Research map / Image generation / Image restoration
Score-based Self-supervised MRI Denoising
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Corruption2Self (C2S), a score-based self-supervised MRI denoiser trained on noisy data via a generalized denoising score matching loss, with noise-level reparameterization, detail refinement, and multi-contrast extension.
Problem
Supervised MRI denoisers need high-SNR clean labels that are rarely available, while existing self-supervised methods (blind-spot/masking based) oversmooth fine detail and often underperform supervised methods.
Approach
Generalizes denoising score matching to ambient (noisy-only) settings by training a network to predict the conditional expectation of a higher-SNR version of the observed noisy image given a further corrupted version, with a reparameterized noise schedule, a non-zero target noise level for detail refinement, and NVC-MSA noise-conditioned attention in a DDPM-style U-Net.
Key results
- On M4Raw test, C2S with detail refinement reaches 32.77 dB/0.919 (T1), 32.33 dB/0.890 (T2), 32.51 dB/0.876 (FLAIR), best among self-supervised methods.
- Multi-contrast C2S reaches 33.89 dB/0.922 on T1 (FLAIR & T1 inputs), exceeding best supervised Noise2Noise (32.59 dB) on the same target.
- Reparameterization of noise levels improves validation PSNR on M4Raw from 31.14 to 34.43 dB for T1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add detail refinement (ablation) | C2S · M4Raw test, T1 | PSNR: 32.59 -> 32.77 | Helped |
| Add detail refinement (ablation) | C2S · M4Raw test, T2 | PSNR: 32.28 -> 32.33 | Helped |
| Add detail refinement (ablation) | C2S · M4Raw test, FLAIR | PSNR: 32.43 -> 32.51 | Helped |
| Add detail refinement (ablation) | C2S · fastMRI test, PD, σ = 13/255 | PSNR: 33.36 -> 33.48 | Helped |
| Add detail refinement (ablation) | C2S · fastMRI test, PD, σ = 25/255 | PSNR: 30.62 -> 30.67 | Helped |
| Add detail refinement (ablation) | C2S · fastMRI test, PDFS, σ = 13/255 | PSNR: 30.72 -> 30.91 | Hurt |
| Add noise-level reparameterization (ablation) | C2S · M4Raw validation, T1 | PSNR: 31.14 -> 34.43 | Helped |
| Add noise-level reparameterization (ablation) | C2S · M4Raw validation, T2 | PSNR: 30.53 -> 33.82 | Helped |
| Add noise-level reparameterization (ablation) | C2S · M4Raw validation, FLAIR | PSNR: 30.43 -> 32.56 | Helped |
| Use time conditioning and NVC-MSA rather than a DDPM architecture (ablation) | C2S · M4Raw T1 | PSNR: 34.82 -> 34.91 | Helped |
| Use time conditioning and NVC-MSA rather than a DDPM architecture (ablation) | C2S · fastMRI PD | PSNR: 33.48 -> 33.63 | Helped |
| Use multi-contrast inputs (T1 and FLAIR) for T1 denoising | C2S · M4Raw T1 | PSNR: 32.77 -> 33.89 | Helped |
Surprising findings
- Supervised SwinIR and Restormer trained on 3-repetition averages do not significantly outperform self-supervised methods on higher-SNR test labels.
- Blind-spot self-supervised methods such as PUCA and LG-BPN perform worse than classical BM3D on MRI data.
Limitations
- Relies on Gaussian noise assumption, which may not hold for Rician noise in low-SNR regions; VST is suggested but not evaluated.
- Requires noise level estimate; robustness to ±50% error is shown but blind operation depends on external estimators.
- M4Raw test labels are averages of repetitions rather than true clean images, and fastMRI labels still contain inherent noise.
Experimental setup
Datasets: M4Raw fastMRI
Benchmarks: M4Raw fastMRI
Metrics: PSNR SSIM
Baselines: NLM BM3D SwinIR Restormer Noise2Noise Noise2Void Noise2Self PUCA LG-BPN Noisier2Noise Recorrupted2Recorrupted
Builds on: Denoising Score Matching Ambient Denoising Score Matching Noisier2Noise DDPM Noise2Void
Closest papers by meaning
- Patch2Self2: Self-supervised Denoising on Coresets via Matrix Sketching (CVPR 2024 · similarity 0.91)
- Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement (ICLR 2025 · similarity 0.90)
- Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising (CVPR 2025 · similarity 0.86)
- Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum (NeurIPS 2025 · similarity 0.85)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.85)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.