Artificial Intelligence-Based Detection of Colonic Angiodysplasia Can Improve Endoscopic Reporting Consistency.

BACKGROUND AND STUDY AIMS: Colonic angiodysplasia (AD) is often found incidentally during colonoscopy and frequently omitted from reports, despite being a key cause of lower gastrointestinal bleeding. Artificial intelligence (AI) may improve reporting completeness. In this study, we developed and validated an AI system to support AD documentation. METHODS: The system combines a computer vision (CV) model to detect AD in colonoscopy images and videos with a locally deployed open-weight Large Language Model (LLM) to assess whether AD is mentioned in reports. It was externally validated on independent image datasets and prospectively collected videos and used to identify unreported AD cases. RESULTS: The CV model was trained on 10,908 annotated images (986 with AD) from 993 colonoscopies. External validation of 3534 images from 239 procedures showed 94.5% accuracy, 84.4% sensitivity, and 95.9% specificity. The LLM detected AD documentation with 95.8% accuracy, 97.7% sensitivity, and 92.4% specificity. The combined system identified 88 cases where AD was visible but unreported. In 126 full-length videos, 5 cases were physician-reported, while the system detected 23 AD-positive cases. CONCLUSIONS: Our AI system accurately detects AD and cross-checks reports, improving documentation completeness and supporting standardized reporting while preserving privacy through local deployment.