This Ph.D thesis presents new paradigms in the field of segmentation by mean of deformable models. Classic model-based segmentation approach consists of iteratively deform complex models in the space of the scanner slices until the organ to segment is fitted. The proposed works tries to improve the process by decomposing the problem in three distinct phases : First, anatomical landmarks of the organ to segment are automatically detected using a fast ray-casting technique ; A basic deformable model is next initialized to fit these landmarks. It allows approximately fitting the organ contour. This raw approximation can optionally be refined by a conventional registration ; Finally, an active contour method (snakes) allows to accurately delimiting the contour of the organ. Although contour delimitation by classic snake resolution usually converges to a local minimum, this problem is avoided by using dynamic programming. A main advantage of our approach is to propose a good compromise between segmentation speed and quality of the result. It also allows coming back to an interactive segmentation of the organ anytime. CT femoral segmentation of degenerative hips has been used to validate our approach. Good results were achieved on most of the patients. Only highly degenerated femurs with osteoporosis or arthritis were not fully automatically segmented. Thanks to the goods results achieved, this new segmentation paradigm could practically be used in prosthesis planning software like femur resurfacing.