AI ANALYSIS: MEDICAL DIAGNOSTICS

The results demonstrate Report-Angel's strong performance across various metrics. In prospective internal validation, 79.3% of reports were clinically acceptable, rising to 83.3% in the external cohort. Case-level completeness was 88.51% and accuracy was 78.93%. Lesion-level accuracy ranged from 83.94% to 91.92% across datasets, with high reliability for critical findings like high-risk lesions. Processing time averaged 1.5 seconds per lesion, allowing for rapid report generation. Subjective evaluations showed Report-Angel's reports were comparable to senior endoscopists and superior to junior ones in accuracy and completeness.

The discussion contextualizes Report-Angel within the evolving landscape of AI in medical imaging, emphasizing its novel hybrid architecture and ability to generate coherent narrative reports. It underscores the system's role as an assistive tool for clinicians, enhancing efficiency and standardization. Error analysis provides insights for future refinement, particularly in balancing sensitivity for neoplastic lesions against false positives. The study acknowledges limitations, including dataset size and the subjective nature of clinical acceptability, while laying groundwork for broader MLLM application in gastroenterology.

The methods section details the observational study design, including data collection from single- and multi-center retrospective and prospective datasets (RMI-Train, RMI-Test, PMV, PSV, PEV). It describes the data annotation process by expert endoscopists to create gold standards and the system construction, integrating deep learning models for video processing, lesion detection, anatomical landmark identification, and a fine-tuned MLLM (Model 6) for generating descriptions and diagnoses. The prospective validation design and statistical analysis approach are also outlined, ensuring robust evaluation of clinical utility and performance.