Segmentation of anatomical structure from DT-MRI

A general framework for the automatic segmentation of anatomical structures from diffusion tensor MRI is presented here. We adopt an energy based approach to segmentation assuming a piecewise-smooth image model that allows tensors to change orientation inside bundles, complemented by adequate modeling of image statistics. Energy minimization is carried out using a greedy region-growing algorithm that is both efficient and robust. Although the framework is general and any tensor metric is supported, we use a simplified tensor representation that adapts well to the DTI setting and further improves computational performance. Segmentation results are generated automatically from a single seed point as illustrated on several real and synthetic datasets