Morphometrics with

SPM12
John Ashburner

Wellcome Trust Centre for Neuroimaging,

12 Queen Square, London, UK.
What kind of differences are we looking for?
• Usually, we try to localise regions of difference.
• Univariate models.
• Typically involves fitting a GLM
• Typically localising volumetric differences
• Some anatomical differences can not be localised.
• Need multivariate models.
• Differences in terms of proportions among measurements.
• Where would the difference between male and female
faces be localised?

• Need to select the best model of difference to use,

before trying to fill in the details.
Overview
• Voxel-Based Morphometry
• Diffeomorphic Registration
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Voxel-Based Morphometry
• Produce a map of statistically significant differences
among populations of subjects.
• e.g. compare a patient group with a control group.
• or identify correlations with age, test-score etc.
• The data are pre-processed to sensitise the tests to
regional tissue volumes.
• Usually grey or white matter.
Volumetry

T1-Weighted MRI Grey Matter

Original Warped Template
“Modulation” – change of variables.

Deformation Field Jacobians determinants

Encode relative volumes.
 Smoothing
Each voxel after smoothing effectively
becomes the result of applying a weighted
region of interest (ROI).

Before convolution Convolved with a circle Convolved with a Gaussian

 VBM Pre-processing
in SPM12
• Use Segment for
characterising intensity
distributions of tissue classes,
and writing out “imported”
images that Dartel can use.
• Run Dartel to estimate all the
deformations.
• Dartel warping to generate
smoothed, “modulated”,
warped grey matter.
• Statistics.
 Some Explanations of the Differences

Mis-classify Mis-register

Folding

Thinning
Thickening Mis-register

Mis-classify
Some References

• Ashburner & Friston. “Unified Segmentation”. NeuroImage

26:839-851, 2005.
• Ashburner. “A Fast Diffeomorphic Image Registration
Algorithm”. NeuroImage 38:95-113 (2007).
• Ashburner & Friston. “Computing Average Shaped Tissue
Probability Templates”. NeuroImage 45:333-341, 2009.
• Ashburner. “Computational Anatomy with the SPM software”.
Magnetic Resonance Imaging 27(8):1163-1174, 2009.
Overview
• Voxel-Based Morphometry
• Diffeomorphic Registration
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Diffeomorphism
In mathematics, a
diffeomorphism is an
isomorphism in the category of
smooth manifolds. It is an
invertible function that maps one
differentiable manifold to
another, such that both the
function and its inverse are
smooth.
Wikipedia
Deformations
 A A

B A°θ B θ°φ

C B°φ C A°θ°φ

Composition
Small
Deformation A θ°φ ((θ − Id) + (φ − Id)) + Id

Approximation
The composition:

ϑ°φ
Would be approximated with:

Id +((ϑ-Id) + (φ-Id))
C A°θ°φ A(((θ − Id) + (φ − Id)) + Id)

The inversion:

φ-1
Would be approximated with:

Id -(φ-Id)
Not good approximations for large deformations.
A θ°φ ((θ − Id) + (φ − Id)) + Id

C A°θ°φ A(((θ − Id) + (φ − Id)) + Id)

Diffeomorphic Image Registration
• Minimises two terms:
1. A measure of distance between images
2. A measure of the amount of distortion.

Because we can not simply add displacement fields,

large deformations are generated by composing
many small deformations.

The amount of distortion is computed by summing

up the distortion measures from the small
displacements.
Effect of Different Distortion Measures
Two diffeomorphic approaches in SPM
Dartel.
• Uses the same small
deformation composed
multiple times.
• Faster than Geodesic
Shooting.
• Gives similar deformations
to Geodesic Shooting.
• Currently more additional
utilities.
Geodesic Shooting
• Uses the optimal series of
small deformations, which
are composed together.
• More mathematically correct
than Dartel.
• Gives nicer maps of volume
change than Dartel.
• Likely to replace Dartel in
future.
Dartel & GS Compared
Dartel Geodesic Shooting
µ°χ f ° χ−1 µ°θ f°φ

−1 θ φ
χ χ

−1
θ φ
|Jχ| |Jχ | |J | |J |
Simultaneous registration of GM to GM
and WM to WM

Grey matter
Subject 1
White matter Grey matter Subject 3

White matter

Grey matter

White matter
Grey matter
Template Grey matter
White matter
White matter
Subject 2
Subject 4
 Template
Initial
Average

Iteratively generated
from all subjects in
study
After a few
Begin with rigidly iterations
aligned tissue
probability maps

Final
template
Initial
GM images
Warped
GM images
Evaluations of
nonlinear
registration
algorithms
 LPBA40
R lateral orbitofrontal gyrus
R superior occipital gyrus
L lateral orbitofrontal gyrus
L superior occipital gyrus
L angular gyrus
R angular gyrus
R cuneus
L inferior temporal gyrus
L precuneus
L inferior occipital gyrus
R supramarginal gyrus
L cuneus
L supramarginal gyrus
L middle orbitofrontal gyrus
R middle orbitofrontal gyrus
L gyrus rectus
L middle temporal gyrus
R precuneus
R inferior occipital gyrus
R middle occipital gyrus
R inferior temporal gyrus
L middle occipital gyrus
R gyrus rectus
R parahippocampal gyrus
R cingulate gyrus
R fusiform gyrus
L cingulate gyrus
L parahippocampal gyrus
L fusiform gyrus
R middle temporal gyrus
L postcentral gyrus
L superior parietal gyrus
R superior parietal gyrus
L lingual gyrus
R postcentral gyrus
R inferior frontal gyrus
L inferior frontal gyrus
R lingual gyrus
L superior temporal gyrus
L hippocampus
R precentral gyrus
R caudate
R hippocampus
R superior temporal gyrus
L caudate
R putamen
L precentral gyrus
L putamen
R insular cortex
L middle frontal gyrus
R middle frontal gyrus
L insular cortex
R superior frontal gyrus
brainstem
L superior frontal gyrus
cerebellum

0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

Overlap
Why use diffeomorphic registration?

2×2≠3
This is what you get from approximating a
multiplication using additions.
((2-1)+(2-1))+1 = 3

It almost works for values close to 1.

1.01 × 1.01 = 1.0201
((1.01-1)+(1.01-1))+1 = 1.02
Some References

• Ashburner. “A Fast Diffeomorphic Image Registration

Algorithm”. NeuroImage 38:95-113, 2007.
• Ashburner & Friston. “Computing Average Shaped Tissue
Probability Templates”. NeuroImage 45:333-341, 2009.
• Ashburner & Friston. “Diffeomorphic registration using
geodesic shooting and Gauss–Newton optimisation”.
NeuroImage 55(3):954-967, 2011.
• Klein, Andersson, Ardekani, Ashburner, Avants, Chiang,
Christensen, Collins, Gee, Hellier, Song, Jenkinson, Lepage,
Rueckert, Thompson, Vercauteren, Woods, Mann & Parsey.
“Evaluation of 14 nonlinear deformation algorithms applied to
human brain MRI registration”. NeuroImage 46:786-802, 2009.
Overview
• Voxel-Based Morphometry
• Diffeomorphic Registration
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Some 2D Shapes
Shapes aligned to their average
These were the deformations for that
and these are the Jacobian determinants
Cross-Sectional Data

Used 550 T1w brain MRI from

IXI (Information eXtraction
from Images) dataset.
• http://www.brain-
development.org/

Data from three different

hospitals in London:
• Hammersmith Hospital
using a Philips 3T system
• Guy’s Hospital using a
Philips 1.5T system
• Institute of Psychiatry using
a GE 1.5T system5T system
Segmentation
Segmented into GM and WM.
Approximately aligned via rigid-body.
Diffeomorphic Alignment
All GM and WM were diffeomorphically aligned to their common average-
shaped template.
 Divergence Maps

• Used maps of initial velocity

divergence.

• Similar to logarithms of Jacobian

determinants.
• Encode a sort of “growth rate”
Mass-Univariate Analysis – shrinkage with age
 Large T statistics (> 15) – but not very predictive

T Statistic Image The most predictive single voxel

Fitted responses
age−
0.4

0.3

0.2

response at [−7.5, −16.5, 1.5]

0.1

−0.1

−0.2

−0.3

−0.4
−40 −30 −20 −10 0 10 20 30 40 50
age
Some References

• Ashburner & Friston. “Unified Segmentation”. NeuroImage

26:839-851, 2005.
• Ashburner & Friston. “Computing Average Shaped Tissue
Probability Templates”. NeuroImage 45:333-341, 2009.
• Ashburner & Friston. “Diffeomorphic registration using
geodesic shooting and Gauss–Newton optimisation”.
NeuroImage 55(3):954-967, 2011.
Overview
• Voxel-Based Morphometry
• Diffeomorphic Registration
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Multivariate models of form
• In theory, assumptions about structural covariance
among brain regions are more biologically plausible.
Form determined (in part) by spatio-temporal modes of
gene expression.
• Empirical evidence in (eg)
Mechelli, Friston, Frackowiak & Price. Structural
covariance in the human cortex. Journal of Neuroscience
25(36):8303-8310 (2005).

• We should work with the most accurate modelling

assumptions available.
• If a model is accurate, it will make accurate predictions.
Fisher’s Linear Discriminant Analysis

• A multivariate −1
p(x,y=0) = p(x|y=0) p(y=0)
−1
p(x,y=1) = p(x|y=1) p(y=1)

model. −2 −2

• Special case of −3 −3

Feature 2

Feature 2
−4 −4

canonical −5 −5

−6 −6
variates analysis.
•
−7 −7
0 2 4 0 2 4

A generative Feature 1 Feature 1

p(x) = p(x,y=0) + p(x,y=1) p(y=0|x) = p(x,y=0)/p(x)

model. −1 −1

−2 −2

−3 −3
Feature 2

Feature 2
−4 −4

−5 −5

−6 −6

−7 −7
0 2 4 0 2 4
Feature 1 Feature 1
Other linear discrimination approaches
• Can also use
discriminative models. −1
Ground truth
−1
FLDA

−2 −2

• Anatomical differences −3 −3

Feature 2

Feature 2
−4 −4

are encoded by the −5 −5

vector orthogonal to the −6 −6

separating hyper-plane. −7
0 2 4
−7
0 2 4
Feature 1 Feature 1

• The most accurate

SVC Simple Logistic Regression
−1 −1

−2 −2
model of difference is −3 −3
Feature 2

Feature 2
the one that best −4 −4

separates the groups. −5 −5

−6 −6

−7 −7
0 2 4 0 2 4
Feature 1 Feature 1
Regression
• For predicting a continuous variable
 Predicting Age – univariate v multivariate

Single Voxel Combining All Voxels

100

Fitted responses
age− 90
0.4

80
0.3

70
0.2
response at [−7.5, −16.5, 1.5]

60

Predicted Age
0.1

50
0

40
−0.1

30
−0.2

20
−0.3

10
−0.4
−40 −30 −20 −10 0 10 20 30 40 50
age 0
0 10 20 30 40 50 60 70 80 90 100
True Age
Weight Map

For linear classifiers, predictions are made by:

where: y is the prediction

x1, x2, x3 etc are voxels in the image to classify
a1, a2, a3 etc are voxels in a weight map
b is a constant offset.

The weight map can be visualised

Maps
Multivariate weight map Simple T statistic image

100

90

80
Prettier – but
70
does not
60

accurately
Predicted Age

50

40

30
characterise the
20 effects of age.
10

0
0 10 20 30 40 50 60 70 80 90 100
True Age
Some References

• Bishop. Pattern Recognition and

Machine Learning. 2006.
• Rasmussen & Williams. Gaussian
Processes for Machine Learning. MIT
Press, 2006. ISBN-10 0-262-18253-X,
ISBN-13 978-0-262-18253-9.
http://www.gaussianprocess.org/gpml/

• Ashburner & Klöppel. “Multivariate

models of inter-subject anatomical
variability”. NeuroImage 56(2):422-439,
2011.
Overview
• Voxel-Based Morphometry
• Diffeomorphic Registration
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Distances
Biologically plausible measures of anatomical similarity.
Nonlinear distance measures
The 2D shapes (again)
“Scalar momentum” – encodes the original shapes
The 2D shapes (yet again)
Reconstructed from scalar momentum and template.
“Scalar momentum” – encodes the original shapes
Template and it’s gradients Residuals Momentum Velocity Deformation Jacobian
Some References

• Younes, Arrate & Miller. “Evolutions equations in

computational anatomy”. NeuroImage
45(1):S40-S50, 2009.

• Singh, Fletcher, Preston, Ha, King, Marron,

Wiener & Joshi (2010). Multivariate Statistical
Analysis of Deformation Momenta Relating
Anatomical Shape to Neuropsychological
Measures. T. Jiang et al. (Eds.): MICCAI 2010,
Part III, LNCS 6363, pp. 529–537, 2010.

• Various textbooks
Overview
• Voxel-Based Morphometry
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
IXI Data
Original Images Rigidly Aligned Grey Matter
VBM-type Features
Warped Grey Matter “Modulated” Warped GM
Volumetric Measures from Deformation Fields
Jacobian determinants Initial Velocity Divergence
Scalar Momentum
1st Component 2nd Component
Age Prediction - Best Result
Scalar Momentum (10mm FWHM)
100

90

Predicted Age (years) 80

70

60

50

40

30

20

10

0
0 10 20 30 40 50 60 70 80 90 100
Actual Age (years)
Age Prediction – Comparison Among Features
8−Fold Cross−Validation
8.5
Jacobians
Divergences
Rigid GM
Unmodulated GM
Modulated GM
8 Scalar Momentum
RMS error (years)

7.5

6.5

6
0 2 4 6 8 10 12 14 16 18 20
Smoothing (mm)
 Age Prediction – Model Log Likelihoods
Bayesian Model Evidence
−1820
Jacobians
Divergences
Differences > 4.6 −1840 Rigid GM
indicate “decisive” Unmodulated GM
evidence in favour −1860
Modulated GM
Scalar Momentum
of one approach
over another.
−1880
Log Likelihood

−1900

−1920

−1940

−1960

−1980

−2000

−2020
0 2 4 6 8 10 12 14 16 18 20
Smoothing (mm)
Sex Prediction – Best Result
1

0.9

0.8

0.7

0.6
Prediction

0.5

0.4

0.3

0.2

0.1

0
Label
Sex Prediction – Best Result
ROC Curve (AUC=0.9769)
1

0.8

0.6
Specificity

0.4

0.2

0 0.2 0.4 0.6 0.8 1

Sensitivity
Sex Prediction – Comparison Among Features
Gaussian Process (EP)
0.98

0.97

0.96

0.95
AUC

0.94

Jacobians
0.93
Divergences
Rigid GM
Unmodulated GM
0.92 Modulated GM
Scalar Momentum

0.91
0 2 4 6 8 10 12 14 16 18 20
Smoothing (mm)
 Sex Prediction – Model Log Likelihoods
Bayesian Model Evidence
−120

Differences > 4.6

indicate “decisive” −140
evidence in favour
of one approach
over another. −160
Log Likelihood

−180

−200

Jacobians
Divergences
−220
Rigid GM
Unmodulated GM
Modulated GM
−240
Scalar Momentum

−260
0 2 4 6 8 10 12 14 16 18 20
Smoothing (mm)
References

• Singh, Fletcher, Preston, Ha, King, Marron, Wiener & Joshi

(2010). Multivariate Statistical Analysis of Deformation
Momenta Relating Anatomical Shape to Neuropsychological
Measures. T. Jiang et al. (Eds.): MICCAI 2010, Part III, LNCS
6363, pp. 529–537, 2010.

• Rasmussen & Williams. Gaussian Processes for Machine

Learning. MIT Press, 2006. ISBN-10 0-262-18253-X, ISBN-13
978-0-262-18253-9.
http://www.gaussianprocess.org/gpml/
Overview
• Voxel-Based Morphometry
• Tensor-Based Morphometry
• Multivariate Approaches
• Scalar Momentum
• Some Evaluations
• Longitudinal Registration
Longitudinal Registration

• Unified model combines:

• Nonlinear diffeomorphic
registration.
• Rigid-body registration.
• Intensity inhomoheneity
correction.

• All made as mathematically

coherent as possible.
OASIS Data
OAS2 0048
66 year old male with
dementia (MMSE=19,
CDR=1).
Five scans collected over
40 months.

Marcus, D., A. Fotenos, J. Csernansky, J.

Morris, and R. Buckner (2010). Open
access series of imaging studies:
longitudinal MRI data in nondemented
and demented older adults. Journal of
cognitive neuroscience 22 (12), 2677–
2684.
OASIS Data
OAS2 0048
66 year old male with
dementia (MMSE=19,
CDR=1).
Five scans collected over
40 months.

Difference between time

point and first scan.
OASIS Data
OAS2 0048
66 year old male with
dementia (MMSE=19,
CDR=1).
Five scans collected over
40 months.

Expansion/contraction.
Two Longitudinal Scans

Two scans taken 6 years apart

(after rigid registration).

Average and difference images.

Shape average and map of

expansion/contraction
(after nonlinear registration)
Oasis Data
OAS2 0002
75 year old male,
with MCI
(MMSE=22,
CDR=0.5).
Oasis Data
OAS2 0002
75 year old male,
with MCI
(MMSE=22,
CDR=0.5).
Oasis Data
OAS2 0048
66 year old male, with MCI (MMSE=19, CDR=1).
Oasis Data
Data from first 82 subjects (OAS2 0001 to OAS2 0099).
Computed average expansion/contraction rates for each subject.
Warped all data to common anatomical space.
Generated averages.

Mean image Control Dementia

intensity subjects subjects
References

• Ashburner & Ridgway (2013). Symmetric diffeomorphic

modelling of longitudinal structural MRI. Frontiers in
Neuroscience 6(197).