Computer-Aided Diagnosis Using the Large-Scale Visual Language Models in Screening of Digestive Endoscopy

China's response to the high prevalence of digestive tract diseases involves conducting large-scale endoscopic screening programs. However, the substantial diagnostic workload generated by these programs highlights a low staff-to-patient ratio. To address this challenge, we leverage artificial intelligence (AI) to assist in the computer-aided diagnosis (CAD) of endoscopic images. Specifically, we propose a CAD method that utilizes pre-trained large-scale vision-language models to support digestive endoscopy screening. This approach enables effective classification through the fusion of visual and textual information channels, making it particularly suitable for analyzing large-scale endoscopic datasets. Our model is trained on multicenter data collected from four hospitals, comprising 41,191 images that span 19 disease categories and four normal conditions. In comparative evaluations, our model outperforms eight state-of-the-art models. Furthermore, in a human-AI collaborative reading experiment, the model demonstrates superior diagnostic performance and faster decision-making times compared to six experienced endoscopists. These results underscore the model's effectiveness in CAD for digestive endoscopy. Despite certain limitations, this study highlights our continued efforts to advance AI-assisted diagnostics and envisions a future where AI-driven endoscopic diagnosis becomes a clinical reality.