We have a few ppts for which pydeface.py didn’t work quite right (the registration did weird things and masking got out of hand). In those cases, we did a rough skullstrip and dilated the brain mask by 2cm and that worked.
I’m wondering, then: since that approach certainly isn’t more complicated than the whole “transform, mask, transform” one, is there a reason it isn’t used more commonly?
Registration is reasonably robust to a lot of variance in structural data acquisitions, which makes it a pretty reliable approach, if you want something you can run a lot of subjects through and not think too much about. Comparatively speaking, skull-stripping can very easily strip far too much, though perhaps a 2cm dilation would compensate for that possibility.
Some part of the decision may rest on how much your later processing steps depend on extra-cerebral features. It may be considered safer to try to remove as few voxels as possible while destroying personally-identifying information.
If 2cm is enough to leave skull intact (for future precise skull stripping and MEG skull surface reconstruction) I don’t see a problem (as long as the face is removed).