An Ambient Denoising Score Matching Based Self-supervised Denoising Approach for Multicontrast Low-Field MRI

Motivation: Multicontrast MRI offers unique capabilities for diagnosis and tissue characterization, but it often has more limited trade-offs in speed/resolution/SNR, especially in low-field MRI. Goal(s): We developed a self-supervised denoising method for multicontrast MRI without requiring clean/high-SNR labels. Approach: An ambient denoising score matching based approach is proposed to denoise the target contrast image by effectively leveraging both its noisy counterparts and noisy images of other contrasts. Results: The proposed method effectively enhanced SNR while preserving fine details, achieving state-of-the-art performance on the M4Raw dataset for different target contrasts. Impact: Our method represents a new approach for self-supervised multicontrast MRI denoising. It may offer better trade-offs in SNR, resolution, and speed to benefit many low-field applications.