Artificial intelligence-assisted endoscopic ultrasound diagnosis of esophageal subepithelial lesions

• DOI: 10.1007/s00464-025-11767-5
• Publication Date: 2025-05-07T16:57:24Z
• Authors: Ai-meng Zhang; Dai-min Jiang; Shu-peng Wang; Wen Liu; Bei-bei Sun; Zhe Wang; Guo-yi Zhou; Yao-fu Wu; Qing-yun Cai; Jin-tao Guo; Si-yu Sun
• Keywords: Endoscopic Ultrasound

Abstract

Background Endoscopic ultrasound (EUS) is one of the most accurate methods for determining the originating layer of subepithelial lesions (SELs). However, the accuracy is greatly influenced by the expertise and proficiency of the endoscopist. In this study, we aimed to develop an artificial intelligence (AI) model to identify the originating layer of SELs in the esophagus and evaluate its efficacy.

Methods A total of 1445 cases of esophageal SELs were used to develop the model. An AI model stemming from YOLOv8s-seg and MobileNetv2 was developed to detect esophageal lesions and identify the originating layer. Two seniors and two junior endoscopists independently diagnosed the same test set.

Results The precision, recall, mean average precision @ 0.5, and F1-score of the AI model were 92.2%, 73.6%, 0.832, and 81.9%, respectively. The overall accuracy of the originating layer recognition model was 55.2%. The F1-scores of the second, third, and fourth layers were 47.1%, 51.7%, and 66.1%, respectively. The accuracy of the AI system in differentiating layers 2 and 3 from four was 76.5% and was similar to that of senior endoscopists (74.9–79.8%, P = 0.585) but higher than that of junior endoscopists (65.6–66.7%, P = 0.045).

Conclusions The EUS-AI model has shown high diagnostic potential for detecting esophageal SELs and identifying their originating layers. EUS-AI has the potential to enhance the diagnostic ability of junior endoscopists in clinical practice.

Graphical abstract