Statistical shape model-based segmentation of brain MRI images.

We propose a segmentation method that automatically delineates structures contours from 3D brain MRI images using a statistical shape model. We automatically build this 3D Point Distribution Model (PDM) in applying a Minimum Description Length (MDL) annotation to a training set of shapes, obtained by registration of a 3D anatomical atlas over a set of patients brain MRIs. Delineation of any structure from a new MRI image is first initialized by such registration. Then, delineation is achieved in iterating two consecutive steps until the 3D contour reaches idempotence. The first step consists in applying an intensity model to the latest shape position so as to formulate a closer guess: our model requires far less priors than standard model in aiming at direct interpretation rather than compliance to learned contexts. The second step consists in enforcing shape constraints onto previous guess so as to remove all bias induced by artifacts or low contrast on current MRI. For this, we infer the closest shape instance from the PDM shape space using a new estimation method which accuracy is significantly improved by a huge increase in the model resolution and by a depth-search in the parameter space. The delineation results we obtained are very encouraging and show the interest of the proposed framework.