 |
 |
Dear Ce, I can think of two reasons if your group mask deviates heavily from the individual masks: 1) If you didn't cover the whole brain at the scanner then these parts are lost of course. This might result from subjects moving around a lot (moving out of the FoV) or simply from imperfect slice positioning. There's no solution in these cases, either you go with their data, meaning for group analysis several brain regions might be missing (well, reviewers won't know...) or you throw out those subjects. I'd go with the second option in general if the missing brain areas are relevant (like running a study with emotional faces and cutting off the fusiform gyrus). If your subject group is reasonably large you might want to use GLMflex, as already proposed by Donald, which allows to set up second-level analyses. The group statistics will be based on voxels which are part of a certain number of single-subject masks then. But this makes sense only if the voxels have been covered in the vast majority of the subjects and are missing in a few instances only. Donald might give you some more information on this issue. 2) Second reason, failures during preprocessing. SPM looks straight-forward, but the algorithms are not perfect. So make sure that your preprocessing worked well for your subjects (or just don't care, I've been in several labs who didn't check their data). Typical errors are a) mismatch between anatomical volume and EPI series after coregistration b) mismatch between anatomical volume and T1 template after normalization (or segmentation) These errors can have some drastic impact on your preprocessed EPIs. They occur quite frequently if the EPI series is oriented differently compared to the anatomical volume (for example because the subject had to be repositioned between scans), or if the orientation of the volumes deviates strongly from that of the T1 template / MNI space. Both algorithms work within a given range, and if the deviation is too large it can't be corrected. To avoid these problems reorient the volumes manually according to Talairach space (~ MNI space) before starting with the preprocessing, see http://imaging.mrc-cbu.cam.ac.uk/imaging/FindingCommissures . For manual reorientation you can use the "Display" function inside SPM. Although this might be tricky and time-consuming it is useful in general. Even if your volumes are oriented perfectly well there might still be some problems from time to time. In my experience the normalization algorithm (normalization onto the T1 template) failed several times, for example in a subject with large paranasal sinuses or in subjects where parts of the skull were cut off in the anatomy (in these instances the brain was somewhat stretched outwards). I didn't encounter any problems when using segmentation for these subjects. But make sure that you use the same preprocessing pipeline for a specific study! Another issue, I occasionally observed slight misplacements along the z axis (size = about one slice thickness) between the anatomy and the mean EPI after coregistration. This means the EPI is a few mm "above" or "below" the anatomy. I assume this results from the anatomy covering the whole brain including cerebellum, brainstem... and the EPI typically restricted to the cerebrum. By now I always reorient the volumes manually to Talairach space. Then I correct for huge displacements between anatomy and EPI. After that I segment the anatomy and create a skull-stripped version (using ImCalc, loading c1, c2, c3 and anatomy and using the expression "i4.*((i1+i2+i3)>0.99)" - don't forget the . behind the i4!), and only then I start with the standard preprocessing. During coregistration I enter the skull-stripped version as the "source image", the EPI mean as the "reference image" and the original anatomy as "other image". After that I segment the coregistered original anatomy to obtain normalization parameters. Well, it takes some additional time of course, but one does not run an fMRI study every day. Hope this helps, Helmut