Domain specific multimodal large language model for automated endoscopy reporting with multicenter prospective validation - PubMed
3 hours ago
- #Endoscopy Automation
- #Medical Reporting
- #AI in Healthcare
- Developed Report-Angel, an AI system using a multimodal large language model and deep learning for automated endoscopy report generation.
- Trained on 20,617 image-text pairs to draft reports for upper gastrointestinal endoscopies.
- Achieved clinically acceptable report rates: 79.3% internal and 83.3% external cohort.
- Report completeness and accuracy were 88.51% and 78.93%, respectively.
- Lesion-level reporting accuracies ranged from 83.94% to 91.92% across datasets.
- System processes lesions in 1.5 seconds on average, standardizes reports, and reduces workloads.