The reliable and accurate segmentation of anatomical volumes (normal or pathological) from imaging systems remains an important goal in the medical information processing because it is the first step in the chain of analysis leading to perform diagnosis to the study of internal structures morphology, for detection and quantification of lesions. Segmentation is a difficult step, especially for MRI images, for which it is primarily performed manually when there are high precision constraints. In this context, the objective of this PhD is the filtering of MRI bony joint of the human body, to achieve efficient and accurate segmentation. The first contribution is the development of a robust iterative 3D anisotropic filtering. This method can diffuse the voxels intensity according to their belonging to two populations: high diffusion for voxels located inside homogeneous regions in order to reduce noise while preventing the diffusion through the transition areas (surfaces). A rigorous comparison of several stopping functions was performed to insert the most relevant one in the iterative model. The threshold parameter splitting these two populations was estimated in an original and effective way in relation to the state of the art. The original diffusion scheme was confronted to the front propagation methods. The identification of these two models, currently separated in the literature, helped to merging the two approaches into a comprehensive anisotropic diffusion scheme governed by a front propagation. This formulation has helped offset the disadvantages of each approach namely i) the difficulty of stopping the iterative diffusion scheme, ii) the difficulty of implementing the level-sets approach facing to prohibitive calculation times. The proposed iterative scheme combines four terms related to diffusion, contrast, initial data, and to the local geometry. In particular, we have shown the importance of the ''geometrical term'' which fixes the problem of the edges discontinuities after filtering. Applications of MRI data of shoulder joints, hip and knee are presented in the testing and validation chapter. The results are given using two evaluation functions concerning edges and regions. These functions demonstrate robustness and accuracy of proposed model in removing noise and preserving edges.
