A recent study published in [n]npj Digital Medicine[/n] introduces an advancement in the application of artificial intelligence for endoscopy reporting. A research team has developed the Report-Angel system, which integrates a Multimodal Large Language Model with deep learning techniques to automatically generate detailed draft reports for upper gastrointestinal endoscopy examinations. Trained on 20,617 image-text data pairs, the system completed a multi-center prospective validation, achieving clinical acceptance rates from 79.3% to 83.3% and processing lesions in just 1.5 seconds.
Practical Challenges in Endoscopy Reporting
Upper gastrointestinal endoscopy is a key procedure for diagnosing and managing diseases of the upper digestive tract, and accurate endoscopy reports are crucial for patient diagnosis and treatment. However, errors and omissions in reports are common. The routine process of preparing reports is time-consuming and laborious, adding to the daily workload of physicians.
Report-Angel System: Development and Training
Researchers built the Report-Angel integrated AI system based on a Multimodal Large Language Model (MLLM) and conventional deep learning models. The system was trained on 20,617 pairs of endoscopy images and corresponding text, enabling it to automatically analyze images and generate professionally formatted text reports.
Core Results from Multi-Center Prospective Validation
In the prospective internal cohort, Report-Angel’s clinical acceptance rate for reports was 79.3% (95% CI: 74.4-83.5%). In the external cohort, this rate reached 83.3% (95% CI: 78.7-87.3%). At the case level, report completeness was 88.51% (95% CI: 84.64-92.38%) and accuracy was 78.93% (95% CI: 73.98-83.88%), with an average processing time of just 1.5 seconds per lesion.
Lesion-Level Accuracy Assessment
Lesion-level report accuracy reached 91.92% (95% CI: 90.58-93.25%) on the retrospective image dataset, 89.07% (95% CI: 87.57-90.57%) on the prospective single-center video dataset, and 83.94% (95% CI: 81.58-86.31%) on the multi-center video dataset. These results confirm the system’s high performance across different data scenarios.
Summary of System’s Report Generation Capability
Report-Angel can generate expert-level draft endoscopy reports and has demonstrated robust generalization capabilities. The system provides a reliable foundation for standardizing the endoscopy reporting workflow.
