I’d say that the most reliable is generally
antsBrainExtraction.sh, since it is atlas-based. In our experience with fMRIPrep, that reliability outweighs the long run time and some precision loss with non-standard cases (the most extreme examples would be brains with large lesions, or resected areas).
That said, it is true that FreeSurfer offers a really great balance between run-time and precision. If I’m not wrong, their skull-stripping method is watershed-based and it probably has some kind of atlas-based refinement. I can’t tell how it works with non-standard brains.
bet2 if I recall correctly, is kind of an active-contour that is initialized as an ellipsoid or an sphere centered in the brain and then grown. For this reason it is extremely sensitive to the parameters you use, and as you mention, to intensity normalization.
I don’t know the internals of
3dSkullStrip. It seems to implement atlas-based extraction with an extremely fast registration process. In our experience it is worth reading through the documentation and check for recommendations on parameters and intensity normalization.
However, and it should’ve been mentioned before, the most important issue that challenges brain extraction is illumination non-uniformity (INU) or “bias field”. ANTs’ skull-stripping has the N4 algorithm built-in. FreeSurfer runs skull-stripping on the image after N3 correction (an older implementation of N4), and after that, the watershed should be rather insensitive to intensity normalization. And, if you are trying
3dSkullStrip without running INU correction first that probably explains the difference.
Maybe this paper would be of your interest - https://gigascience.biomedcentral.com/articles/10.1186/s13742-016-0150-5