Score-based Self-supervised MRI Denoising
Authors: Jiachen Tu, Yaokun Shi, Fan Lam
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on publicly available datasets, including a low-field MRI dataset, to evaluate the performance of C2S. Our results demonstrate that C2S achieves state-of-the-art performance among self-supervised methods and, after extending to multi-contrast on the M4Raw dataset, shows state-of-the-art performance among both self-supervised and supervised methods. Notably, we are among the first to comprehensively analyze and compare self-supervised and supervised learning approaches in MRI denoising. Our findings reveal that C2S not only bridges the performance gap but also offers robust performance under varying noise conditions and MRI contrasts. This indicates the potential of self-supervised learning to achieve competitive performance with supervised approaches when the latter are trained on practically obtainable higher-SNR labels, particularly in scenarios where perfectly clean ground truth is unavailable, offering a practical and robust solution adaptable to broader clinical settings. |
| Researcher Affiliation | Academia | Jiachen Tu, Yaokun Shi & Fan Lam Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign Champaign, IL 61820, USA EMAIL |
| Pseudocode | Yes | Algorithm 1 Corruption2Self Training Procedure Algorithm 2 Multi-Contrast Corruption2Self Training Procedure Algorithm 3 Detail Refinement Training Procedure |
| Open Source Code | No | The project website is available at: https://jiachentu.github. io/Corruption2Self-Self-Supervised-Denoising/. The provided URL leads to a project website, not an explicit statement of source code release or a direct link to a code repository. |
| Open Datasets | Yes | We conduct extensive experiments on publicly available datasets, including a low-field MRI dataset, to evaluate the performance of C2S. In-vivo Dataset (M4Raw): The M4Raw dataset Lyu et al. (2023) contains multi-channel k-space images across three contrasts (T1-weighted, T2-weighted, and FLAIR) from 183 participants. Simulated Dataset (fast MRI): We utilized single-coil knee data from the fast MRI dataset Zbontar et al. (2018), selecting patient entries matching those in MINet Feng et al. (2021) to ensure contrast correspondence. |
| Dataset Splits | Yes | For the M4Raw dataset, optimization was performed using Adam with learning rate 1e-4 and weight decay 1e-4. For the fast MRI dataset, Adam was configured with learning rate 1e-4 and weight decay 5e-2. Critical hyperparameters (learning rate, weight decay, batch size, maximum noise level T) were optimized based on validation performance. Early stopping was implemented to prevent overfitting, and final models were selected based on optimal validation metrics before test set evaluation. |
| Hardware Specification | Yes | All experiments were conducted on NVIDIA A6000 GPUs. |
| Software Dependencies | No | In our implementation, we utilize the skimage package Van der Walt et al. (2014) for noise level estimation, which proves sufficient for optimal performance. The paper mentions 'skimage package' and 'Adam optimizer' but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | For the M4Raw dataset, optimization was performed using Adam with learning rate 1e-4 and weight decay 1e-4. For the fast MRI dataset, Adam was configured with learning rate 1e-4 and weight decay 5e-2. Critical hyperparameters (learning rate, weight decay, batch size, maximum noise level T) were optimized based on validation performance. Early stopping was implemented to prevent overfitting, and final models were selected based on optimal validation metrics before test set evaluation. |