Voxel-based morphometry studies are frequently cited as having the advantage of being objective compared to region-of-interest methods. This statement assumes, however, that all regions are treated equally both in controls and diseased cohorts. This study aimed to test whether this statement is correct by analyzing fiducial landmarks in controls, Alzheimer's disease (as a model of mild generalized atrophy model); Frontotemporal Dementia (focal atrophy model) and Semantic Dementia (extreme focal atrophy model). Standard SPM5 and DARTEL were evaluated using either raw or skull-stripped/bias corrected scans. The results indicated that with all methods there was variability in the degree of misregistration across regions and that there was a disease grouping interaction-most severely in the extreme focal atrophy model (Semantic Dementia). Preprocessing improved VBM outputs both with standard SPM and DARTEL. In the latter case, this occurred to an extreme degree-DARTEL using raw data was grossly insensitive to a ground truth (manually verified hippocampal atrophy in AD) whereas DARTEL after preprocessing yielded excellent results with respect to this yardstick.