Dear Helmut,
I think I made the same observation than you about a decade ago: SPM's
small volume correction (as implemented in spm_VOI.m) will use the
search space _as defined by the user_, allowing to make use of the
analytical results for the intrinsic volumes for simple shapes (sphere,
box, ...) (see spm_resels.m).
With this, for fixed smoothness, the number of resels augments with the
radius of the search space sphere, as predicted by the theory. Also,
remember that the results from the random field theory take into account
the geometry of the search space so "it is sometimes more appropriate to
surround a small but highly convoluted search space by an encompassing
sphere with slightly higher volume but less surface area", see Worsley
et al, Human Brain Mapping, 1996.
If you nevertheless want to restrict your search space further, you can
explicitly create an image mask composed of the intersection of a
sphere with the analysis mask using ImCalc and use this as an input to
small volume correction.
While the resel count increases, the peak p-values stop evolving at some
point because there is a comparison between the p-value computed with
the random field theory and the p-value returned by Bonferroni (see
spm_P.m), which makes use of the number of voxels.
Best regards,
Guillaume.
On 18/05/15 18:13, H. Nebl wrote:
> Hello everyone,
>
> When setting up a VOI for small volume correction the VOI seems not be masked with the mask image from the whole-brain GLM for some of the calculations. For illustrative purpose:
>
> Whole-brain analysis: 161,479 voxels = 955.7 resels
> SVC based on the mask image: 161,479 voxels = 955.7 resels
> SVC based on a spherical VOI with radius = 100 mm: 161,173 voxels = 3,334.2 resels
> SVC based on a spherical VOI with radius = 200 mm: 161,479 voxels = 26,673.6 resels
> SVC based on a spherical VOI with radius = 1,000 mm: 161,479 voxels = 3,334,200 resels
> SVC based on a spherical VOI with radius = 2,000 mm: 161,479 voxels = 26,673,631.6 resels
>
> Thus, with increasing size of the VOI the number of voxels does not exceed that of the whole-brain analysis, but the number of resels increases, and "Expected number of clusters", FWEc, and FWE-corrected cluster p values are affected. The current radii are meaningless, but
> 1) one might indeed go with some large sphere to e.g. cover the occipital cortex
> 2) sometimes spheres are placed near the border of the mask image (frontal eye fields)
> with some parts of or even most of the sphere containing no data. This is somewhat problematic, it would be better to estimate the resel count based on the intersection of the VOI and the mask image (which seems to be done for some of the statistics anyway, e.g. the no. of voxels does not increase, nor does the FWEp).
>
> Best
>
> Helmut
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|