Segmentation of anatomical structures from 3D brain MRI using automatically-built statistical shape models

We propose a twofold method that first automatically builds a statistical shape model of anatomical 3D brain structures of interest, then uses this model for delineating structure contours onto any patient MRI. First of all, an estimated training set of shapes is inferred by registration of a 3D anatomical atlas over a set of brain MRIs, then automatically landmarked using the "Minimum Description Length" based method developed by Davies et al., (2002). A 3D "Point Distribution Model" is then established and used to constrain the delineation process. It is lead by a novel intensity model specifically developed to overcome the estimated nature of our training set in exploiting only local intensities.