In this paper, we introduce a novel approach for segmenting MRI images into the three classes of tissue of the musculoskeletal system : bone, fat and muscle. This approach is guided by a prior anatomical knowledge modeled using a tree structure. This tree aims at representing the natural nested topology of the musculoskeletal anatomy, and is used to hierarchically segment images. At each level of the hierarchy, a standard two-classes classification process is performed using texture-based descriptors and support vector machines (SVM). The classification is refined using topological constraints (connexity and neighborhood) derived from anatomy. We evaluate the performance of our approach by comparing the constrained approach with the original hierarchical algorithm. We achieve an excellent classification(78%) and shows that the use of texture analysis combined with simple topological constraints can improve the segmentation.