Step-Calibrated Diffusion for Biomedical Optical Image Restoration
Authors: Yiwei Lyu, Sung Jik Cha, Cheng Jiang, Asadur Zaman Chowdury, Xinhai Hou, Edward S. Harake, Akhil Kondepudi, Christian Freudiger, Honglak Lee, Todd C. Hollon
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | RSCD outperforms other widely used unpaired image restoration methods on both image quality and perceptual evaluation metrics for restoring optical images. Medical imaging experts consistently prefer images restored using RSCD in blinded comparison experiments and report minimal to no hallucinations. Finally, we show that RSCD improves performance on downstream clinical imaging tasks, including automated brain tumor diagnosis and deep tissue imaging. We evaluate the perceptual quality of image restoration on 12K unpaired low-quality images sampled from the largest public SRH dataset, Open SRH (Jiang et al. 2022). We evaluate RSCD against the above baseline methods and ablations. We use Frechet Inception Distance (FID) (Heusel et al. 2017) and CLIP Maximum Mean Discrepancy (CMMD) (Jayasumana et al. 2024) between the restored images and the 4.5K expert-selected high-quality images as metrics. We report all results in Table 1. |
| Researcher Affiliation | Collaboration | 1University of Michigan 2Western Michigan University 3 Invenio Imaging 4 LG AI Research EMAIL |
| Pseudocode | Yes | Algorithm 1: Restorative Step-Calibrated Diffusion with Step Calibration and Dynamic Recalibration: Sampling |
| Open Source Code | Yes | Code https://github.com/MLNeurosurg/restorative stepcalibrated diffusion |
| Open Datasets | Yes | We evaluate the perceptual quality of image restoration on 12K unpaired low-quality images sampled from the largest public SRH dataset, Open SRH (Jiang et al. 2022). |
| Dataset Splits | Yes | Training data was generated from approximately 2500 patients who underwent intraoperative SRH imaging to evaluate tissue during surgery (Orringer et al. 2017). Whole slide SRH images are approximately 6000 6000 pixels, which are then divided into 256 256 pixel patches, resulting in 1 million total patches. To obtain high-quality SRH images, optical imaging experts manually selected 4.5K high-quality patches, and then we automatically filtered through the remaining patches to obtain 840K relatively high-quality patches, using the 4.5K patches as guidance. ... We evaluate the perceptual quality of image restoration on 12K unpaired low-quality images sampled from the largest public SRH dataset, Open SRH (Jiang et al. 2022). ... we obtained 2,135 pairs of near-registered low-quality/high-quality SRH patches. |
| Hardware Specification | No | This research was also supported, in part, through computational resources and services provided by Advanced Research Computing, a division of Information and Technology Services at the University of Michigan. |
| Software Dependencies | No | The paper mentions several models and frameworks like Res Net-50, Cycle GAN, DDPMs, but does not specify software dependencies with version numbers. |
| Experiment Setup | Yes | In practice, we set T = 1000. ... In practice, we set T = 200 due to the distribution in of tpred as discussed in the previous section. ... During training, a step number t U(0, T) is sampled and t steps of Gaussian noise are added to a high-quality image according to a cosine schedule (Ho, Jain, and Abbeel 2020). The calibrator is trained to predict t using an L2 loss between the prediction tpred and t. |