Distributed Memory-Efficient Physics-Guided Deep Learning Reconstruction for Large-Scale 3d Non-Cartesian MRI

Physics-guided deep learning (PG-DL) reconstruction has emerged as a powerful strategy for accelerated MRI. However, adopting PG-DL on 3D non-Cartesian MRI remains a challenge due to GPU hardware limitations. In this paper, we utilize multiple memory-efficient techniques to accomplish PG-DL on large-scale 3D kooshball coronary MRI. We first leverage a recently proposed approach to keep only one unrolled step on GPUs. We then utilize a Toeplitz approach to represent the multi-coil encoding operator. Subsequently, we distribute the most memory-consuming data consistency operations into multiple GPUs, enabling conjugate gradient iterations without necessitating coil compression. Finally, we employ mixed-precision training to further reduce memory consumption.The combination of these methods enable training of high-quality PG-DL reconstruction for 3D kooshball trajectories, and our results show reconstruction improvement compared to existing strategies.