Bridging Deep Learning and Physical Knowledge for Self-Supervised Magnetic Resonance Image Denoising
MRI is a vital diagnostic tool, offering excellent soft-tissue contrast without ionizing radiation, but its image quality is fundamentally limited by signal-to-noise ratio, which drops especially low in certain settings like low-field MRI or diffusion-weighted imaging (DWI). Low-field MRI systems are becoming more attractive for their lower cost and better accessibility, but they inherently produce weaker signals, making good denoising essential to keep the images diagnostically useful. Deep learning has become the go-to solution for MRI denoising, but most methods rely on supervised training with paired noisy and clean images, which are rarely available in real clinical settings. This thesis instead focuses on self-supervised denoising, which learns to reduce noise without ever needing a noise-free reference image.
The central contribution extends Stein’s Unbiased Risk Estimator (SURE), a self-supervised denoising framework, to MRI by incorporating spatially resolved noise maps that capture how noise characteristics vary across a reconstructed MR image. These noise maps guide training and also let the denoising strength be adjusted dynamically at inference time. Extensive experiments show this self-supervised approach performs competitively against supervised methods. Building on this, the thesis extends the framework to DWI, where multiple diffusion directions and repeated measurements add complexity, introducing a new way to estimate noise maps directly from the acquired data to make the method more practically usable. To evaluate all of this reliably without clean ground truth, the thesis also proposes a self-supervised quantitative assessment strategy that checks whether the noise removed during denoising preserves the anatomically relevant structures.
To improve robustness and computational efficiency, the thesis explores combining these self-supervised techniques with low-parameter models, in particular a known-operator architecture built around bilateral filters, which improves both interpretability and how well the method generalizes across different imaging conditions. Finally, the thesis takes a critical look at diffusion models, originally built for image generation but increasingly explored for medical image denoising, and finds that their iterative sampling process often progressively degrades image fidelity, raising real concerns about whether they’re actually suitable for clinical use.
Together, these contributions offer a thorough investigation of self-supervised MRI denoising, connecting theoretical development with practical implementation, and showing how combining domain knowledge with deep learning can improve MR image quality while keeping the results robust, generalizable, and clinically trustworthy.