Comparative Analysis of Generative and Diffusion Models for Mitosis Synthesis in Histopathology Images

Accurate cancer diagnosis depends on identifying key cellular features, including mitotic figures, which have a major influence on treatment decisions and patient outcomes. Histopathological evaluation remains the clinical gold standard for detecting mitotic figures, but any manual evaluation is laborious and often hindered by large and unpredictable inter-observer variability. The use of Generative Artificial Intelligence (GenAI) continues to gain traction as a viable means of producing synthetic medical images for the purpose of creating data augmentations to feed model training and reduce dependence on manually labelling images. However, the application of GenAI within histopathology has not been thoroughly assessed. In this study, we assess two leading GenAI architectures, Denoising Diffusion Probabilistic Models (DDPM) and StyleGAN3, on their ability to synthesise mitotic figures using the MIDOG++ dataset, which consists of multiple tumour types from various species. Our results show that DDPM captures fine structural morphology very well, and StyleGAN3's color reproduction was superior. Quantitative evaluation using Fréchet Inception Distance (FID), Structural Similarity Index (SSIM), and expert-based Receiver Operating Characteristic analysis showed that DDPM achieved lower FID and higher structural similarity, while both DDPM and StyleGAN3 produced synthetic images that were largely indistinguishable from real samples by expert pathologists (area under the curve ≈ 0.5). Although diffusion models exhibit improved structural fidelity relative to generative adversarial networks-based approaches, they introduce higher computational demands and unresolved questions regarding generalisability across institutions and imaging pipelines. This functional distinction illustrates complementary capabilities that can be judiciously utilised depending upon the augmentation needs of the dataset. By incorporating these high-quality synthetic images into training pipelines, researchers can reduce manual annotation burden, reduce observer variance, and improve the robustness of mitosis detection models. In contrast to prior histopathology studies that have largely focused on GAN-based synthesis, this work presents a systematic comparative evaluation of diffusion models (DDPM) and StyleGAN3 for mitotic figure generation using the MIDOG++ dataset, clarifying their respective strengths and limitations for synthetic data augmentation.