Self-Supervised Denoising of Diffusion MRI Data Via Spatio-Angular Noise2Noise

Diffusion MRI (DMRI) suffers from heavy noise that reduces the accuracy and reliability of the derived diffusion metrics. Existing Deep Learning (DL) methods for DMRI denoising usually rely on training with paired noisy-clean data, which are unavailable in a clinical setting. To this end, we propose a self-supervised DL denoising method, called Spatio-Angular Noise2Noise (SAN2N). We utilize a network trained with paired noisy data that can capture the essential information of underlying clean data for noise reduction. Specifically, SAN2N generates angular neighboring DMRI data based on the geometric structure of q-space sampling points. The resulting data and the original one are then fed to two neighborhood-based x-space sub-samplers to extract 4D similar patches in the spatio-angular domain. Finally, these patches are employed to train our SAN2N with a regularized denoising loss. Extensive experiments on simulated and real datasets demonstrate the superiority of SAN2N over existing DMRI denoising methods.