Transcript VBM - UCL

VBM
Voxel-based morphometry
Nicola Hobbs & Marianne Novak
Thanks to Susie Henley
Overview
• Background
• Pre-processing steps
•
•
•
•
Analysis
Multiple comparisons
Pros and cons of VBM
Optional extras
Background
• VBM is a voxel-wise comparison of local tissue
volumes within a group or across groups
• Whole-brain analysis, does not require a priori
assumptions about ROIs; unbiased way of
localising structural changes
• Can be automated, requires little user intervention
 compare to manual ROI tracing
Basic Premise
1. Spatial normalisation (alignment) into standard
space
2. Segmentation of tissue classes
3. Modulation - adjust for volume changes during
normalisation
4. Smoothing - each voxel is a weighted average of
surrounding voxels
5. Statistics - localise & make inferences about
differences
VBM Processing
Step 1: normalisation
• Aligns images by warping to standard stereotactic space
• Affine step – translation, rotation, scaling, shearing
• Non-linear step
• Adjust for differences in
• head position/orientation in scanner
• global brain shape
• Any remaining differences (detectable by VBM) are due to
smaller-scale differences in volume
ORIGINAL
IMAGE
SPATIAL
NORMALISATION
TEMPLATE
IMAGE
SPATIALLY
NORMALISED
IMAGE
2. Tissue segmentation
• Aims to classify image as GM, WM or CSF
• Two sources of information
SPATIALLY
a) Spatial prior probability maps
NORMALISED
b) Intensity information in the image itself
IMAGE
GREY MATTER
WHITE MATTER
CSF
a) Spatial prior probability maps
• Smoothed average of GM
from MNI
• Intensity at each voxel
represents probability of
being GM
• SPM compares the original
image to this to help work
out the probability of each
voxel in the image being
GM (or WM, CSF)
b) Image intensities
• Intensities in the image fall into roughly 3 classes
• SPM can also assign a voxel to a tissue class by seeing
what its intensity is relative to the others in the image
• Each voxel has a value between 0 and 1, representing the
probability of it being in that particular tissue class
• Includes correction for image intensity non-uniformity
Generative model
• Segmentation into tissue types
• Bias Correction
• Normalisation
• These steps cycled through until normalisation
and segmentation criteria are met
Step 3: modulation
• Corrects for changes in volume induced by normalisation
• Voxel intensities are multiplied by the local value in the
deformation field from normalisation, so that total GM/WM
signal remains the same
• Allows us to make inferences about volume, instead of
concentration
Modulation
i
i
i / δV
X δV
normalisation
modulation
• E.g. During normalisation TL in AD subject expands to
double the size
• Modulation multiplies voxel intensities by Jacobian from
normalisation process (halve intensities in this case).
• Intensity now represents relative volume at that point
Is modulation optional?
• Unmodulated data: compares “the proportion of grey or
white matter to all tissue types within a region”
• Hard to interpret
• Not useful for looking at e.g. the effects of degenerative disease
• Modulated data: compares volumes
• Unmodulated data may be useful for highlighting areas of
poor registration (perfectly registered unmodulated data
should show no differences between groups)
Step 4: Smoothing
• Convolve with an isotropic Gaussian kernel
• Each voxel becomes weighted average of surrounding voxels
• Smoothing renders the data more normally distributed (Central Limit
theorem)
• Required if using parametric statistics
• Smoothing compensates for inaccuracies in normalisation
• Makes mass univariate analysis more like multivariate analysis
• Filter size should match the expected effect size
• Usually between 8 – 14mm
Smoothing
SMOOTH
WITH 8MM
KERNEL
8 mm
VBM: Analysis
•
•
•
•
•
What does the SPM show in VBM?
Cross-sectional VBM
Multiple comparison corrections
Pros and cons of VBM
Optional extras
VBM: Cross-sectional analysis overview
• T1-weighted MRI from one or more groups at a single time
point
• Analysis compares (whole or part of) brain volume
between groups, or correlates volume with another
measurement at that time point
• Generates map of voxel intensities: represent volume of,
or probability of being in, a particular tissue class
What is the question in VBM analysis?
AD
Control
• Take a single voxel, and ask: “are the intensities in the AD
images significantly different to those in the control images
for this particular voxel?”
• eg is the GM intensity (volume) lower in the AD group cf
controls?
• ie do a simple t-test on the voxel intensities
Statistical Parametric Maps (SPM)
• Repeat this for all voxels
• Highlights all voxels where intensities (volume) are
significantly different between groups: the SPM
• SPM showing regions where
Huntington’s patients have lower
GM intensity than controls
• Colour bar shows the t-value
VBM: group comparison
• Intensity for each voxel (V) is a function that models the different
things that account for differences between scans:
• V = β1(AD) + β2(control) + β3(covariates) + β4(global volume) + μ + ε
• V = β1(AD) + β2(control) + β3(age) + β4(gender) + β5(global volume) + μ + ε
• In practice, the contrast of interest is usually t-test
between β1 and β2
• eg “is there significantly more GM in the control than in
the AD scans?”
VBM: correlation
• Correlate images and test scores (eg Alzheimer’s patients
with memory score)
• SPM shows regions of GM or WM where there are
significant associations between intensity (volume) and
test score
• V = β1(test score) + β2(age) + β3(gender) + β4(global volume) + μ +
ε
• Contrast of interest is whether β1 (slope of association
between intensity & test score) is significantly different to
zero
Correcting for Multiple Comparisons
• 200,000 voxels per scan ie 200,000 t-tests
• If you do 200,000 t-tests at p<0.05, by chance 10,000 will be
false positives
• Bad practice…
• A strict Bonferroni correction would reduce the p value for each
test to 0.00000025
• However, voxel intensities are not independent, but correlated
with their neighbours
• Bonferroni is therefore too harsh a correction and will lose true
results
Familywise Error
• SPM uses Gaussian Random Field theory (GRF)1
• Using FWE, p<0.05: 5% of ALL our SPMs will contain a false
positive voxel
• This effectively controls the number of false positive regions rather
than voxels
• Can be thought of as a Bonferroni-type correction, allowing for multiple
non-independent tests
• Good: a “safe” way to correct
• Bad: but we are probably missing a lot of true positives
1
http://www.mrc-cbu.cam.ac.uk/Imaging/Common/randomfields.shtml
q value
FDR
False Discovery Rate
• FDR more recent
q<0.05
Voxel
• It controls the expected proportion of false positives among
suprathreshold voxels only
• Using FDR, q<0.05: we expect 5% of the voxels for each SPM to
be false positives (1,000 voxels)
• Bad: less stringent than FWE so more false positives
• Good: fewer false negatives (ie more true positives)
• But: assumes independence of voxels: avoid….?
VBM Pros
1. Objective analysis
2. Do not need priors – more exploratory
3. Automated
VBM Cons
1. False positives: misregistration, FDR
2. False negatives: FWE
3. More difficult to pick up differences in areas with high
inter-subject variance: low signal to noise ratio
Other VBM Issues
• Optimised VBM: GM to GM warping, then applied to
whole brain image (better GM alignment); Good et al,
Neuroimage 2001 (SPM 2)
• Diffeomorphic warping: DARTEL
• Multivariate techniques: including classification/SVM
• Longitudinal scan analysis: two time points especially
18 iterations to form
template
Ashburner
Neuroimage
2007
Standard preprocessing: areas of decreased volume in depressed subjects
DARTEL preprocessing: areas of decreased volume in depressed subjects
Longitudinal VBM
• Baseline and follow-up image are registered together nonlinearly (fluid registration), NOT using spm software
• Voxels at follow-up are warped to voxels at baseline
• Represented visually as a voxel compression map
showing regions of contraction and expansion
Fluid Registered Image
FTD
(semantic
dementia)
Voxel
compression
map
1 year
contracting
expanding
Optimised VBM
1. Affine registration
to SPM2 T1 template
Native space images
Standard space images
2. Segmentation
GM segments
3. Estimate normalisation
parameters for GM
segments to SPM2 GM
template
Normalisation parameters
GM to GM
4. Normalisation using parameters
from step 3; GM is well-aligned
Standard space images
5. Segmentation
GM segments
6. Modulation: correcting
for spatial changes
introduced in normalisation
Mod GM
7. Masking: segments are
multiplied by binary
region to exclude any
non-brain
Masked GM
8. Smoothed at
8mm FWHM
Smoothed, Masked, mod GM
Resources and references
• http://www.fil.ion.ucl.ac.uk/spm (the SPM homepage)
• http://imaging.mrc-cbu.cam.ac.uk/imaging/CbuImaging (neurimaging wiki
homepage)
• http://www.mrc-cbu.cam.ac.uk/Imaging/Common/randomfields.shtml (for multiple
comparisons info)
• Ashburner J, Friston KJ. Voxel-based morphometry--the methods. Neuroimage
2000; 11: 805-821 (the original VBM paper)
• Good CD, Johnsrude IS, Ashburner J, Henson RN, Friston KJ, Frackowiak RS.
A voxel-based morphometric study of ageing in 465 normal adult human brains.
Neuroimage 2001; 14: 21-36 (the optimised VBM paper)
• Ridgway GR, Henley SM, Rohrer JD, Scahill RI, Warren JD, Fox NC. Ten
simple rules for reporting voxel-based morphometry studies. Neuroimage 2008.