Catalog of Regulatory Science Tools to Help Assess New Medical Devices
Technical Description
The HistoGen Tool generates synthetic histopathology images using cell nuclei mask based on conditional denoising diffusion probabilistic model. The tool allows users to create realistic Hematoxylin and Eosin (H&E) stained tissue images guided by cell nuclei segmentation masks as input. This addresses the challenge of limited availability of annotated whole slide images in computational pathology.
The model architecture uses a conditional U-Net where segmentation masks are injected using a spatially-adaptive normalization module to guide the generation process and iteratively convert a Gaussian noise array into a histopathology image. The segmentation mask used for guidance during the diffusion process serves as the annotation for all images generated using the mask.
Intended Purpose
The HistoGen Tool can be used for augmenting training datasets for AI model development in digital pathology, generating synthetic data for rare diseases or underrepresented cases where real data is limited, and supporting performance assessment of nuclei segmentation and classification algorithms. The tool helps alleviate the burden of costly manual annotation by pathologists while maintaining clinically relevant features. The clinical application areas include AI development and validation for digital pathology in nuclei segmentation, cancer diagnosis, and tissue classification tasks.
Testing
The HistoGen Tool has been tested on multiple publicly available histopathology datasets including MoNuSeg, TNBC, 2018 Data Science Bowl, and PanNuke datasets. The output results have been verified through quantitative metrics. The synthetic images generated by the HistoGen tool have been confirmed by a pathologist to be realistic and contain structures observed in real histopathology data. The tool demonstrates superior performance with fine-tuning compared to coarse training and successfully avoids mode collapse issues reported in other generative models. The HistoGen Tool methodology and verification results have been presented in:
- Kahaki, Seyed, et al. "Assessment of Conditional Image-to-Image Diffusion Model for Synthetic Histopathology Image Generation" Medical Imaging 2026: Digital and Computational Pathology. SPIE, 2026.
- Kahaki, Seyed, et al. "Assessment of Conditional Image-to-Image Diffusion Model for Synthetic Histopathology Image Generation" FDA Science Forum 2025.
Limitations
The current evaluation metrics are based on models trained on natural images which may not fully capture histopathology-specific features, and development of domain-specific metrics is ongoing. The tool has been primarily validated on H&E-stained images and requires further validation for other staining protocols. The image generation process requires access to GPU resources and may be slow for very large mask sizes. The tool currently supports 2D histopathology images only.
Supporting Documentation
The documentation is embedded in the tool as a user guide for HistoGen tool GUI and function usage. Please see the online documentation: https://github.com/DIDSR/HistoGen
Contact
Tool Reference
- RST Reference Number: RST26DP02.01
- Date of Publication: 05/04/2026
- Recommended Citation: U.S. Food and Drug Administration. (2026). HistoGen: Histopathology Cell Nuclei Image Generation Tool (RST26DP02.01). https://cdrh-rst.fda.gov/histogen-histopathology-cell-nuclei-image-generation-tool