Sorry, I forgot to comment on the 2nd approach you mentioned: use a two-factor ANOVA model, and apply your first approach of 3dANOVA3 for patients and normal controls separately. You could do that. However, if you want to do T tests about the group differences, you would have to consider both two factors fixed since both mixed and random models would only give you the estimates of the variance components, but not the estimates of level means. And again, no matter what, this approach would not work for those statistics that use MSE since, again, with this design the sample size would be 1 (n=1) for all cells.