Brain segmentation from 3D MRI using statistically learned physics-based deformable models

The authors introduce a statistical deformable model for the segmentation of the brain structure in 3D MRI. Their approach relies on a physically deformable multimodel that embeds information on head (skull and scalp) and brain by parameterizing these structures by the amplitudes of vibration of an initial spherical mesh. The spatial relation between head and brain is then statistically learned through an off-line training procedure using a representative population of 3D MRI. In order to segment the brain from a MR image not belonging to the training set, the authors first segment the head surface. The brain contour coordinates are then iteratively recovered using their statistical relations to the head coordinates.