Hello fsl users, I am replicating a published work where a group of healthy controls (CT head 3D volumes) is compared to individual subject-cases in a voxelwise fashion. All the controls are registered and smoothed to MNI space. To check if a subject-case contains any lesion (blood which is of high intensity) as opposed to the healthy controls, the subject-case is also registered to the MNI space and smoothed. Then the intensity differences (voxel-wise) of the subject-case with the controls group is calculated as a t-test (Crawford-Howell) t-test = subject –mean(controls) / stdev(controls) *sqrt((n + 1) / n), with n= number of controls. The two tailed p_values are given by p-val = tails*(1 - t_distr_cdf( abs(t-test) ) ) , with tails =2 So a 3D t-test map (t-test-map.nii.gz) and a p-val map (p-val-map.nii.gz) are created. To correct for the multiple comparisons that paper uses a Bonferroni corrected threshold p<0.05. I am puzzled if I could use a bonferroni correction in this scheme and I have the following questions: If I use Bonferroni correction I need to threshold at 0.05/number-of-meaningful-voxels. The number of meaningful voxels (not background) are about 5.000.000. With such a number of voxels isn’t it that the threshold will be so conservative that is going to be totally useless? Even if the amount of meaningful voxels were very small, I have the impression that I still cannot use the bonferroni correction since the data are smoothed and registered. The voxels of smoothed and/or registered images are correlated and the assumption of independent tests for bonferroni doesn’t hold. What do you think? Would you recommend FDR, cluster based inference , randomise? If Bonferroni is not an option then could I use FDR? fdr –i p-val-map.nii.gz –q 0.05 If applicable to my data scheme, I would like to try also the GRF for voxel-wise tests and cluster based inference as described in the lecture slides of fsl but I couldn’t figure out how to do it for my data with the fsl programs. If I use: smoothest -z t-test.nii.gz -m brain_mask.nii.gz # get the smoothness_estimate cluster –I t-test.nii.gz –t arbitrary_thresh –p 0.05 –d smoothness_estimate –volume=brain_mask_voxels_num > cluster_t_stat.txt #calculate cluster t_stat is this a correct way for cluster based inference ? How could I go for GRF voxel-wise tests? Which fsl utility to use and how? Finally is it possible to use randomise and permutations on my scheme? If yes how could I do that? Thank you in advance Antonios Konstantinos