Adaptive graph cuts with tissue priors for brain MRI segmentation

We describe a novel framework for automatic brain MRI tissue segmentation. To overcome inherent difficulties associated with this particular segmentation problem, we use a graph cut/atlas-based registration methodology optimized within an iterative mode. The basic graph cut algorithm guarantees a global or near-global minimum of an energy function associated with a Markov random field (MRF). For atlas-based graph cuts, we tailor this energy function to incorporate both a priori information derived from registered brain atlases as well as region and boundary information derived directly from the images. The iterative algorithm adaptively alternates segmentation and inhomogeneity correction. The proposed method can be extended to multispectral image segmentation. We validate our method in both simulated adult and real neonatal brain MR images corrupted by significant noise and intensity inhomogeneities