Segmentation of magnetic resonance images (MRI) of skeletal striated muscles is of crucial interest when studying myopathies. Diseases understanding, therapeutic followups of patients, etc. rely on discriminating the muscles in MRI anatomical images. However, delineating the muscle contours manually is an extremely long and tedious task, and thus often a bottleneck in clinical research. Typical automatic segmentation methods rely on finding discriminative visual properties between objects of interest, accurate contour detection or clinically interesting anatomical points. Skeletal muscles show none of these features in MRI, making automatic segmentation a challenging problem. In spite of recent advances on segmentation methods, their application in clinical settings is difficult, and most of the times, manual segmentation and correction is still the only option. In this thesis, we propose several approaches for segmenting skeletal muscles automatically in MRI, all related to the popular graph-based Random Walker (RW) segmentation algorithm. The strength of the RW method relies on its robustness in the case of weak contours and its fast and global optimization. Originally, the RW algorithm was developed for interactive segmentation: the user had to pre-segment small regions of the image – called seeds – before running the algorithm which would then complete the segmentation. Our first contribution is a method for automatically generating and labeling all the appropriate seeds, based on a Markov Random Fields formulation integrating prior knowledge of the relative positions, and prior detection of contours between pairs of seeds. A second contribution amounts to incorporating prior knowledge of the shape directly into the RW framework. Such formulation retains the probabilistic interpretation of the RW algorithm and thus allows to compute the segmentation by solving a large but simple sparse linear system, like in the original method. In a third contribution, we propose to develop a learning framework to estimate the optimal set of parameters for balancing the contrast term of the RW algorithm and the different existing prior models. The main challenge we face is that the training samples are not fully supervised. Specifically, they provide a hard segmentation of the medical images, instead of the optimal probabilistic segmentation, which corresponds to the desired output of the RW algorithm. We overcome this challenge by treating the optimal probabilistic segmentation as a latent variable. This allows us to employ the latent Support Vector Machine (latent SVM) formulation for parameter estimation. All proposed methods are tested and validated on real clinical datasets of MRI volumes of lower limbs.
