A Generative Foundation Model for Scalable Cytology Image Synthesis in AI-Powered Diagnostics

Cytology is a cornerstone of pathologic diagnosis. However, the use of artificial intelligence (AI) models for cytology-based diagnostics remains constrained by limited data availability and stringent privacy regulations. This study aims to develop COIN, a controllable cytology image generation foundation model, to address these challenges by synthesizing high-quality cytology images to enhance AI diagnostics and support clinical applications.

The COIN model was trained on a large-scale dataset of 112,226 cytology image-report pairs from 16 anatomic sites. Using diagnostic textual reports, it generates high-fidelity cytology images with morphologically and semantically coherent features. Expert cytologists evaluated the generated images for anatomic and diagnostic authenticity. The model's utility was assessed through data augmentation experiments, AI model training under data-scarce conditions, and content-based image retrieval applications.

Expert evaluations confirmed the high anatomic and diagnostic fidelity of the images generated by COIN. When used for data augmentation, COIN significantly improved the performance of diagnostic AI models across various tasks. Under data-scarce conditions, models trained exclusively on COIN-generated images demonstrated effective generalization to real-world datasets. Furthermore, COIN supported content-based image retrieval, offering a novel tool for case referencing and clinical decision support.

COIN represents a robust and privacy-preserving framework for scalable cytology data generation. Its ability to synthesize realistic images and enhance AI diagnostics highlights its broad applicability in computational pathology, providing a valuable tool to accelerate the development and implementation of AI-based diagnostic solutions.