Self-gated self-supervised ADMM unrolling enables mesoscale high-resolution motion-robust diffusion-weighted imaging

Motivation: High-resolution and motion-robust diffusion-weighted imaging (DWI) is clinically demanding. A self-supervised image reconstruction model that leverages spatial-diffusion complementary sampling and convolution is beneficial to high-quality clinical DWI. Goal(s): To develop an efficient self-supervised algorithm unrolling technique for high-resolution and motion-robust DWI. Approach: We unroll the alternating direction method of multipliers (ADMM) to perform scan-specific self-supervised learning for deep DWI reconstruction. Results: We demonstrate that (1) ADMM unrolling is generalizable across slices, (2) ADMM unrolling outperforms compressed sensing with locally-low rank (LLR) regularization in terms of image sharpness, tissue continuity and motion robustness, (3) ADMM unrolling enables clinically feasible inference time. Impact: Our proposed ADMM unrolling enables whole brain DWI of 21 volumes at 0.7 mm isotropic resolution and 10 minutes scan, and shows higher signal-to-noise ratio (SNR), clearer tissue delineation, and improved motion robustness, which make it plausible for clinical translation.