A deep learning-based, real-time image report system for linear EUS.

Abstract

Background and aims: The integrity of image acquisition is critical for biliopancreatic EUS reporting, significantly affecting the quality of EUS examinations and disease-related decision-making. However, the quality of EUS reports varies among endoscopists. To address this issue, we developed a deep learning-based EUS automatic image report system (EUS-AIRS), aiming to achieve automatic photodocumentation in real-time during EUS, including capturing standard stations, lesions, and puncture procedures.

Methods: Eight deep learning models trained and tested using 235,784 images were integrated to construct the EUS-AIRS. The performance of EUS-AIRS was tested through man-machine comparisons at 2 levels: a retrospective test (include internal and external testing) and a prospective test. From May 2023 to October 2023, a total of 114 patients undergoing EUS at Renmin Hospital of Wuhan University were consecutively recruited for prospective testing. The primary outcome was the completeness of the EUS-AIRS for capturing standard stations.

Results: In terms of completeness in capturing biliopancreatic standard stations, EUS-AIRS exceeded the capabilities of endoscopists at all levels of expertise in retrospective internal testing (90.8% [95% confidence interval (CI), 88.7%-92.9%] vs 70.5% [95% CI, 67.2%-73.8%]; P < .001) and external testing (91.4% [95% CI, 88.4%-94.4%] vs 68.2% [95% CI, 63.3%-73.2%]; P < .001). EUS-AIRS exhibited high accuracy and completeness in capturing standard station images. The completeness of the EUS-AIRS significantly outperformed manual endoscopist reports (91.4% [95% CI, 89.4%-93.4%] vs 78.1% [95% CI, 75.1%-81.0%); P < .001).

Conclusions: EUS-AIRS exhibits exceptional capabilities in real-time, capturing high-quality and high-integrity biliopancreatic EUS images. This showcases the potential of applying an artificial intelligence image report system in the EUS field.