Evaluating StyleGAN2-ADA and f-AnoGAN for GAN based unsupervised anomaly detection in gastrointestinal endoscopy

Manual interpretation of gastrointestinal (GI) endoscopy images is a subjective and time-intensive process, often leading to variability between clinicians and the risk of missed abnormalities. To address these challenges, we investigate the use of StyleGAN2-ADA as a generative model for unsupervised anomaly detection in gastrointestinal endoscopy images. StyleGAN2-ADA combines adaptive discriminator augmentation with style-based generative modelling, enabling the model to capture both global anatomical structures and fine-grained mucosal textures. In this work, the performance of StyleGAN2-ADA is empirically compared with the widely used f-AnoGAN framework to assess its suitability for anomaly detection in gastrointestinal endoscopy images. The model was trained and evaluated on the HyperKvasir dataset, a large and diverse collection of endoscopic images representing both normal and pathological findings. Our method achieved a ROC-AUC of 0.945 and an F1 score of 0.847, outperforming the established f-AnoGAN baseline in anomaly discrimination. These results highlight the ability of StyleGAN2-ADA to model complex multimodal distributions of healthy tissue and to emphasize abnormal regions as reconstruction defects. Clinically, this framework has the potential to support early detection of GI diseases and to integrate into real-time endoscopy workflows, reducing cognitive load for physicians, assisting less experienced clinicians, and accelerating biopsy decision-making.