orGAN: A Synthetic Data Augmentation Pipeline for Simultaneous Generation of Surgical Images and Ground Truth Labels

Abstract

Deep learning in medical imaging faces obstacles: limited data diversity, ethical issues, high acquisition costs, and the need for precise annotations. Bleeding detection and localization during surgery is especially challenging due to the scarcity of high-quality datasets that reflect real surgical scenarios. We propose orGAN, a GAN-based system for generating high-fidelity, annotated surgical images of bleeding. By leveraging small "mimicking organ" datasets, synthetic models that replicate tissue properties and bleeding, our approach reduces ethical concerns and data-collection costs. orGAN builds on StyleGAN with Relational Positional Learning to simulate bleeding events realistically and mark bleeding coordinates. A LaMa-based inpainting module then restores clean, pre-bleed visuals, enabling precise pixel-level annotations. In evaluations, a balanced dataset of orGAN and mimicking-organ images achieved 90% detection accuracy in surgical settings and up to 99% frame-level accuracy. While our development data lack diverse organ morphologies and contain intraoperative artifacts, orGAN markedly advances ethical, efficient, and cost-effective creation of realistic annotated bleeding datasets, supporting broader integration of AI in surgical practice.