Supervised Segmentation of Fiber Tracts

In this work we study the problem of supervised tract segmentation from tractography data, a vectorial representation of the brain connectivity extracted from diffusion magnetic resonance images. We report a case study based on a dataset where for each tractography of three subjects the segmentation of eight major anatomical tracts was manually operated by expert neuroanatomists. Domain specific distances that encodes the dissimilarity of tracts do not allow to define a positive semi-definite kernel function. We show that a dissimilarity representation based on such distances enables the successful design of a classifier. This approach provides a robust encoding which proves to be effective using a linear classifier. Our empirical analysis shows that we obtain better tract segmentation than previously proposed methods.