Presentation on theme: "VBM Susie Henley and Stefan Klöppel Based on slides by John Ashburner"— Presentation transcript:
1 VBM Susie Henley and Stefan Klöppel Based on slides by John Ashburner
2 Overview Voxel-based Morphometry Problems with VBM Alternative Approaches
3 VBM (voxel-based morphometry) VBM: whole-brain analysis, does not require a priori assumptions about ROIs; unbiased way of localising structural changesDoes a voxel by voxel comparison of local tissue volume.
4 Pre-processing for VBM -should discuss why each of the steps is done, e.g. “align brains up, divide tissue types, correct for changes in volume introduced in warping, and smooth to further correct for registration error”
5 VBM Preprocessing in SPM5 It uses a generative model, which involves:Segmentation into tissue typesGM, WM and CSFBias CorrectionCorrects intensity inhomogeneities in imagesNormalisationAligns images, puts them into the same (standard) spaceThese steps are cycled through until normalisation and segmentation criteria are met
6 SegmentationUses information from tissue probability maps (TPMs) and the intensities of voxels in the image to work out the probability of a voxel being GM, WM or CSFMention that maps are deformed during segmentationICBM Tissue Probabilistic Atlases. These tissue probability maps are kindly provided by the International Consortium for Brain Mapping, John C. Mazziotta and Arthur W. Toga.
7 Bias correction Warping Estimates a function to correct for bias in the image and applies itWarpingThe tissue probability maps (which are in standard space) are warped to match the imagethis gives parameters for registering the image into standard space later
8 The generative modelKeeps doing these steps iteratively until the objective function is minimisedResults in images that are segmented, bias-corrected, and registered into standard spaceDoes the objective function model error? Does it keep going until it thinks it can’t reduce error any more/isn’t changing parameters any more and therefore it has found the best parameters for the data?
9 Modulation Vox[i, v] normalisation modulation Vox[i, v*δV] Vox[i/δV, v*δV]modulationVox[i, v]During modulation voxel intensities are multiplied by the local value in the deformation field from normalisation, so that total GM/WM signal remains the sameChange of intensity now represents volume relative to templateDuring normalisation some volumes will change, e.g. TL of AD “stretches” and doubles in sizeDuring modulation the voxel intensities are multiplied by the Jacobians from the normalisation process, so in this case the intensities are halvedIntensity at each point now represents the relative volume at that point
10 How optional is modulation ? Unmodulated data: compares “the proportion of grey or white matter to all tissue types within a region”Hard to interpretTherefore not very useful for looking at e.g. the effects of degenerative diseaseModulated data: compares volumesUnmodulated data may be useful for highlighting areas of poor registration (perfectly registered unmodulated data should show no differences between groups)
11 SmoothingReasons:Each voxel becomes weighted average of surrounding onesData are more normally distributedSmooth out incorrect normalisationMost studies use a kernel between 8 and 14 mm. depending on the size of the expected effect.
12 VBM: analysisControlADTake a single voxel, and ask e.g. “are the intensities in the AD images significantly lower than those in the control images for this particular voxel?”i.e. do a simple t-test on the voxel intensities
13 VBM: group comparisonAt each voxel intensity is actually modelled as a function of explanatory or confounding variablesV=β1(AD) + β2(control)+ β3(age) +β4(gender) + β5(TIV) + μ + εIn practice most models are set up with similar covariates as above, with the “contrast” of interest being the t-test between β1 and β2
14 SPMHighlight all voxels where intensities (volume) in patient images are significantly lower than controls: this is a statistical parametric mapThe colour bar shows the t-value
15 Correcting confoundsBigger brains will have bigger GM or WM volumes which could confound comparisonsInclude TIV as covariate to correct for differences due to head sizeWith TIV as a covariate we can compare GM assuming no differences in head sizeHere one brain is bigger than the other (and possibly has more GM because of that)
16 Global or local change? A B Brains are of similar size but GM differs globally and locallyAs it stands we would find greater volume in B relative to A except in the thin area on the right-hand sideABIncluding total GM or WM volume as a covariate adjusts for global atrophy and looks for regionally-specific changesWith global GM as a covariate we will find greater volume in A relative to B only in the thin area on the right-hand side
17 Which to use?Comparisons should usually be adjusted for head size (TIV)Inferences may then be based on global differencese.g. what’s the global effect of disease X?Alternatively you may wish to look at regionally specific changese.g. having adjusted for overall atrophy, are there any regions which still show relative sparing or loss of tissue?di
18 Some Explanations of the Differences FoldingMis-classifyMis-registerThickeningThinningMis-classifyMis-register
19 Validity of the statistical tests in SPM Errors (residuals) need to be normally distributed throughout brain for stats to be validAfter smoothing this is usually true BUTInvalidates experiments that compare one subject with a groupCorrection for multiple comparisonsValid for corrections based on peak heights (voxel-wise)Not valid for corrections based on cluster extentsThis requires smoothness of residuals to be uniformly distributed but it’s not in VBM because of the non-stationary nature of underlying neuroanatomyBigger blobs expected in smoother regions, purely by chance
20 Alternatives Improve normalisation use multivariate approaches Lao: ‘Morphological classification of brains via high-dimensional shape transformations and machine learning methods‘ (2004) NeuroImage.
21 Multivariate Approaches An alternative to mass-univariate testing (SPMs)Shape is multivariateGenerate a description of how to separate groups of subjectsUse training data to develop a classifierUse the classifier to diagnose test data
22 Points to think about What do results mean? VBM generally Limitations of spatial normalisation for aligning small-volume structures (e.g. hippo, caudate)VBM in degenerative brain diseases:Spatial normalisation of atrophied scansOptimal segmentation of atrophied scansOptimal smoothing width for expected volume loss
23 Useful refsAshburner & Friston. VBM – the Methods. Neuroimage Jun;11(6 Pt 1):805-21Good et al. A Voxel-based morphometric study of ageing in 465 normal adult human brains. Neuroimage Jul;14(1 Pt 1):21-36