Artificial intelligence-assisted colonoscopy: A prospective, multicenter, randomized controlled trial of polyp detection

Abstract

Background: Artificial intelligence (AI) assistance has been considered as a promising way to improve colonoscopic polyp detection, but there are limited prospective studies on real-time use of AI systems. Methods: We conducted a prospective, multicenter, randomized controlled trial of patients undergoing colonoscopy at six centers. Eligible patients were randomly assigned to conventional colonoscopy (control group) or AI-assisted colonoscopy (AI group). AI assistance was our newly developed AI system for real-time colonoscopic polyp detection. Primary outcome is polyp detection rate (PDR). Secondary outcomes include polyps per positive patient (PPP), polyps per colonoscopy (PPC), and non-first polyps per colonoscopy (PPC-Plus). Results: A total of 2352 patients were included in the final analysis. Compared with the control, AI group did not show significant increment in PDR (38.8% vs. 36.2%, p = 0.183), but its PPC-Plus was significantly higher (0.5 vs. 0.4, p < 0.05). In addition, AI group detected more diminutive polyps (76.0% vs. 68.8%, p < 0.01) and flat polyps (5.9% vs. 3.3%, p < 0.05). The effects varied somewhat between centers. In further logistic regression analysis, AI assistance independently contributed to the increment of PDR, and the impact was more pronounced for male endoscopists, shorter insertion time but longer withdrawal time, and elderly patients with larger waist circumference. Conclusion: The intervention of AI plays a limited role in overall polyp detection, but increases detection of easily missed polyps; ChiCTR.org.cn number, ChiCTR1800015607.