Dear Adriana,
Yes, you can create masks to restrict the analysis to GM/WM voxels. You could segment the individual T1 volumes (if there are any, otherwise you will have to turn to the PET data) and then create individual masks based on the c1 (= GM) and c2 (= WM) volumes. One option is to go with Imcalc, you might start with something like (i1 + i2) > .75 and then test different thresholds. For the group mask you would combine the different individual masks, something like (i1+i2+...)/n > n to restrict the analysis to voxels which have (non-CSF) data in all of your n subjects. However, the masking will also affect voxels on the cortical surface (due to the CSF nearby).
I'm not really sure whether this is a good approach in your case though. It is basically hiding something you're not happy with. If you have some unexpected findings for the ventricles, how to make sure whether the rest of the brain is meaningful? Some of the data sets might be corrupted entirely, not just the ventricles (in fMRI you might detect "ventricle activation" due to severe head motion), there might be some physiological effects affecting oxygen supply, ...
Thus I would suggest to check your data for outliers, and also the settings for global normalisation/scaling during model specification. It could also be due to bad normalisation (larger ventricles/higher atrophy in patients relative to controls, which might result in artificial metabolism differences due to CSF voxels in patients and GM voxels in controls), see e.g. Reig et al. (2007, Neuroimage). Maybe you have to switch to different templates like those recently described in Della Rosa et al. (2014, Neuroinformatics).
If you still detect some unexpected findings, maybe extract the time course in the ventricles (averaged across corresponding voxels) and add this as a regressor. You would thus control for the signal changes on whole brain level (as other voxels might be affected as well, although maybe to a lesser extent).
Hope this helps a little,
Helmut
|