Voxelwise comparisons are going to be problematic because of differences in atrophy and CSF. You can do an analysis of each group using their own appropriate template, and then compare the results in a combined template space, but you will have to carefully account for differences just from atrophy, as you have noticed. You can potentially mitigate this by using a measure of CSF/atrophy as a covariate. An ROI analysis using ROIs from transformed atlases or manually drawn regions, instead of a voxelwise analysis, may help.