In this thesis, we are interested in using deformable surface models for 3D and 4D image segmentation. Our contribution is twofold. First, we have constrained the class of model deformation to improve the segmentation robustness to noise and outliers. We have used the simplex mesh representation to define local regularizing constraints. We have developped an evolutive deformation process combining a global transformation with few degrees of freedom and a local deformation field. It allows to control the amount of admissible deformation of a deformable model in a simple and efficient manner. We have also introduced an a priori knowledge on the data by using shape constrained deformations. This makes the 3D reconstruction process more robust especially in area where image data are noisy or lacking. In addition to studying the theoritical convergence of the deformation scheme, we have developed algorithms allowing automatic topological changes comparable to the level-set method. Second, we have investigated different strategies for computing the external force for various 3D image types. We have studied different medical image geometries on which a deformable model can be deformed. For instance, we have defined region based and intensity-profile based external forces for monomodal segmentation and multimodal image registration. Finally, we have extended the framework of deformable modelling to include time serie of images (4D images) for the segmentation of 2D and 3D cardiac image sequences. The introduction of time continuity allows to introduce new constraints in the deformation process. We illustrate our method results by segmenting heart images or image sequences acquired using different imaging modalities.
