For each of the intensity-corrected T1, T2, and
FLAIR images (left), we generated 12 rotationally-invariant image
features at every voxel consisting of the Gaussian, magnitude of the
gradient, Laplacian, and magnitude of the Hessian (middle panels), each
at three spatial scales or levels (2, 4, and 8 times lower resolution)
(right panels for the T1-weighted image), resulting in a 39-feature
vector for each voxel.