Research map / Image generation / Image restoration
Step-Calibrated Diffusion for Biomedical Optical Image Restoration
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 RSCD, an unpaired diffusion restoration method for Raman-based optical (SRH) images that uses a ResNet-50 step calibrator to pick how many reverse-diffusion steps to run, with dynamic recalibration during sampling.
Problem
Restoring degraded stimulated Raman histology (SRH) optical images, where degradation is stochastic, tissue-dependent, and paired clean/degraded data cannot be collected; prior unpaired methods hallucinate or produce perceptually poor outputs.
Approach
Use a trained step calibrator to predict, per image, how many reverse-diffusion steps to apply to the degraded image (instead of full generation from noise), recalibrate the step count every d steps during sampling, and train with a region-mixing augmentation so the calibrator tracks the most degraded region.
Key results
- On 12K unpaired OpenSRH images, RSCD achieves FID 32.02 and CMMD 0.128, best among baselines (best baseline CycleGAN FID 37.26, CMMD 0.196).
- On near-registered pairs, RSCD achieves best FID 21.05 and LPIPS 0.280 and second-best PSNR 29.06 and SSIM 0.791, behind Deep Image Prior on PSNR/SSIM.
- Inference takes ~14.5 s per 256x256 patch (34 NFE) vs 431 s (1000 NFE) for full conditional diffusion, roughly 30x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic recalibration (ablation) | RSCD · OpenSRH unpaired images | FID: 34.62 -> 32.02 | Helped |
| replace the nonparametric noise estimator with the learned step calibrator (ablation) | RSCD · OpenSRH unpaired images | FID: 35.53 -> 32.02 | Helped |
| use learned step calibration instead of a fixed 10-step schedule (ablation) | RSCD · OpenSRH unpaired images | FID: 36.41 -> 32.02 | Helped |
| use learned step calibration instead of a fixed 50-step schedule (ablation) | RSCD · OpenSRH unpaired images | FID: 42.06 -> 32.02 | Helped |
| use a cosine instead of linear noise schedule (ablation) | RSCD · OpenSRH unpaired images | FID: 38.71 -> 32.02 | Helped |
| train the step calibrator with augmentation (ablation) | RSCD · OpenSRH unpaired images | FID: 33.14 -> 32.02 | Helped |
| change recalibration interval from d=10 to d=5 (ablation) | RSCD · OpenSRH unpaired images | FID: 32.02 -> 32.30 | No clear effect |
| change recalibration interval from d=10 to d=20 (ablation) | RSCD · OpenSRH unpaired images | FID: 32.02 -> 32.64 | No clear effect |
Surprising findings
- Deep Image Prior and median blur attain higher PSNR/SSIM despite visibly oversmoothed outputs, so pixel metrics are misleading for restoration quality.
- Degradation acts like adversarial noise for the tumor classifier: even small degradations can flip predictions, and restoration corrects many non-diagnostic and wrong predictions.
Limitations
- Each patch is restored independently, ignoring neighboring regions, so restoration consistency across the full megapixel image is not enforced.
- Trained with Gaussian noise although real SRH degradation is non-Gaussian; generalization relies on the assumption that DDPM steps make the residual more Gaussian-like.
- Evaluated on a single SRH modality and institution; generalization to other imaging settings is untested.
Experimental setup
Datasets: SRH high-quality patches from ~2500 patients (840K filtered) 4.5K expert-selected high-quality SRH patches OpenSRH (12K unpaired low-quality images) 2,135 near-registered LQ/HQ SRH pairs (evaluation only) 425 surgical specimen z-stacks 8,372 labeled low-quality images for tumor diagnosis
Benchmarks: OpenSRH unpaired LQ images near-registered SRH pairs SRH tumor classification z-stack SRH restoration
Metrics: FID CMMD PSNR SSIM LPIPS classification accuracy expert preference hallucination rate
Baselines: CycleGAN Synthetic noise / Noise2Noise Conditional Diffusion CCDF Regularized Reverse Diffusion Deep Image Prior Median Blur
Builds on: DDPM cosine noise schedule CCDF Regularized Reverse Diffusion CycleGAN
Closest papers by meaning
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.84)
- IR-CM: The Fast and General-purpose Image Restoration Method Based on Consistency Model (NeurIPS 2024 · similarity 0.84)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.83)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.83)
- Blind Image Restoration via Fast Diffusion Inversion (NeurIPS 2024 · similarity 0.83)
- Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 2025 · similarity 0.82)
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.