Hi there,
I have a question about the difference between statistical analysis in SPM versus FSL.
I had previously carried out all the optimized VBM processing in SPM2, but as I was interested in applying permutation analysis to correct for multiple comparisons, I thought I could just analyse my SPM normalized and smoothed images in FSL - instead of going through the whole process again from scratch in FSL.
Here’s what I did: I took the normalized, modulated and smoothed images from SPM, converted them to nii.gz and created a 4Dimage. I specified a straight forward GLM to determine the effect of ‘group’ (patient or control) on intensity at each voxel. I ran randomise (t=2, n=10,000). I was surprised to see much larger cluster from permutation analysis than what I had seen using FDR in SPM2 (widespread significance compared to 3 or 4 small clusters).
When I viewed the uncorrected p-value maps, FSL-VBM's appears to be much more significant (and smoother) than in SPM (p<0.05). Is there an obvious reason why this might be?
I also applied the same FDR correction (q=0.05) to the uncorrected p-values in FSL (as described - www.fmrib.ox.ac.uk/fsl/randomise/fdr.html) as I had done in SPM. The result was now much more in keeping with the permutation findings (and also much much larger than the SPM FDR corrected findings).
Thanks,
Cathy
|