Diffusion orientation histograms (Doh) for diffusion weighted image analysis

Abstract

This paper proposes a novel keypoint descriptor for Diffusion Weighted Image (DWI) analysis, the Diffusion Orientation Histogram (DOH). The DOH descriptor quantizes local diffusion gradients into histograms over spatial location and orientation, in a manner analogous to the quantization of image gradients in the widely used Histogram of Oriented Gradients (HOG) technique. Diffusion gradient symmetry allows representing half of the orientation space at double the angular resolution, leading to a compact but highly informative descriptor. Quantitative preliminary experiments evaluate descriptors for the task of automatically identifying familial links (twins, non-twin siblings) from DWI keypoint correspondences. The DOH descriptor is found to be complementary to traditional HOG descriptors computed from scalar fractional anisotropy (FA) images, where concatenated DOH and HOG descriptors result in the highest rates of correct family member identification. Twin-twin descriptor correspondences are generally more concentrated about major white matter tracts, e.g. the internal capsule, in comparison to twin/non-twin sibling correspondences.