Reproduce “CytoSyn: a Foundation Diffusion Model for Histopathology -- Tech Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Thomas Duboudin, Xavier Fontaine, Etienne Andrier, Lionel Guillou, Alexandre Filiot, Thalyssa Baiocco-Rodrigues, Antoine Olivier, Alberto Romagnoni and 2 more
Owkin researchers introduce CytoSyn, a foundation latent diffusion model specifically optimized for generating histopathology images. It produces highly realistic and diverse H&E-stained images, outperforming existing models in generation quality and exhibiting strong robustness on out-of-distribution datasets.
Key insights
- Seemingly minor preprocessing details, such as image size and JPEG compression, profoundly impact generative model performance and evaluation metrics in histopathology, leading to order-of-magnitude FID score discrepancies.
- Scaling the training dataset from 40 million to 108 million images did not result in performance gains for CytoSyn, suggesting that larger training sets do not automatically translate to better generative quality in this specific context.
- Employing histopathology-specific feature extractors (like H0-mini) for both representation alignment and guidance is crucial for developing robust generative models and accurately assessing their quality in the medical image domain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.