Generative adversarial networks for balancing and expanding data resources for computer-aided detection in colonoscopy.

Background and aims: Developing computer-aided detection (CADe) algorithms in colonoscopy requires high-quality databases sourced from real-world data. However, such databases demand costly investments and display imbalanced distributions that underrepresent flat adenomas. Therefore, we evaluated whether StyleGAN2-ADA, a generative adversarial network, can generate synthetic images as alternative training resources for CADe, and whether a modified StyleGAN2-ADA can enhance synthesis control for better balanced datasets. Methods: Two synthetic datasets were generated using the original StyleGAN2-ADA and our modified version with feature-clustered conditioning vectors. Generative adversarial networks were trained on images from our local colonoscopy database, and their outputs were evaluated through Fréchet Inception Distance where lower scores indicate more realistic and diverse data. Three CADe models were trained using synthetic or real-world data and tested in two external databases comprising only polyp images. CADe performance included mean average precision (mAP) score as critical metric for polyp identification. Results: The modified StyleGAN2-ADA achieved a lower Fréchet Inception Distance-score (7.54 vs. 13.68) than the original StyleGAN2-ADA. During external testing, CADe trained on synthetic data from the modified version achieved performance comparable to the model trained on real-world data across all metrics ( Conclusions: StyleGAN2-ADA incorporating feature-clustered conditioning vectors can synthesize better balanced colonoscopy databases that offer alternative training resources to real-world data for CADe training and validation.