Chromosome Mask-Conditioned Generative Inpainting for Atypical Mitosis Classification

Sweta Banerjee, Viktoria Weiss, Thomas Conrad, Taryn A. Donovan, Jonas Ammeling, Rutger H.J. Fick, Jonas Utz, Robert Klopfleisch, Christopher Kaltenecker, Christof Bertram, Katharina Breininger, Marc Aubreville

Keywords: Atypical Mitoses, Generative Inpainting, Classification, Denoising diffusion models, Conditional GANs

TL;DR: Introducing a triple-conditioned inpainting framework using a context patch, inpainting and chromosome segmentation masks to synthesize diverse normal and atypical mitoses via GANs and diffusion, substantially boosting classification performance.

Abstract: Atypical mitoses are critical prognostic markers for tumor proliferation, yet classification efforts are compromised by class imbalance, data scarcity, and noisy labels. Our work focuses on hematoxylin and eosin (H\&E)-stained histopathology images, where identifying such mitoses is particularly challenging due to overlapping morphological features and stain variability. We address these challenges with a novel approach for biologically informed inpainting, conditioned on a histological context patch, an inpainting mask, and a chromosome segmentation mask. This triple-conditioned generative strategy allows disentanglement of the mitotic figure shape information from the cellular context and enables the utilization of large-scale datasets that do not contain atypical sub-classification for training classification models. We evaluate both adversarial and denoising diffusion-based inpainting strategies. Our approach mitigates the lack of data diversity and label noise, thereby substantially improving classification performance for atypical vs. normal mitoses - as demonstrated by downstream classification with EfficientNet-B0 and Low-rank adaptation (LoRA) fine-tuned foundation models. We provide the complete source code, including all our methods, at our github repository: https://github.com/DeepMicroscopy/ChroMa-GI.

Submission Number: 27