Resolution Enhancement of Diffusion-Weighted Images via Unified x-q Space Learning

Low resolution is a major issue restricting the application of Diffusion-Weighted Imaging (DWI) in neuroscience research and clinical routine. Super-resolution provides a viable solution to enhance the resolution of DWIs at the post-acquisition stage. Existing methods for DWI super-resolution primarily rely on the information in the x-space (i.e., spatial domain), but fail in exploiting the angular relationships in q-space (i.e., diffusion wavevector domain). In this work, we propose a Unified X-Q space Learning (UXQL) framework that makes full use of x-space and q-space information. Building upon a message-passing scheme, we employ 3D residual convolutional blocks to learn correlations in x-space, while utilizing a spatial attention mechanism to achieve effective q-space learning. Additionally, the T1-w MR image is incorporated into our framework for additional information to assist DWI super-resolution. We conduct experiments on the DWIs from the widely-used Human Connectome Project (HCP). Experimental results demonstrate the effectiveness of UXQL in improving DWI super-resolution, both quantitatively and qualitatively.