Zero-Shot Recon2Recon: Data-Free Unsupervised Denoiser Learning for Plug-and-Play Magnetic Resonance Imaging Reconstruction

This work presents a novel zero-shot learning method for undersampled magnetic resonance imaging (MRI) reconstruction. The proposed method utilizes a plug-and-play approach, wherein the denoiser neural network, serving as the image prior, is trained solely with the single acquired undersampled k-space data. Specifically, the training of the denoiser employs the Noise2Noise and Self2Self frameworks; the acquired k-space is divided into two disjoint subsets using complementary subsampling masks, which are then reconstructed to serve as the training data and target labels, respectively. We provide theoretical justification for the proposed method, demonstrating the feasibility of training an effective denoiser without the need for extra training data or fully sampled ground truths. The trained denoiser is subsequently utilized as a plug-and-play denoiser for undersampled MRI reconstruction. We have evaluated and compared the performance of our method with several other reconstruction techniques, thereby demonstrating the advantages of the proposed approach.