Currently, all atlases are 3-tissue segmentations (GM, WM, CSF). It would probably be better for N4 estimation, as well as would be nice, to have the deep GM as a separate tissue class.