Clinical Validation of a Generative AI System for Diagnosing Ampullary Lesions: A Multicenter Study

Abstract

Background Accurate histologic classification of ampullary lesions is essential for guiding therapeutic decisions; however, conventional biopsy is limited by sampling error and false-negative results. We evaluated the clinical utility of a generative artificial intelligence (AI)-based computer-aided diagnosis (CAD) system that integrates real and synthetic endoscopic images to improve diagnostic performance. Methods In this retrospective multicenter study conducted across seven hospitals, duodenoscopic images were classified as Normal, Adenoma, or Cancer. A generative AI-based CAD system using latent diffusion synthesized 500 images per class for data augmentation and was trained to predict histologic classes from endoscopic images. External validation assessed accuracy, sensitivity, specificity, positive and negative predictive values, and area under the receiver operating characteristic curve. A reader study involving five expert and five trainee endoscopists compared diagnostic performance with and without CAD assistance. Results The generative AI-based CAD system demonstrated high overall diagnostic accuracy (91.57%) and strong performance in identifying adenomas (accuracy, 88.76%). In the reader study, CAD assistance significantly increased adenoma sensitivity (63.47%-70.56%; p < 0.01), with corresponding improvements in predictive values. Both expert and trainee endoscopists benefited from CAD support, with reduced interobserver variability and consistent improvements across Normal, Adenoma, and Cancer classifications. Conclusions A generative AI-based CAD system improved diagnostic accuracy and consistency in the evaluation of ampullary lesions. These findings support its potential as a clinically useful adjunct to routine duodenoscopy, particularly for improving adenoma recognition and supporting therapeutic decision-making.