Research map / Image generation / Image restoration
Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement
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.
Di-Fusion is a single-stage, fully self-supervised diffusion model for dMRI denoising that uses a Fusion forward process, a noise term built from differences between adjacent volumes, training on only the latter diffusion steps, and adaptive early-stopping sampling.
Problem
Denoising dMRI without clean ground truth or paired data, where existing self-supervised diffusion approaches (e.g., DDM2) require multi-stage training and tend to over-denoise and perform poorly on downstream tasks.
Approach
Builds a diffusion chain from x' (previous volume) to x (target volume) via a Fusion process that interpolates the two measurements, replaces Gaussian noise with a zero-mean, spatially shuffled difference noise (the "Di-" process), trains only the last T_c diffusion steps, and samples with Run-Walk acceleration plus an adaptive termination criterion (CSNR).
Key results
- On simulated data, Di-Fusion achieves the best SSIM in the first four simulation settings (e.g., 39.05 vs 26.02 for DDM2 in Simulation 1) and the best PSNR in the three highest-SNR-degradation settings are mixed, with P2S best PSNR at the mildest noise levels (e.g., 33.87 in Simulation 5).
- Di-Fusion achieves the best R^2 for DTI and CSD microstructure fitting across all four settings on in-vivo data and produces the fewest streamlines while keeping high FBC in OR tractography, where DDM2 yields fewer streamlines but misses high-FBC fibers.
- Adaptive termination combined with Run-Walk accelerated sampling greatly reduces sampling time while sampling quality stays largely unchanged.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Di-Fusion instead of the strongest reported baseline | dMRI denoising · Simulations 1–5 | SSIM (%): 39.05 vs 26.02; 62.02 vs 50.83; 77.36 vs 71.77; 83.43 vs 78.61; 89.52 vs 87.31 | Helped |
| use Di-Fusion instead of the strongest reported baseline | dMRI denoising · Simulations 1–3 | PSNR (dB): 19.91 vs 17.92; 23.60 vs 22.13; 26.96 vs 26.80 | Helped |
| use Di-Fusion instead of the strongest reported baseline | dMRI denoising · Simulations 4–5 | PSNR (dB): 28.69 vs 30.26; 30.63 vs 33.87 | Hurt |
| remove the Fusion process (ablation) | Di-Fusion · In-vivo dMRI denoising | image drift | Hurt |
| add the Fusion process (ablation) | Di-Fusion · In-vivo dMRI denoising | image drift | Helped |
| remove the Di- process (ablation) | Di-Fusion · In-vivo dMRI denoising | high-frequency information | Hurt |
| add the Di- process (ablation) | Di-Fusion · In-vivo dMRI denoising | high-frequency information | Helped |
| train without restricting training to the latter diffusion steps (ablation) | Di-Fusion · In-vivo dMRI denoising | denoising quality | Hurt |
| train only the latter diffusion steps (ablation) | Di-Fusion · In-vivo dMRI denoising | denoising quality | Helped |
| vary Tc within the reported reasonable range (Tc<500) (ablation) | Di-Fusion · Microstructure model fitting | R² | No clear effect |
| use Run-Walk accelerated sampling (ablation) | Di-Fusion · dMRI sampling | sampling quality | No clear effect |
| add adaptive termination and Run-Walk accelerated sampling (ablation) | Di-Fusion · dMRI sampling | sampling time | Helped |
Surprising findings
- Patch2Self attains the highest PSNR at the mildest simulated noise levels, outperforming Di-Fusion there, despite lower SSIM.
- DDM2 yields the fewest streamlines in tractography yet misses high-FBC fibers, indicating over-denoising.
Limitations
- Inference can be long because diffusion sampling is slow, and adaptive termination adds computation.
- Assumes additive, thermal-like noise; signal-dependent noise and physiological artifacts are not addressed.
- Possible hallucinations inherent to generative models, mitigated but not eliminated by training only the latter diffusion steps.
Experimental setup
Datasets: Stanford HARDI Sherbrooke 3-Shell PPMI fastMRI
Benchmarks: Stanford HARDI Sherbrooke 3-Shell PPMI fastMRI simulated data
Metrics: PSNR SSIM R^2 (DTI and CSD microstructure fitting) SNR CNR FBC (fiber bundle coherency)
Baselines: ASCM Deep Image Prior Noisier2Noise Patch2Self DDM2 MPPCA Noise2Score Recorrupted2Recorrupted Patch2Self2
Builds on: DDPM DDIM Noise2Noise Noise2Self DDM2 Noisier2Noise
Closest papers by meaning
- Patch2Self2: Self-supervised Denoising on Coresets via Matrix Sketching (CVPR 2024 · similarity 0.91)
- Score-based Self-supervised MRI Denoising (ICLR 2025 · similarity 0.90)
- Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution (CVPR 2024 · similarity 0.87)
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.87)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.87)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.86)
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.