Automated segmentation and analysis of human cerebral cortex imagery

Surface models of the cerebral cortex are useful for making atlases of the human brain, for performing morphometric analysis of the brain, for visualizing brain activity, and for source modeling in magnetoencephalography (MEG) and electroencephalography (EEG). The cerebral cortex is a single sheet of grey matter enclosing the telencephalon. If we close this surface at the brainstem, we can describe it as a deformed sphere. After this closing, surface representations of the cerebral cortex should be topologically equivalent to a sphere. We describe a new sequence of image analysis methods that produces cortical surface representations with spherical topology from magnetic resonance (MR) images of the human brain. Our sequence consists of four stages. The first stage removes skull and other non-brain tissues from the MR volume using a combination of anisotropic diffusion filtering, Marr-Hildreth edge detection, and morphological operators. The second stage corrects for image nonuniformity by computing local tissue intensity properties and relating them to global properties. A cubic B-spline representation of image nonuniformity is then generated and this estimate is removed from the volume. The corrected volume is labeled by a Maximum A Posteriori classifier that models spatial tissue properties of the human brain with a Gibbs prior. Finally, we apply a topological constraint to the region bounded by the white/grey matter interface. This constraint ensures that a resulting tessellation is homeomorphic to a sphere. We validate our methods on real and phantom MR data. We also describe BrainSuite, our new software package for MR image analysis and visualization that incorporates the methods described in this dissertation.