Automated Sagmentation of Cerebral Structures Incorporating Explicit Knowledge

Overview & Synopsis

We propose in this thesis an automated segmentation system for medical images (most especially in vivo brain MRI). We have focused on devising a segmentation methodology where maximum use is made of the available a priori medical expertise. We approach the issue of boundary finding as a process of fitting a group of deformable templates (simplex meshes) to the contours of the target structures. Those templates evolve in parallel under the supervision of a series of rules derived from the analysis of both the template's dynamics and medical experience. We subject the templates to a variety of constraints also designed based on prior medical knowledge about the textural, shape and underlying histological properties of the target structures. Textural information is extracted by a linear/non-linear classifier in the form of a 2-stage neural network whose hybrid architecture and dynamic learning phase help produce a better classification map. An original learning approach to the dense matching of n-D objects enables the introduction of prior knowledge in the computation of shape models for the target structures. We also introduce a novel shape descriptor, the observed transport descriptor, whose noise robustness and enhanced discriminating power make it a good candidate for our matching strategy. Finally, a more faithful model of the transformations induced by histological processes, the piecewise affine approach, allows the design of a biomedical registration algorithm better suited to the reconstruction of 3-D histological volumes, a first step towards hybrid combined MRI/histology atlases.