Hybrid Graph Convolutional Neural Networks for Super Resolution of DW Images

Abstract

Deep learning based super-resolution has been proved as a promising technique in increasing the resolution of Diffusion-weighted (DW) images, which is widely used in brain white matter analyses. Existing models using single type of convolutional neural networks (CNN) cannot support effective learning in heterogeneous data space composed of grid structure and wavevector domain. We propose a novel technique that employs a graph model in a deep network to improve the spatial resolution of DW image. The model is composed of a residual CNN to learn from spatial information in 3D grid structural domain and graph CNN (GCNN) to emphasize diffusion angular information in non-Euclidean domain. Given a low resolution DW image, for each direction of diffusion gradients, 3D convolutions and residual CNN are firstly performed to generate coarse-level super resolution image. Then the learning outputs from grid structure space are stacked and refined by the features in diffusion gradient space modelled by GCNN. We evaluate the proposed hybrid graph CNN model using real brain data from Human Connectome Project. Extensive experiments demonstrate improved results with richer fiber tracts that are closest to the ground truth. Our hybrid graph CNN model benefits the learning of spatial and angular features in complex or heterogeneous spaces of DWI.