Motivation: Current deep learning-based MRI denoising methods either produce over-smoothed results or fail to generalize well across different scanners and protocols, limiting their clinical applicability. Goal(s): To develop a zero-shot MRI denoising framework that can handle complex real-world noise patterns while preserving critical diagnostic details, without requiring paired training data or scanner-specific fine-tuning. Approach: We combine a noise transformation network with a pretrained diffusion model, using range-null space decomposition during sampling to ensure both noise removal and detail preservation. Results: Our method achieves 485% SNR improvement on out-of-distribution high noise data, significantly outperforming existing approaches while maintaining diagnostic quality across diverse MRI protocols. Impact: The proposed zero-shot method enables immediate deployment across different MRI protocols without retraining, enabling reduced scan times or lower field strengths while maintaining diagnostic quality. The approach opens possibilities for applying diffusion models to MRI challenges without extensive data collection.
Zhang et al. (Tue,) studied this question.