Robust Plug-and-Play Methods for highly Accelerated Non-Cartesian MRI Reconstruction: A reproducible benchmark.

Motivation: Non-Cartesian MRI scans allow for faster scan time and high-resolution, but requires expensive reconstruction. Traditional algorithms delivers suboptimal results and deep-learning methods have improved image quality but are expensive to train and prone to overfitting. Goal(s): Propose a novel algorithm for multicoil, non-Cartesian MRI reconstruction by leveraging Plug-and-Play (PnP) methods and Half-Quadratic Splitting (HQS), with better reconstruction quality, stability, and generalization. Approach: 1)Boost the learned denoiser by preprocessing fastMRI dataset. 2) Generalized preconditioned PGD algorithm to HQS scheme. Results: Preconditioning improves image quality, with HQS providing the best. At 16x acceleration factor, all PnP methods outperform NCPDNet and variational approaches without further finetuning. Impact: The proposed algorithm's improved image quality and accelerated reconstruction times at a minimal cost, and generalizes to any sampling pattern and acceleration factor.