 |
 |
> I am doing VBM on two study groups. For each subject I have gray and white > matter segments, members of each group having been segmented from fully > normalized raw images (optimized VBM) using a template specific to their > respective groups (I am hoping to get the best possible segmentation this > way, by assuming that the groups have distinct gray/white matter > distributions). If you plan to campare one group with another, then I would suggest treating all the data in the same way - i.e. using the same tissue probability maps as priors for all data. > I would like to normalize these segments to an average > template I have created from all subjects in the study, so as not to create > a bias toward one group or the other in my analysis. I didn't expect > normalization to be a problem as I can invert the deformations performed on > each subject and normalize everyone to the average template. However, I > have read a posting from 9 Feb 2004 indicating that resampling using a > logit transform would be appropriate if I am going to warp probability > images. While I understand what a logit transform is, I am not clear on how > to apply this transformation to my data. In particular when I have > experimented with using Imcalc on one image, the voxels with probability = > 1 cause division by zero and an error. Has anyone used this method? Thanks > for your help. Here is that email for reference... http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind04&L=spm&O=D&F=&S=&X=0616F44B877C1A18C3&Y=spm%40fil.ion.ucl.ac.uk&P=41935 The use of the logit transform is a little speculative, but you could apply the logit transform to the data, do the resampling and then do an inverse logit transform at the end. Before doing the logit transform, you'd probably need to add a small value to the images and rescale them slightly - as a logit transform of zero is -Inf, and a transform of one is +Inf. Maybe someone else can comment on the best way of interpolating images such that the interpolated values fall between zero and one. I speculated a little bit more about this in the recent NeuroImage paper on "Unified Segmentation". Best regards, -John