Fibre population-guided pre-training for 3D spatial super-resolution on multimodal brain diffusion MR imaging

Diffusion-weighted imaging (DWI) is a non-invasive imaging technique to analyze the diffusivity in the human brain to provide insights of the underlying tissue microstructure. Multimodal derivatives of DWI, including the estimation of fibre orientation distribution (FOD), whole-brain tractography, and connectomics, can measure the global changes in brain connectivity quantitatively and reveal the connectivity between brain regions. However, the reliability of these downstream modalities is heavily reliant upon the quality of the original DWI. Low spatial resolution is a common issue when acquiring DWIs for clinical use due to the restrictions on the scanning protocols, which leads to significant differences between the connectomics constructed by high and low spatial resolution DWIs. To this end, we propose a 3D multimodal Fixel-guided Spatial Super-Resolution network (FSSR-Net) for FOD to mitigate differences introduced by the FODs generated by low spatial resolution DWIs, where fixel refers to fibre population within a voxel. FSSR-Net integrates a fixel-guided pre-training to better encode the FOD representation in complex fibre scenarios, and an SSR-Net for the main super-resolution task. In addition to FSSR-Net, we present a comprehensive set of evaluation metrics for multimodal FOD derivatives, which can serve as benchmarks when comparing DWI-related derivatives. The experimental results show that our FSSR-Net not only improves the quality of FOD but also mitigates the faulty disparities in multimodal FOD derivatives when compared to the ground truth.