Artificial Intelligence-Assisted Flexure and Landmark Identification Improves Reporting Completeness in Colonoscopy

Introduction Colorectal cancer is a major global health concern, with colonoscopy as the gold standard for detection and prevention. Accurate localization of findings remains challenging, limiting reporting objectivity and completeness, which can be supported by artificial intelligence (AI). We present an AI system that identifies the appendiceal orifice, ileocecal valve, and flexures, enabling automated colon segment localization. Methods The AI was trained on 7264 manually annotated images from 991 patients and internally evaluated on 1238 images from 215 patients. Performance was assessed using accuracy, sensitivity, specificity, and F1-score. External validation was conducted on 78 public videos. The impact on standardized reporting was evaluated on 14 videos recorded in a different external clinic that did not contribute training data. These were independently evaluated by two expert endoscopists who reached a consensus that was used as the gold standard for flexure detection. Results On the internal image test set, the AI achieved 94.5% accuracy, 76.7% sensitivity, 97.0% specificity, and 0.79 F1-score. In external validation on videos from a public dataset, the AI identified 91.7% of cecum segments and 66% of flexures, with 81.4% detected within 30 s of annotation. AI integration improved flexure identification in the external video dataset from the second clinic, increasing detected flexures by 53.8% and videos with both flexures identified by 133.3%. Conclusions The AI reliably detects key anatomical landmarks, supporting colon segment localization. Its integration into reporting pipelines could enhance lesion localization, improve report completeness, and reduce documentation workload.