Development and Validation of an Automatic Image-Recognition Endoscopic Report Generation System: A Multicenter Study

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Conventional gastrointestinal (GI) endoscopy reports are time-consuming and prone to heterogeneity or omissions.
• This variability can hinder efficiency and multicenter collaboration in GI diagnostics.

Purpose of the Study:

• To develop and validate an image recognition-based structured report generation system (ISRGS).
• To assess the diagnostic performance of ISRGS using a multicenter database.

Main Methods:

• Developed an ISRGS integrating real-time video capture, site identification, lesion detection, and structured report generation.
• Trained and validated the system on 46,987 images from 9 hospitals, with further prospective testing on 5,699 images.

Main Results:

• Achieved high overall accuracy in diagnosing various GI lesions: 0.8841 (multicenter) and 0.8965 (prospective).
• Demonstrated excellent accuracy in site identification (cecum: 0.9978, upper GI: 0.8513) and staining discrimination (0.9489).
• Reported a low relative error of 4.04% for size measurement.

Conclusions:

• ISRGS is a reliable computer-aided system for generating standardized and accurate endoscopy reports.
• The system can assist endoscopists at all hospital levels, enhancing diagnostic consistency and efficiency.