The treatment of musculoskeletal disorders (MSD) is of paramount importance, as MSDs are chronic pathologies accounting for the largest fraction of temporary and permanent joint disabilities in industrialized societies. Accurate modeling of musculoskeletal structures is crucial for diagnosis and pre-operative planning and usually relies on Magnetic Resonance Imaging (MRI) for its ability to simultaneously image soft and bony structures composing the articulations. However, MRI images present, in addition to ubiquitous image artifacts, highly heterogeneous bone tissue intensities and articular structures with fuzzy boundaries. Furthermore, clinical MRI images are often acquired with limited field of view (FOV) or low image resolution, seriously compromising its use in automated image processing. In this thesis, we focus on the design of efficient image segmentation methods to create subject-specific musculoskeletal models from MRI data. We propose a novel prior knowledge construction for use in discrete deformable models to yield an effective modeling of musculoskeletal structures.