Low Latency and Generalizable Dynamic MRI via L+S Alternating GD and Minimization

Abstract

In this work, we develop novel MRI reconstruction approaches that are accurate, fast and low-latency for a large number of dynamic MRI applications, sampling schemes and sampling rates; without any problem-specific parameter tuning. We refer to this property of a single algorithm, without parameter tuning, being accurate and fast for many settings as generalizability. Generalizability is possible only for simple (few parameter) models such as low-rank (LR) or LR plus sparse (L plus S), and for simple few parameter algorithms based on these models, which is what we develop and evaluate in this work.