A sparse coding approach for the efficient representation and segmentation of white matter fibers

A sparse coding method is proposed for the representation and segmentation of multi-subject white matter fiber tracts. Instead of representing bundles as a single centroid, this method learns a compact dictionary of training fibers, capable of describing the whole dataset, and encodes bundles as a sparse combination of dictionary prototypes. This provides an efficient and accurate way to segment new fiber data, without explicitly embedding fibers. A strategy based on the Nyström method is used to approximate the pairwise similarities of training fibers. Experiments using dMRI data from the Human Connectome Project show the ability of our method to identify white matter bundles across subjects, and illustrates the impact of sparsity on the performance of the proposed method.