Human–machine interaction in optical polyp diagnosis: decision-making after CADx polyp diagnosis

Abstract Background Artificial intelligence (AI)-based computer-aided diagnosis (CADx) systems are increasingly used to classify colorectal polyps during colonoscopy, but their impact depends on how endoscopists interpret AI predictions. This study aimed to evaluate endoscopists’ decision-making after reviewing CADx outputs, using histopathology diagnosis as ground truth. Patients and methods We performed a secondary analysis of a prospective study (NCT06822816) conducted at our centre between 2022 and 2025. All patients and polyps ≤10 mm with documented CADx output and endoscopist diagnosis were included. Endoscopists either accepted CADx classifications (neoplastic vs hyperplastic) or provided alternatives. Our primary outcome was the accuracy of accepted CADx diagnoses; secondary outcomes assessed reclassification patterns after CADx output rejection. Results Our cohort included 754 polyps <10 mm from 367 patients. Accepted CADx diagnoses were correct for 89.1% of neoplastic and 68.7% of hyperplastic CADx predictions. Endoscopists accepted incorrect CADx classifications more frequently when these were neoplastic compared to hyperplastic (64.9% vs 26.5%). Upon rejecting correct CADx neoplastic predictions, endoscopists misclassified 23.4% as sessile serrated lesions (SSLs). Similarly, 29.5% of rejected correct CADx hyperplastic predictions were incorrectly reassigned as SSLs. Conclusions Endoscopists exhibit a diagnostic bias toward neoplasia, particularly misclassifying hyperplastic polyps as SSLs. While this may enhance safety, it may risk unnecessary surveillance colonoscopies and increase costs.