Prediction of the Histology of Ampullary Adenoma Using Deep Learning Algorithm

초록

Introduction: Ampullary adenoma (AA) is a rare gastrointestinal neoplasm arising from the major papilla. Due to its premalignant potential, AA must be completely removed, either endoscopically or surgically, with treatment decisions guided by histology and depth of invasion. Endoscopic biopsy is usually performed before the removal, but diagnostic discrepancy is reported up to 30%. We developed a convolutional neural network (CNN)–based AI model to aid in the diagnosis of ampullary adenoma. Methods: We retrospectively collected images of ampullary adenoma in four tertiary hospitals: Hanyang University Hospital, Korea University Anam, Guro, and Ansan Hospitals. All images collected were taken using duodenoscope with en face view of the ampulla. Lesions were classified as normal, adenoma, or carcinoma. Normal cases were defined by consensus of two expert endoscopists and stable appearance for over 1 year. Adenomas were confirmed through complete resection and pathology; carcinomas were identified by any malignant histologic findings. To improve diagnostic robustness, only cases with at least three high-quality images were included. Results: We developed a CNN-based computer-aided diagnostic (CAD) system using an automated AI platform. The model achieved an accuracy of 81.3%, precision of 79.4%, recall of 78.9%, and F1 score of 79.4%. The system outperformed trainee endoscopists and showed comparable or superior performance to expert endoscopists in selected cases. Class activation mapping highlighted image regions critical for decision-making (Figure 1). Conclusion: Our CNN-CAD system demonstrated reliable histologic classification of ampullary lesions and may serve as an effective adjunct to endoscopic diagnosis, potentially improving clinical outcomes.