I'd guess that the behaviour of the tissue segmentation when applied to EPI data depends on the image contrast, artefacts, etc, so may not always work well.
I haven't seen so many coreg failures with data from our department, so haven't built up so much expertise on what is the best thing to try. My guess is that when skull stripping helps, it helps more when applied to the structural (because of all the non-brain signal from the scalp, face, neck, etc).
All the best,
John

"H. Nebl" <[log in to unmask]> wrote:
>Dear John,
>
>I already rely on bias-corrected means and structures in current projects, as accounting for intensity inhomogenities in an intensity-based registration seemed to be a good idea to me ;-) , and the results are somewhat more promising, although the differences are usually not huge   but consistent (mainly driven by a different rotation in yz plane).
>
>I have doubts on the skull-stripping though, as for consistency reasons one should probably also remove eye-balls and other non-brain tissue visible in the EPI volume. I've tried to do so with BET for both structure and EPI, but this is never really good. Segment on the other hand works reasonably well for structures (e.g. something like masking with c1+c2+c3 > 0), but when trying to construct a similar mask based on the EPI segmentation files the boundary voxels are lost, possibly meaning one would have to go with c1+c2+c3+c4. I've also tried to go with a mask based on c1+c2+c3 > 0 from EPI segmentation, then apply some smoothing, threshold again, apply the mask to the EPI, thus mimicking some "growing", which leads to quite nice results. But then again, I would like to keep this as similar as possible for EPI and structure. Probably one would need extra tissue probability maps when segmenting EPI data, or a different number of Gaussians, no idea.
>
>Best
>
>Helmut