian Haselgrove, Tony Harris, Evelina Busa, Maureen
Glessner, and Nouchine Hadjikhani
![Page 3: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/3.jpg)
Definition 1: The manner in which the neuroanatomical structure of thebrain facilitates or carries out computations.
Computational Neuroanatomy:
Definition 2: The application of computational techniques to modelneuroanatomical structures.
![Page 4: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/4.jpg)
Definition 1: The manner in which the neuroanatomical structure of thebrain facilitates or carries out computations.
Computational Neuroanatomy:
Definition 2: The application of computational techniques to modelneuroanatomical structures.
![Page 5: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/5.jpg)
Definition 1: The manner in which the neuroanatomical structure of thebrain facilitates or carries out computations.
Computational Neuroanatomy:
Definition 2: The application of computational techniques to modelneuroanatomical structures.
![Page 6: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/6.jpg)
Warning!
There are no textbooks on computational neuroanatomy:
Much of what you hear in this lecture will be opinion!
![Page 7: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/7.jpg)
Talk Outline
• The Spatial Structure of Retinotopic Cortex.
• Cortical Analysis.
• Subcortical Analysis.
![Page 8: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/8.jpg)
Talk Outline
• The Spatial Structure of Retinotopic Cortex.
• Cortical Analysis.
• Subcortical Analysis.
![Page 9: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/9.jpg)
How is the Visual Field Representedin Mammalian Cortex?*
(Physically Flattened Macaque V1)Stimulus 2-DG map of V1
*thanks to Eric Schwartz for this slide
![Page 10: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/10.jpg)
What is the form of the retino-cortical map function?
First insight: Burkhardt Fischer (1970):
If retinal cell density/length is 1/r
Then several possible optic tract exist maps,one of which is (z=retina, w=cortex):
221(,)(,)log()log()tan/wuxyivxyzxyiyxzxiy−=+==++=+
![Page 11: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/11.jpg)
Problems With Log(z)Hypothesis
• In cat, V1 not really log polar.
• Retinal cell density doesn’t necessarilydetermine the cortical map. This point stilluncertain in both monkey and cat!
• Log(z) has a singularity at the origin – themost important place!
![Page 12: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/12.jpg)
Add a small constant, and map eachhemifield separately: W=log(z+a)
Removal of the Foveal Singularity
Eccentricityà
Polar A
ngleà
*thanks to Eric Schwartz for this slide
![Page 13: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/13.jpg)
Conformal Maps
• A map function is said to be conformal if- It preserves local angles (equivalent to…)- The jacobian of the map function is non-singular
• Riemann map theorem: a conformal map is uniquelydetermined by one point correspondence, one angle, andboundary of the two domains (retina and cortex).
• Log(z) is not conformal, but Log(z+a) is.• Can only meaningfully talk about magnification function
if the map is conformal!
![Page 14: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/14.jpg)
Riemann fit to V1Includes eye position regression and
geodesic brain flattening
*thanks to Eric Schwartz for this slide
![Page 15: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/15.jpg)
What Do Images Look Like inCortex?
Original image “Retinal” image “Cortical” image
*thanks to Eric Schwartz for this slide
![Page 16: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/16.jpg)
Summary of CurrentKnowledge of Spatial Maps
• They exist and are strongly space-variant incat, owl, monkey, human etc.
• They are approximately conformal (V1).
• We don’t know if they are “functional” ornot.
• We don’t know how to do visual computationon SV maps in biology or in computers.
![Page 17: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/17.jpg)
Talk Outline
• The Spatial Structure of Retinotopic Cortex.
• Cortical Analysis.
• Subcortical Analysis.
![Page 18: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/18.jpg)
Talk Outline
• The Spatial Structure of Retinotopic Cortex.
• Cortical Analysis.
• Subcortical Analysis.
![Page 19: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/19.jpg)
None of the preceding analysis of the spatialstructure of the representation of the visualfield in V1 could have been done withoutknowing the position and orientation of thecortex.
Why Is a Model of theCortical Surface Useful?
![Page 20: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/20.jpg)
Why Is a Model of theCortical Surface Useful?
Local functional organization of cortex is largely 2-dimensional!
From (Sereno et al, 1995, Science).
![Page 21: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/21.jpg)
Flat Map of Monkey Visual Areas
D.J. Felleman and D.C. Van Essen, CC, 1991
![Page 22: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/22.jpg)
Why Is Constructing aModel of The Cortical
Surface Difficult?
The cortex is highly folded!
• Partial voluming.
• Subject motion.
• Susceptibility artifacts.
• Bias field.
• Tissue inhomogeneities.
Intensity of a tissueclass varies as a
function of spatiallocation
![Page 23: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/23.jpg)
Sources of within-classintensity variation
• Partial voluming – a single voxel may contain more than one tissue type.
• Bias field – effective flip angle or sensitivity of receive coil
may vary across space.
• Tissue inhomogeneities – even within tissue type (e.g. cortical gray
matter), intrinsic properties such as T1, PD
can vary (up to 20%).
![Page 24: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/24.jpg)
Contrast-to-Noise Ratio
Higher CNR values imply the class distributions overlap less.
All the previous effects reduce the CNR.
( )22
2
BA
BACNR+−=
For two classes, A and B, the contrast-to-noise ratio (CNR) is
given by (one possible definition):
![Page 25: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/25.jpg)
Assigning tissue classes to voxels can be difficult
T1 weighted MR volume
![Page 26: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/26.jpg)
Goal: Reconstruction of theCortical Surface
Generate a geometrically accurate and topologicallycorrect model of the cerebral cortex.
Uses of the surface reconstruction include:
• Visualization of functional and structural neuroimaging data.
• Calculation of morphometric properties of the cortex.
• High-resolution averaging of cortical data across subjects.
• Increasing spatial resolution of EEG/MEG data.
![Page 27: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/27.jpg)
Which Surface toReconstruct?
Pial surface is ultimate goal, but pretty much impossibleto directly generate a representation of from MRI images(many have tried!).
Alternative: construct an interim representation of theinterface between gray matter and white matter, and useit to infer the location of the true cortical surface (Daleand Sereno, 1993).
![Page 28: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/28.jpg)
Skull Stripping and building ofBoundary Element Models
![Page 29: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/29.jpg)
Conductivity Boundaries for BEM
Inner Skull Outer Skull Outer Skin
![Page 30: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/30.jpg)
MRI Segmentation and SurfaceReconstruction
![Page 31: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/31.jpg)
Surface RepresentationsTwo Choices:
• Lagrangian – generate an explicit representation ofthe surface through a tessellation. Surfacedeformations are then carried out by computing themovement of points (vertices) on the surface.
• Eulerian – represent the surface by embedding it in ahigher-dimensional space. The surface is representedimplicitly as the set of points with constant value in thehigher dimensional function (the “level-set” approach ofOsher and Sethian).
![Page 32: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/32.jpg)
TessellationTessellation - a covering of a space with a pattern suchthat the elements of the pattern do not overlap.
In our case (and typically), we cover the cortex withtriangles. The tessellation is thus made up of vertices(points), faces (triangles) and edges (line segments).
![Page 33: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/33.jpg)
Tessellation: example
![Page 34: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/34.jpg)
Surface Inflation: Equations
� �=
−=V
i iNnin
tind dd
VJ
1 )(
20 )(4
1 tn
ti
tind xx −=Metric Distortion Term:
SSd JJJ +=Complete Energy Functional:
Smoothness Term: � �=
−=V
i iNnniS iN
J1 )(
2)(
)(#
1xx
Where N(i) is aneighborhood functionthat returns the set ofneighbors of the ithvertex.
To “inflate” surface model: compute gradient of J with respect tothe coordinates of each vertex xi, and move vertex in oppositedirection (gradient descent), while constraining the total surfacearea to be constant.
![Page 35: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/35.jpg)
Surface Inflation
![Page 36: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/36.jpg)
Gray-white boundary
Pial surface
White matter and pial surfaces
![Page 37: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/37.jpg)
Gray-white boundary
Pial surface
Representing the pial surface
![Page 38: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/38.jpg)
Quasi-Isometric Flattening:Equations
� �=
−=V
i iBnin
tind
r
ddV
J1 )(
20 )(4
1 tn
ti
tind xx −=Metric Distortion Term:
TTd JJJ +=Complete Energy Functional:
Topology Term: 0i
ti
iA
AR =�
=√√↵
����
−+=
F
ii
kR
T Rk
e
FJ
i
1
)1log(1
Note: distances din are for macroscopic geodesics: vertices iand n are not necessarily neighbors.
![Page 39: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/39.jpg)
Quasi-Isometric Flattening:Equations (cont)
Where:
-0.1 0 0.10
0.1Ai
t – oriented area of ith face in tessellation
F – number of faces in tessellation
k – positive real constant
Topology Term: 0i
ti
iA
AR =�
=√√↵
����
−+=
F
ii
kR
T Rk
e
FJ
i
1
)1log(1
![Page 40: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/40.jpg)
superior temporal
Inflated surface with cuts
Metrically optimal flat map
calcarine
central
sylvian
anterior
posterior
Surface Flattening – WholeHemisphere
![Page 41: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/41.jpg)
Borrowed from (Halgren et al., 1999)
![Page 42: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/42.jpg)
Inter-subject Registration
Problem: this information is in general unavailable
Typical solution: align image intensities and hope this resultsin alignment of function/structure as well.
Goal: align functionally homologous points across subjects(e.g. hippocampus with hippocampus, amygdala withamygdala, etc…).
![Page 43: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/43.jpg)
Inter-Subject Registration: Standard Formulation
Some typical forms for f:
-Linear/Affine (many groups)
-Polynomial (Woods et al. AIR)
-Discrete Cosine Transform (Ashburner and Friston, SPM)
-Navier Stokes (Miller)
Find f that minimizes(T is target image, I is input image, r is spatial coordinate)
rrr dTfI��� − 2))())(((
![Page 44: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/44.jpg)
Some Definitions
p(A|B) is called the likelihood of A given B. If p(A|B) isexponential (e.g. Gaussian) in form, the log of the likelihoodis much easier to work with. Usually A is some observed dataand B is a set of model parameters that we want to estimate.
The B that maximizes p(A|B) is called the maximumlikelihood estimate (MLE) of B.
The value of B that maximizes p(B|A) is called the maximuma posteriori (MAP) estimate of B (more on this later).
![Page 45: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/45.jpg)
What does Mean-Squared ErrorEstimation mean from a
Probabilistic Perspective?
Assume rrr dTfITfIp ��� −= 2))())((()),|(log(
Then: ⊆ −=2))())(((),|( rr TfIeTfIp
f is the maximum likelihood solution assuming the
image can be modeled as a set of random variables with
means T(r) and equal variances.
![Page 46: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/46.jpg)
Talairach CoordinatesCan mean many things, but most common is lineartransform to align input image with a target image that isaverage of many individuals aligned with the atlas ofTalairach and Tournoux (1988).
Not Good For Cortex!
• Typical transform is too low dimensional to account forvariability in cortical folds.
• Landmarks are subcortical (and far from much of cortex).
• Implicit assumption that 3D metric is appropriate one.
![Page 47: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/47.jpg)
Average of 40
Single subject
Talairach averaging
![Page 48: Computational Neuroanatomy. - MITweb.mit.edu/hst.583/www/course2001/LECTURES/hst583_lect9...Surface Representations Two Choices: • Lagrangian – generate an explicit representation](https://reader034.fdocuments.us/reader034/viewer/2022042210/5eaeb83ad9ae325edf4b1514/html5/thumbnails/48.jpg)
How to align different corticalsurfaces?