Simon Jeremy Damion Prince1, Daniel C. Alexander1
1Computer Science, University College London, London, UK
We aim to improve the spatial resolution of fibre-orientation estimates in diffusion MRI by learning a prior over 3x3x3 voxel patches that assumes self-similarity across scale. In training we align and cluster patches of fibre orientations from a larger scale using a mixture of Watsons model. The resulting canonical fibre configurations describe homogenous regions, bending, fanning etc. To super-resolve, we find the canonical patch and mean orientation, diffusivity and volume fraction that best describe the voxel measurements. We compare our results to nearest-neighbour interpolation and demonstrate that it is possible to successfully determine sub-voxel structure.