Noise-Adaptive MRI Denoising Using Self-Supervised Learning with Average-to-Average (Avg2Avg) Loss

Motivation: Low-field MRI scanners offer improved accessibility but are plagued by low-SNR, often requiring averaging of scans which prolongs scan times and reduces accessibility. Goal(s): To introduce SNAPER, a framework which accelerates the conventional low-SNR MRI acquisition pipeline (scan averaging) via scan denoising. Approach: SNAPER proposes the collection of two imaging averages, and trains a noise-adaptive denoising neural network with an average-to-average training. After training, averages are combined before denoising to yield improved image quality. Results: We validate on synthetically contaminated fully-sampled and 2x GRAPPA-reconstructed data. On 0.55T T2-weighted accelerated Prostate image data, preliminary results indicate SNAPER can achieve 4x acceleration by denoising. Impact: The proposed denoising technique could greatly encourage the use of 0.55T MRI and other low-SNR MRI scanners, making imaging more affordable and accessible.