AI Optical Diagnosis of Diminutive Colorectal Polyps for Resect-and-Discard and Diagnose-and-Leave Strategies

Abstract Diminutive colorectal polyps are frequently encountered during colonoscopy, yet their individual risk of advanced histology or invasive cancer is low. This has prompted resect-and-discard and diagnose-and-leave strategies aimed at reducing unnecessary histopathology. Artificial intelligence-assisted computer-aided diagnosis (AI-CADx) may standardize optical diagnosis, but evidence ranges from selected retrospective image analyses to real-time multicenter studies and a single strategy-level randomized trial. Diagnostic performance is often high in controlled settings; however, prospective studies show clinically important abstention rates, poor specificity, weak recognition of sessile serrated lesions, and little incremental benefit over competent endoscopists. Two recent meta-analyses found no material increase in pathology avoidance or reduction in the number of neoplastic polyps erroneously left in situ when CADx was added to endoscopist assessment. The evidence for resect-and-discard is more mature than that for diagnose-and-leave, for which an incorrect nonneoplastic prediction has a greater clinical consequence. AI-CADx should therefore remain an adjunct rather than an autonomous authority and should be restricted to well-visualized lesions within the validated scope of a system. Lesions with low-confidence or discordant assessments, as well as proximal, serrated-appearing, or otherwise atypical lesions, should be resected and submitted for histopathology. Safe implementation also requires training, documentation, audit, confidence calibration, and prospective monitoring of clinically relevant outcomes. This narrative review critically evaluates the readiness of AI-CADx for the management of diminutive colorectal polyps and identifies the evidence required for wider adoption.