Manifold-driven Grouping of Skeletal Muscle Fibers

In this report, we present a manifold clustering method for the classification of fibers obtained from diffusion tensor images (DTI) of the human skeletal muscle. To this end, we propose the use of angular Hilbertian metrics between multivariate normal distributions to define a family of distances between tensors that we generalize to fibers. The obtained metrics between fiber tracts encompasses both diffusion and localization information. As far as clustering is concerned, we use two methods. The first approach is based on diffusion maps and k-means clustering in the spectral embedding space. The second approach uses a linear programming formulation of prototype-based clustering. This formulation allows for classification over manifolds without the necessity to embed the data in low dimensional spaces and determines automatically the number of clusters. The experimental validation of the proposed framework is done using a manually annotated significant dataset of DTI of the calf muscle for healthy and diseased subjects