Automatic skeletal muscle segmentation through random walks and graph-based seed placement

In this paper we propose a novel skeletal muscle segmentation method driven from discrete optimization. We introduce a graphical model that is able to automatically determine appropriate seed positions with respect to the different muscle classes. This is achieved by taking into account the expected local visual and geometric properties of the seeds through a pair-wise Markov Random Field. The outcome of this optimization process is fed to a powerful graph-based diffusion segmentation method (random walker) that is able to produce very promising results through a fully automated approach. Validation on challenging data sets demonstrates the potentials of our method.