Large-scale High-Dimensional Image Reconstruction via Delayed and Distributed Computing with BART

Motivation: BART is a flexible toolbox for various high-dimensional MRI reconstruction tasks but is limited by a high peak-memory requirement. Goal(s): Generically reduce the BART's peak-memory requirements for large-scale, high-dimensional reconstruction problems while maintaining its flexibility and allowing for distributed computing. Approach: Implement a delayed execution and distributed computing backend for BART that transparently optimizes numeric operations in the background. Results: Peak memory requirements are reduced multiple-fold for motion-resolved 5D cardiac MRI and motion-resolved cardiac T1-mapping. Impact: BART is a versatile toolbox for MRI reconstruction that is widely used due to its flexibility and performance. It is now more memory-efficient and supports distributed computing on multiple GPUs and nodes, allowing researchers to explore even higher-dimensional datasets.