Group Sparse Kernelized Dictionary Learning for the Clustering of White Matter Fibers

This paper presents a novel method that combines kernelized dictionary learning and group sparsity to efficiently cluster white matter fiber tracts obtained from diffusion Magnetic Resonance Imaging (dMRI). Instead of having an explicit feature representation for the fibers, this method uses a non-linear kernel and specialized distance measures that can better learn complex bundles. Through the use of a global sparsity prior, the method also provides a soft assignment of fibers to bundles, making it more robust to overlapping fiber bundles and outliers. Furthermore, by using a group sparsity prior, it can automatically discard small and uninteresting bundles. We evaluate our method both qualitatively and quantitatively using expert labeled data, and compare it with state of the art approaches for this task.