Accelerated Romer-EPTI using physics-driven, joint x-q attention-network regularized reconstruction for fast mesoscale diffusion MRI

Motivation: Mesoscale dMRI will enable investigation of brain's fine-scale structures. Goal(s): We aim to achieve fast mesoscale distortion-free in-vivo dMRI (<= 500-μm-iso) within clinically acceptable time. Approach: To accelerate mesoscale dMRI acquisition based on Romer-EPTI, we develop a novel physics-driven, spatial and q-space (x-q) attention-network regularized reconstruction combined with a x-q undersampling strategy to leverage the data redundancy in the joint x-q domain. Two x-q attention denoiser networks, one in image domain and the other in feature domain, are developed as plug-and-play priors. Results: We demonstrate four-fold accelerated Romer-EPTI capable of acquiring 500-μm isotropic dMRI under 20 minutes on clinical 3T scanners. Impact: The proposed physics-driven, joint spatial and q-space attention-network regularized reconstruction enables accelerated Romer-EPTI, achieving mesoscale dMRI at 500-μm isotropic resolution within 20 minutes.