of GM structure: VBM, DBM, cortical

unclesamnorweiganAI and Robotics

Oct 18, 2013 (3 years and 7 months ago)

103 views

Techniques for the analysis
of GM structure:

VBM, DBM, cortical
thickness

Jason Lerch

Why should I care about
anatomy?

Nieman et al, 2007

Dickerson et al, 2008

Verbal Learning

Anatomy
-

behaviour

The methods.


Manual segmentation/volumetry.


Voxel Based Morphometry (VBM).


Deformation/Tensor Based Morphometry
(DBM).


optimized VBM.


automated volumetry.


cortical thickness.

Processing Flow

Manual Segmentation


Identify one or more regions of interest.


Carefully segment these regions for all
subjects.


Statistics on volumes.

Segmentation example

And it was good.


Cons:


Labour intensive and time consuming.


Need to compute inter and intra rater
reliability measures.


Pros:


Can be highly accurate.


Can discern boundaries still invisible to
machine vision.

Preprocessing

Non
-
uniformity correction

Sled, Zijdenbos, Evans: IEEE
-
TMI Feb 1998

Voxel Classification

T2

PD

T1

MS Lesion
Classification

Positional Differences

Brain 1

Brain 2

Overall Size
Differences


Spatial Normalization

Before
Registration

After
Registration

Voxel Based Morphometry


The goal: localize changes in tissue
concentration.

Tissue Density

Proportion of neighbourhood occupied by tissue class

Real world example

VBM statistics


Tissue density modelled by predictor(s).


I.e.: at every voxel of the brain is there a
difference in tissue density between
groups (or correlation with age, etc.)?


Millions of voxels tested, multiple
comparisons have to be controlled.

Example

Paus et al., Science 283:1908
-
1911, 1999

111 healthy
children

Aged 4
-
18

And it was good.


Pros:


Extremely simple and quick.


Can look at whole brain and different
tissue compartments.


By far most common automated
technique
-

easy comparison to other
studies.


Cons


Hard to explain change (WM? GM?).


Hard to precisely localize differences.


Hard time dealing with different size
brains.

Tensor Based Morphometry


The goal: localize differences in brain
shape.

Non
-
linear deformation


Deformations

Jacobians

Chung et al. A unified statistical approach to deformation
-
based
morphometry. Neuroimage (2001) vol. 14 (3) pp. 595
-
606

Childhoo
d Music

Hyde
et al.,
2008

And it was good.


Pros:


Excellent for simple topology (animal
studies).


Excellent for longitudinal data.


Does not need tissue classification.


Cons:


hard matching human cortex from
different subjects.


Can be quite algorithm dependent.

Optimized VBM


The goal: combine the best of VBM and
TBM

Modulation

x

And it was good.


Pros:


More accurate localization than plain
VBM.


Cons:


Dependent on non
-
linear registration
algorithm.


Is it really better than either VBM or TBM
alone?

Automatic
segmentation


The goal: structure volumes without
manual work.

Segmentation


Backpropagation


And it was good.


Pros:


A lot less work than manual
segmentation.


Excellent if image intensities can be
used.


Excellent if non
-
linear registration is
accurate.


Cons:


Not always accurate for small structures.


Hard time dealing with complex cortical
topology.

Cortical Thickness


The goal: measure the thickness of the
cortex.

Processing Steps in Pictures

Processing Continued

4.5mm

1.0mm

Surface
-
based Blurring

And it was good.


Pros:


Extremely accurate localization of
cortical change.


Sensible anatomical measure.


Sensible blurring.


Cons:


Only covers one dimension of one part of
the brain.


Computationally very expensive and
difficult.

Methods Summary

Method

Computation

Comparisons

Localization

Coverage

manual
segmentatio
n

Manual

one
-
few

depends

ROI

VBM

Easy

millions

poor

cerebrum

TBM

Moderate

millions

OK

brain

optimized
VBM

Moderate

millions

OK

cerebrum

automatic
segmentatio
n

Moderate

few

poor

large
structures

cortical
thickness

Hard

thousands

excellent

cortex

Advice, part 1


MRI anatomy studies need more subjects
than fMRI


aim for at least 20 per group.


Acquire controls on same hardware.


Isotropic sequences are your friend.


T1 is enough unless you’re looking for
lesions.

Advice, part 2


Group comparison, strong hypothesis?


manual segmentation.


automatic segmentation:
FreeSurfer.


Group comparison, few hypotheses?


VBM:
SPM, FSL, MINC tools.


automatic segmentation:
FreeSurfer.


Group comparison, cortical hypothesis?


cortical thickness:
FreeSurfer, MINC tools.


sulcal morphology/shape:
BrainVisa/anatomist.


Lesion/stroke?


manual segmentation.


classification:
MINC tools.



Longitudinal data?


deformations:
SPM (Dartel), ANTS, FSL (SIENA), MINC tools.

Acknowledgements

Alan Evans

Alex Zijdenbos

Krista Hyde

Claude Lepage

Yasser Ad
-
Dab’bagh

Tomas Paus

Jens Pruessner

Veronique Bohbot

John Sled

Mark Henkelman

Matthijs van Eede

Jurgen Germann

Judith Rapoport

Jay Giedd

Dede Greenstein

Rhoshel Lenroot

Philip Shaw

Jeffrey Carroll

Michael Hayden

Harald Hampel

Stefan Teipel