Performance of an artificial intelligence system versus endoscopists for interrogation of scars after piecemeal endoscopic mucosal resection

Abstract Detection and management of recurrence after piecemeal endoscopic mucosal resection (p-EMR) of large colorectal polyps is crucial for preventing post-colonoscopy colorectal cancer. Colonoscopy assisted with artificial intelligence systems for lesion detection (CADe) has shown efficacy in identifying polyps, particularly small lesions (< 10 mm). Our aim was to evaluate sensitivity and negative predictive value of CADe compared with endoscopists in detecting recurrence after p-EMR. Sixty-six pseudo anonymized high-quality videos (25 with recurrence and 15 with previously clipped defects) of > 15 mm polyps post-EMR scars were assessed. Recurrence size was < 10 mm in 19 of 25 cases (76.0%). Six endoscopists—three experts and three non-experts—predicted recurrence (yes/no) with confidence level (high/low). The same videos were analyzed with CAD-e detection mode superimposed. The gold standard was histopathology. Endoscopists had higher sensitivity than CADe (96.0% vs 72.0%; P = 0.03). Experts performed better than non-experts in ruling out recurrence (specificity: 92.7% vs 58.5%; P = 0.001) and accuracy (92.4% vs 72.7%; P < 0.001). In non-clipped defects, endoscopists maintained higher sensitivity than CADe (95.0% vs 65.0%; P = 0.03). Concordance was moderate overall (Kappa 0.57) but substantial among experts (0.70) and when excluding clipped defects (0.61) Endoscopists outperformed CADe in interrogating post-EMR scars. Training CADe with images of post-polypectomy scars may further improve its performance.