Written and maintained by CASRAI Editorial Board
Published
On this page
Who this is for: researchers who have T1-weighted structural MRI scans from two or more groups (or a scan set plus a clinical score) and need to run a voxel-based morphometry analysis, then read the result. This page walks through the documented workflow in two research packages, FSL-VBM and the CAT toolbox for SPM, and explains the choices (modulation, total intracranial volume, smoothing) that decide whether the result means what you think it means.
What voxel-based morphometry is used for in research
Voxel-based morphometry (VBM) is a whole-brain method for finding voxel-wise differences in local grey matter volume or topography, either between populations or in relation to a clinical or behavioural score. Because it tests every voxel rather than a pre-defined region, it needs no a priori hypothesis about where a difference lies. The FSL documentation describes its FSL-VBM pipeline as following the optimised VBM protocol of Good et al. (2001): structural images are segmented into grey matter, registered non-linearly to a standard space, and then “modulated” so that the signal is corrected for local expansion or contraction caused by the registration. Statistics are then run voxel by voxel on the resulting grey matter maps.
VBM is a structural technique. It does not use functional (task or resting-state) data; for those, see the sibling guides on Statistical Parametric Mapping (SPM12) and FSL (FMRIB Software Library). Surface-based thickness and subcortical-volume analysis is a different approach, covered in the FreeSurfer guide.
Access: install from the official source
- FSL-VBM is part of FSL. Install FSL from the FMRIB site (fsl.fmrib.ox.ac.uk); the
fslvbm_*scripts andrandomisecome with it. Documentation: FSL-VBM user guide. - CAT (Computational Anatomy Toolbox) runs inside SPM. Its manual (neuro-jena.github.io/cat12-help) covers the current CAT release, and all menu paths below are from that manual. Check each project’s own site for licence terms.
Step by step, option A: FSL-VBM
Commands below are as given in the FSL-VBM user guide (read 2026-11-07). Run them from a single working directory.
- Prepare the folder. Create a directory and copy every subject’s T1 image into it with a group prefix in the file name, for example
CON_2304.nii.gzandPAT_2042.nii.gz. - Choose the template subjects. Write one image name per line in a file called
template_list. Select an equal number of subjects from each group; the documented example uses five controls and five of the seven patients, so the template is not biased toward the larger group. - Check the inputs. Run
slicesdir `imglob *`and inspect the montage for bad images. - Write the design. Create
design.matanddesign.con(see the randomise manual). The row order indesign.matmust match the order printed byimglob *, because the images are concatenated alphabetically. - Brain-extract.
fslvbm_1_bet -b, orfslvbm_1_bet -Nif the images include a lot of neck. This moves the images intostruc/with_strucappended and creates the*_brainfiles. - Build the template.
fslvbm_2_template -a(affine registration) orfslvbm_2_template -n(non-linear). Checktemplate_GM_4Dinstruc/. - Register, modulate and merge.
fslvbm_3_proc. This creates thestats/directory withGM_merg,GM_mod_mergand the smoothedGM_mod_merg_s2,_s3and_s4(Gaussian sigma of 2, 3 and 4 mm), plusGM_mask. It also runs an initial randomise pass that writes raw t-statistic maps. - Choose the smoothing. The guide leaves the sigma for you to decide and report. For another value, run
fslmaths GM_mod_merg -s 3.5 GM_mod_merg_s3.5. - Run inference with permutation testing. The recommended threshold-free cluster enhancement (TFCE) call is:
randomise -i GM_mod_merg_s3 -m GM_mask -o fslvbm -d design.mat -t design.con -T -n 5000
The cluster-based alternative replaces-Twith-c 2.3.
Step by step, option B: CAT inside SPM
Menu names are from the CAT manual (read 2026-11-07). Start SPM, then open the CAT toolbox.
- Segment. CAT > Preprocessing > Segment, one T1 image per subject. For VBM you want the grey matter output; the modulated, normalised grey matter file is named with the prefix
mwp1. The default modulation setting, “Modulated normalized”, gives regional volume, which is why total intracranial volume (TIV) must be corrected for later. - Check image quality. CAT > Data Quality > Single Slice Display on the
wm*files for a quick look at each segmentation. - Get TIV. CAT > Statistical Analysis > Get TIV, saving “TIV only”. You will use these values as a covariate.
- Check sample homogeneity. CAT > Data Quality > Check Sample Homogeneity on the unsmoothed
mwp1images. It reports a quartic mean Z-score per subject. The manual states that a high score alone is not a reason to exclude a scan; use the “Check worst” option to look at the flagged images before deciding. - Smooth. Use SPM’s Smooth module. The manual suggests starting at an FWHM of 6 mm, with 6-8 mm common, entered as
6 6 6, with prefixs, which gives thesmwp1files. - Specify the model. CAT > Statistical Analysis > Basic Models: choose a two-sample t-test, multiple regression or full factorial design; set TIV correction (ANCOVA, global scaling or none) and an absolute threshold mask of about 0.1. The manual advises global scaling if TIV correlates with the parameter of interest.
- Estimate and view. Run SPM’s Estimate, then Results, and define the contrasts you need.
Output and how to read it
FSL-VBM outputs
| File | Created by | What it is |
|---|---|---|
template_GM (with template_GM_init, template_4D_GM) |
fslvbm_2_template |
The study-specific grey matter template |
GM_merg, GM_mask, GM_mod_merg, GM_mod_merg_s2/s3/s4 |
fslvbm_3_proc |
Merged grey matter data, analysis mask, modulated data, and the smoothed versions |
fslvbm_tstat1 |
randomise (-o fslvbm) |
Raw t-statistic map for contrast 1 |
fslvbm_tfce_corrp_tstat1 |
randomise with -T |
Corrected p-value image for contrast 1 (TFCE) |
fslvbm_clustere_corrp_tstat1 |
randomise with -c |
Corrected p-value image for contrast 1 (cluster extent) |
The guide views the TFCE result in FSLeyes with a display range of 0.949 to 1, and thresholds the cluster result at 0.95 to build mask_pcorrected, which it then applies to the t-statistic map with fslmaths. That reflects how randomise stores corrected results: the image holds values near 1 where the corrected p-value is small, so 0.95 corresponds to p < 0.05. Contrast numbers follow the order of the rows in design.con.
CAT / SPM outputs
Results come from SPM’s standard Results interface for each contrast you defined. Because the input images are modulated, a voxel-level effect is a difference in regional grey matter volume once TIV has been handled by the option you chose.
Smoothing: sigma versus FWHM
The two workflows state their kernels differently. FSL gives Gaussian sigma (2, 3 or 4 mm in the standard outputs); the CAT manual gives FWHM (start at 6 mm). The two are related by FWHM = 2 × √(2 ln 2) × sigma, about 2.355 × sigma, so FSL’s sigma of 2, 3 and 4 mm correspond to roughly 4.7, 7.1 and 9.4 mm FWHM. When you write the methods section, report the kernel in the units your software used and state which one it is.
Troubleshooting
| Symptom | What the documentation says |
|---|---|
| Group labels are swapped or contrasts look wrong in FSL-VBM | The images are concatenated in alphabetical order, so the rows of design.mat must follow the output of imglob *. |
| Poor brain extraction on scans with a long neck | Use fslvbm_1_bet -N instead of -b. |
| Template looks biased toward one group | Build template_list with an equal number of subjects from each group. |
| A subject has a high homogeneity score in CAT | Do not exclude on the score alone; inspect with “Check worst” first. |
| TIV correlates with your variable of interest (CAT) | The manual advises global scaling for TIV correction in that case. |
| You want a different smoothing in FSL-VBM | fslmaths GM_mod_merg -s 3.5 GM_mod_merg_s3.5 (any sigma). |
Tools that run VBM
| Tool | How VBM is run | Inference |
|---|---|---|
| FSL-VBM | Command-line scripts fslvbm_1_bet, _2_template, _3_proc |
Permutation testing with randomise (TFCE or cluster) |
| CAT (Computational Anatomy Toolbox) | Menus inside SPM: Segment, Get TIV, Check Sample Homogeneity, Basic Models | SPM’s Estimate and Results |
See the SPM12 guide for the host environment CAT runs in and the FSL guide for the wider FSL toolset.
FAQ
What does VBM measure?
Voxel-wise differences in local grey matter volume or topography between groups, or in relation to a score, tested across the whole brain without a pre-specified region (FSL-VBM user guide).
Why does total intracranial volume matter?
With modulation, the images express regional volume, which scales with head size. The CAT manual therefore has you compute TIV and correct for it in the model (ANCOVA or global scaling).
Which smoothing kernel should I use?
There is no single required value. FSL provides sigma 2, 3 and 4 mm and leaves the choice to you; CAT suggests starting at 6 mm FWHM with 6-8 mm common. Choose before looking at results and report it.
Should I use TFCE or cluster-based inference in FSL-VBM?
The FSL-VBM guide recommends TFCE and gives the cluster-based call as an alternative.
Sources
- FMRIB Analysis Group, FSL-VBM user guide, read 2026-11-07.
- CAT toolbox documentation, CAT manual, read 2026-11-07.







