Autotuning Plug-and-Play Algorithms for MRI

TL;DR

This paper introduces an automatic tuning method for plug-and-play MRI reconstruction algorithms that leverages noise variance information, achieving near-optimal performance with faster convergence than previous methods.

Contribution

We develop a fast, robust auto-tuning PnP-PDS algorithm for MRI that uses noise variance to optimize performance without manual parameter tuning.

Findings

Converges close to genie-tuned performance

Significantly faster than existing autotuning methods

Effective in MRI image recovery

Abstract

For magnetic resonance imaging (MRI), recently proposed "plug-and-play" (PnP) image recovery algorithms have shown remarkable performance. These PnP algorithms are similar to traditional iterative algorithms like FISTA, ADMM, or primal-dual splitting (PDS), but differ in that the proximal update is replaced by a call to an application-specific image denoiser, such as BM3D or DnCNN. The fixed-points of PnP algorithms depend upon an algorithmic stepsize parameter, however, which must be tuned for optimal performance. In this work, we propose a fast and robust auto-tuning PnP-PDS algorithm that exploits knowledge of the measurement-noise variance that is available from a pre-scan in MRI. Experimental results show that our algorithm converges very close to genie-tuned performance, and does so significantly faster than existing autotuning approaches.