Improving generalization of polyp detection via conditional StyleGAN augmented training

Colorectal cancer outcomes are critically dependent on early diagnosis, yet colonoscopy screening suffers from significant miss rates, particularly for subtle flat or serrated lesions. Although artificial intelligence holds promise for computer-aided detection (CADe), system performance is frequently bottlenecked by the scarcity and imbalance of high-quality annotated datasets. To address this, we employed a conditional StyleGAN architecture to synthesize high-resolution images of colorectal neoplasms, leveraging over 150,000 images aggregated from diverse public datasets. When utilized to train YOLOv5 detection models, this synthetic data demonstrated high fidelity and significantly enhanced diagnostic performance. Hybrid augmentation improved the mean Average Precision from 0.86 to 0.93 on internal testing and markedly reduced the generalization gap on independent external validation sets. Crucially, recall for challenging flat and depressed lesions rose from 0.72 to 0.87. These findings indicate that generative augmentation effectively strengthens model robustness and generalization across diverse clinical scenarios. While currently limited to still imagery, this strategy provides a scalable solution to data limitations, potentially elevating the standard of AI-assisted endoscopic surveillance.