Regarding the swiss cheese part, that will be problematic possibly for the nonlinear alignment. You might start with just an affine alignment to a template to start, where I would expect the "holiness" of the cerebellum not to have a big enough influence on the optimum, single affine matrix transformation for the whole volume. Another approach is to limit the nonlinear alignment to just the initial stages with -maxlev. Alternatively mask the cerebellum off before the alignment or make it an exclusion mask in the nonlinear warping step.