Artificial Intelligence-Assisted Detection and Characterization of Colorectal Polyps during Colonoscopy: Evidence and Future Perspectives

Colonoscopy is central to colorectal cancer prevention because it enables early detection and removal of premalignant lesions. However, lesions can still be missed, and endoscopic differentiation of neoplastic from nonneoplastic polyps remains variable, particularly for diminutive lesions. The potential of artificial intelligence (AI) as a supportive technology for improving both polyp detection and characterization during colonoscopy has therefore been investigated. Since early studies of convolutional neural network-based computer-aided detection (CADe), the focus has shifted from retrospective image-based validation to real-time randomized trials, meta-analyses, and clinical guideline discussions. Accumulating evidence suggests that CADe improves surrogate detection outcomes, including the adenoma detection rate (ADR), polyp detection rate, adenomas per colonoscopy, and adenoma miss rate. Although a higher ADR is associated with a lower risk of postcolonoscopy colorectal cancer, it remains unknown whether the incremental improvement in the ADR achieved by CADe translates into lower postcolonoscopy colorectal cancer incidence or colorectal cancer-related mortality. Recent multicenter data further support the effectiveness of real-time CADe systems in improving detection-related outcomes, while also indicating that benefit varies with lesion characteristics and user experience. In parallel, computer-aided diagnosis (CADx) has been developed to assist the optical characterization of colorectal polyps, particularly for diminutive lesions and potential resect-and-discard or diagnose-and-leave strategies. CADx systems using white-light imaging and structured high-confidence diagnostic rules show promise, but the clinical implementation of CADx remains less mature than that of CADe. This review summarizes current evidence for AI-assisted polyp detection and characterization and discusses limitations and future directions for integrating CADe and CADx into human-centered colonoscopy practices.