Generative Neural Networks for Synthetic Histopathology Image Creation

Miriam Miklánková

Supervisor(s): Ing. Maroš Kollár

Slovak Technical University

Abstract: Mitotic count is a key component of the Nottingham Grading System used for breast cancer prognosis. However, the development of automated detection algorithms is limited by the scarcity of large annotated datasets and the high variability of tissue morphology. To address this challenge, we propose a two-stage generative pipeline for synthesizing high-quality histopathology images with precise control over mitotic figure placement. Our approach builds on Stable Diffusion, adapted to the H&E-stained tissue domain using Low-Rank Adaptation (LoRA) trained on the multi-domain MIDOG++ dataset. We then train a ControlNet model to condition image generation on binary segmentation masks, enabling precise spatial placement of mitotic figures. We evaluate multiple training configurations, with a particular focus on the effects of domain adaptation and transfer learning. Our experiments show that LoRA-based adaptation significantly improves morphological realism. In addition, training ControlNet from scratch produces better results than fine-tuning models pretrained on natural image segmentation tasks. The final pipeline, combined with Vahadane stain normalization, achieved a Kernel Inception Distance (KID) of 0.4 at 256x256 resolution, outperforming configurations without domain adaptation (KID = 1.44). Scaling to 512x512 resolution achieved a KID of 0.75, demonstrating that realistic, spatially controlled synthesis is achievable even at higher resolutions.
Keywords: Computer Vision, Image Processing
Full text:
Year: 2026