BLUE RIBBON POSTER IMAGING: Self-Supervised Low-Field MRI Denoising with Multi-Contrast Structural Guidance
Abstract
Purpose
Low-field MRI suffers from intrinsically low SNR, which limits image quality and slows wide clinical adoption. Deep learning–based denoising shows strong promise, yet supervised training requires paired high-quality references that are rarely available for low-field acquisitions. We develop a self-supervised denoising framework that leverages co-registered multi-contrast MRI as complementary anatomical guidance to enable high-quality multi-contrast denoising from noisy acquisitions alone.
Methods
The 0.3T M4Raw MRI dataset was used with 158 subjects for training and 25 for testing. Co-registered T1/T2 images were jointly used as complementary anatomical guidance. Each noisy image was decomposed into two paired views via complementary sub-sampling, yielding interleaved observations with nearly identical signal content but different noise realizations. The network was trained to predict one view from its paired counterpart. With approximately zero-mean, signal-independent noise, noisy–to-noisy regression drives the network toward the expected signal. T1 and T2 were jointly-denoised using a dual-stream encoder–decoder with a shared bottleneck, while keeping modality-specific branches and intra-modality skips to prevent cross-contrast feature leakage. Training was stabilized by two consistency regularizers: cross-contrast structural consistency (implemented with MIND descriptors) to align anatomical patterns, and shared-sampling consistency to reduce sampling-induced variance by enforcing agreement at shared sampling locations.
Results
Our method achieved the highest performance among the evaluated self-supervised baseline methods (T1: SSIM 0.902, PSNR 32.13 dB; T2: SSIM 0.876, PSNR 31.42 dB), improving SSIM by ~0.02–0.03 and PSNR by ~0.9–1.1 dB over the best-performing baseline method. Qualitatively, zoom-in views show reduced unnatural texture patterns and a smoother, more coherent appearance compared with other self-supervised methods.
Conclusion
A self-supervised multi-contrast MRI denoising network that learns from noisy data without high-quality references was developed for low-field MRI. By jointly denoising co-registered T1/T2 with a leakage-aware dual-stream architecture and consistency regularizers that preserve shared anatomy, the method achieves improved SSIM/PSNR over established self-supervised baselines methods.