Consensus ADMM for Distributed, Constrained Reconstruction with Low-Rank Subspace and Phase Priors

Motivation: Spatiotemporally-undersampled sequences help to accelerate scans and push resolution limits. However, reconstruction of this data is slow and memory-intensive. Goal(s): We aim to accelerate such reconstructions by splitting our objective into independent components and parallelizing the optimization across multiple compute devices. Approach: We implement fast, GPU-accelerated proximal operators for data consistency, Locally-Low-rank, and S-LORAKS. We use a technique called Consensus ADMM to fuse them together in a distributed fashion. Results: Our reconstruction enables reconstruction of high-resolution 3D magnetic resonance fingerprinting datasets and quantitative maps quickly and with good accuracy. Impact: By adding more sophisticated prior knowledge to the reconstruction, we can further accelerate the scan, shortening scan times while maintaining quality.Our techniques are also quite general and apply broadly to a wide variety of reconstruction problems.