Magnetic Resonance Imaging (MRI) is a crucial tool in medical diagnostics, offering exceptional soft tissue contrast and detailed anatomical information without ionizing radiation. However, its image quality is fundamentally constrained by the signalto- noise ratio (SNR), which can be particularly low in specific scenarios, such as low-field MRI or diffusion-weighted imaging (DWI). Low-field MRI systems are gaining interest due to their reduced cost and higher accessibility, yet they inherently produce weaker signals, making effective denoising techniques essential to preserve diagnostic quality. Deep learning-based methods have emerged as a powerful solution for MRI denoising, achieving state-of-the-art performance. However, most deep learning-based approaches rely on supervised training, requiring paired noisy and clean images, which are often unavailable in clinical settings. To circumvent this limitation, this thesis focuses on self-supervised denoising techniques, which enable learning-based noise reduction without the need for noise-free ground truth data. A central contribution of this work is the extension of the Stein’s unbiased risk estimator (SURE) framework to MRI, incorporating spatially resolved noise maps to account for the spatially varying noise characteristics commonly encountered in reconstructed magnetic resonance (MR) images. These noise maps not only guide the training process but also enable dynamic control over the denoising strength during inference. Comprehensive experiments are conducted to compare the proposed self-supervised approach with supervised methods, demonstrating competitive performance. Building on this foundation, the thesis extends the self-supervised framework to DWI, where multiple diffusion directions and repeated measurements introduce additional challenges. A novel approach for estimating noise maps directly from the acquired data is introduced, increasing the method’s practical applicability. To ensure reliable evaluation, a self-supervised quantitative assessment strategy is proposed, analyzing the noise removed during denoising to validate that relevant anatomical structures are preserved. To improve both robustness and computational efficiency, the integration of the proposed self-supervised denoising techniques with low-parameter models is explored, specifically leveraging a known operator-based architecture based on bilateral filters. This approach not only enhances interpretability but also provides improved generalizability across different imaging conditions. Additionally, to further examine the role of model robustness and data fidelity in MRI denoising, the applicability of diffusion models is critically evaluated. Originally designed for image generation, diffusion models have recently been explored for medical image denoising. However, the presented findings reveal that their iterative sampling strategies often lead to a progressive degradation of image fidelity, raising concerns about their suitability for clinical applications. Together, these contributions provide a comprehensive investigation into selfsupervised MRI denoising, bridging theoretical advancements with practical implementation considerations. By integrating domain knowledge with deep learning methodologies, this thesis highlights the potential of self-supervised techniques for enhancing MR image quality while ensuring robustness, generalizability, and clinical reliability.
Laura Pfaff (Thu,) studied this question.