terms of image intensity, their spatial location relative to one another is quite consistent: amygdala is always in front of and above hippocampus, never below or behind it. To encode this, we used a Markov Random Field (MRF) model, which had been previously used extensively in computer vision (Geman and Geman, 1984). Typical MRF modeling at the time was stationary (e.g. the same probabilities for voxel a being label l1 given voxel b being label l2 regardless of spatial location), and isotropic (e.g. the spatial relationship between voxels a and b was not considered other than whether they were neighbors or not). In order to constrain the segmentations to be neuroanatomically plausible, we extended the MRF model to be spatially nonstationary, so that the probabilities were allowed to vary over space, and anisotropic, so that we could model the probability that hippocampus was below amygdala separately from the probability that it was above it. These more sophisticated models allowed us to utilize a training set of manually labeled images to bootstrap a procedure for whole-brain segmentation that was as accurate as extensively trained manual raters were capable of creating, as well as being robust to pathology (Fischl et al., 2002,2004a).
