Assessment of Conditional Diffusion Model for Synthetic Histopathology Image Generation

This study trains conditional denoising diffusion models on four benchmark histopathology datasets and evaluates generated images using conventional FID and Inception Score, pathology-foundation-model variants, precision-recall metrics, and downstream nuclei segmentation. The pathology-adapted Inception Score correlated more strongly with AJI+ performance (r=0.6096, p=0.0122) than the original ImageNet-based score (r=0.0708, p=0.7944). The reported observations also suggest that greater diversity in synthetic training data may contribute more to segmentation performance than higher visual fidelity of individual images.