Computer-Aided Diagnosis of Colorectal Polyps: Clinical Usefulness and Limitations

Computer-aided diagnosis (CADx) systems have emerged as promising tools to support real-time optical characterization of colorectal polyps during colonoscopy. This narrative review critically evaluates their clinical utility and limitations, focusing on two key strategies: "resect and discard" and "leave in situ." While CADx offers potential benefits, such as cost reduction, increased diagnostic consistency, and support for nonexpert endoscopists, its performance in clinical settings remains variable and often below established thresholds by societies like ESGE and ASGE. Key metrics such as positive predictive value, negative predictive value, sensitivity, and specificity fluctuate widely across studies and CADx platforms, influenced by system training data, disease prevalence, and human-AI interactions. Importantly, trust and explainability issues hinder adoption, with studies revealing underutilization of accurate CADx predictions due to clinician skepticism. Additionally, CADx systems struggle to reliably differentiate sessile serrated lesions from hyperplastic polyps, partly due to limitations in histopathological ground truth and data set representation. Cost-effectiveness analyses show promise, but practical implementation is challenged by equipment, regulatory, and training costs. Finally, emerging applications of CADx in predicting invasion depth in colorectal cancer show potential but require more robust validation. Overall, while CADx technologies may enhance diagnostic confidence and aid decision-making, especially for less experienced endoscopists, their widespread clinical integration depends on addressing human-AI interaction challenges, improving system transparency, and refining models to include underrepresented lesion types.