Automatic atlas-based building of point distribution model for segmentation of anatomical structures from brain MRI

We propose a method based on a priori knowledge provided by anatomical atlases to build - almost - automatically a point distribution model (PDM) of internal brain structures. A 3D training set of shapes is constructed by registering chosen atlas over an MRI database, which is then land-marked using a method recently developed by Davies et al. This PDM global optimization process is driven by a minimum description length (MDL) principle-based objective function. Preliminary results of built PDMs are encouraging. Future work might build active appearance models (AAM) from computed PDMs to set up a segmentation method.