S572 Deep Learning and High Resolution Anoscopy: First Trinary Interoperable Model for Detecting and Differentiating Clinically Relevant Anal Lesions

Résumé fourni par la source

Introduction: Predicting the histology of colon polyps during colonoscopy using real-time visual assessment has clinical relevance. Accurate characterization guides immediate clinical decisions, like the need for polypectomy, and reduces unnecessary procedures, risks and costs. Traditionally, polyps are classified by histopathology, but emerging tools, including artificial intelligence, offer real-time, non-invasive characterization. Experienced gastroenterologists can make accurate predictions without these tools. The purpose of this study was to determine the accuracy of visual prediction of colon polyp pathology in a community based colonoscopy setting. Methods: Consecutive patients undergoing screening or surveillance colonoscopy at a community hospital with identified polyps over 2 months were included. Visually predicted pathology on white light (WL) and narrow band imaging (NBI) was recorded. Histopathology results were then reviewed to confirm polyp pathology. Predicted and confirmed pathologies were compared. Colonoscopies were performed with Olympus PCF HQ190L colonoscopes. Results: The 197 patients had a total of 238 polyps. Final pathology classification included hyperplastic polyps, tubular adenomas, and sessile serrated adenomas. WL and NBI predictions were compared to histopathology. Identification of tubular adenomas had AUROC values of 0.92 (NBI) and 0.91 (WL) and sensitivity/specificity of 0.91/0.92 (NBI) and 0.88/0.94 (WL). For hyperplastic polyps, AUROCs were 0.88 (NBI) and 0.87 (WL), with sensitivity/specificity of 0.81/0.95 (NBI) and 0.82/0.93 (WL). Sessile serrated adenomas were identified with area under the receiver operating characteristic curves (AUROCs) of 0.84 (NBI) and 0.78 (WL) and sensitivity/specificity of 0.73/0.94 (NBI) and 0.67/0.92 (WL). Chi-square analysis confirmed statistical significance between both modalities and final pathology (P < 0.001 for WL and NBI). Conclusion: These results demonstrate that experienced gastroenterologists can use real-time visual assessment to achieve high diagnostic accuracy in polyp characterization across all major histological types. When compared to recent AI-based systems, which report AUROC values between 0.83–0.94, these results show that trained endoscopists perform on par with or even exceed the precision of machine learning tools. This study reinforces the enduring value of endoscopic expertise and supports the continued emphasis on training and visual recognition skills among gastroenterologists.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.