Joint MRI Reconstruction and Denoising Using Noise-Adaptive Self-Supervised Learning

Motivation: Deep learning methods are state-of-the-art in accelerated MRI reconstruction. However, they often lack robustness to changing SNR levels between training and inference, which may hinder their successful deployment in low-SNR regimes, and in particular low-field scanners. Goal(s): To introduce LPDSNet, a novel approach for robust deep-learning based joint MRI reconstruction and denoising without ground-truth data. Approach: LPDSNet directly parameterizes an unrolled primal-dual splitting algorithm, and achieves noise-robustness via learned noise-adaptive clipping. Results: LPDSNet demonstrates superior performance in both supervised and self-supervised learning compared to state-of-the-art networks. Additionally, we show novel noise-level robustness in self-supervised joint MRI reconstruction and denoising, where competing methods fail. Impact: LPDSNet surpasses current methods, especially under mismatched noise-level conditions between training and testing, making it highly effective for noisy, limited-sample MRI datasets and promising for low-SNR, low-field MRI applications.