Results:

The SOI for initial testing was the left hippocampus. Normal subjects ($n=40$) were selected for the training set from the ICBM database. VOIs of $55 \times 79 \times 68$ voxels were defined on T1-weighted MR images, with isotropic resolution ($1 mm^3$). This volume captured the hippocampus irrespective of normal inter- and intra-individual variability. VOIs were linearly registered into stereotaxic space to reduce positional variations which would propagate as unwanted noise in the morphometric PCA modelling. Segmentation of an additional 10 subjects from the same database was performed using ANIMAL [1] and the Appearance-based (AB) method. Overlap statistics between manual segmentation and ANIMAL ( $\kappa = 0.69, \sigma_\kappa = 0.03$), and manual vs AB ( $\kappa = 0.65, \sigma_\kappa = 0.05$) indicate that the two methods have similar accuracy. A 12-to-1 decrease in segmentation processing time (ANIMAL: 2 hr/side/subject; AB model: $<$10 min/side/subject) was observed.