Early detection of esophageal cancer using multi-institutional data: An evaluation of AI detection algorithms based on narrowband and white-light imaging data

Abstract Esophageal cancer is one of the most common cancers worldwide. In particular, esophageal squamous epithelial cell carcinoma is primarily diagnosed at the end stage and has a poor prognosis. For the early diagnosis and detection of esophageal cancer, this study aimed to develop an algorithm for detecting tumors in esophageal endoscopic images using innovative artificial intelligence (AI) technology. White-light images and narrowband image data collected at Gachon University Gil Hospital were used, and lesions were detected by applying the YOLOv5 and RetinaNet detection models. This study evaluated the performance of the model using precision, sensitivity, and the number of misdetections per image as evaluation indicators in a large-scale dataset. Additionally, the generalizability of the model was verified using external data collected from various institutions. The results showed that the AI model exhibited high precision and sensitivity in white-light and narrowband images by analyzing not only polyps but also superficial esophageal cancer, and the RetinaNet model exhibited excellent performance. These results are expected to contribute to a reduction in the misdiagnosis rate by enabling a more precise and rapid diagnosis. This study presents an effective method for detecting esophageal tumors through AI-based esophageal endoscopic image analysis, and it is expected to be helpful in investigating AI models that predict the depth of lesion invasion by incorporating real-time video-based detection and synthesis algorithms. These efforts are expected to contribute significantly to the effective diagnosis and treatment of esophageal cancer, thereby promoting the standardization of medical services.