Clinical implications of computer-aided real-time size estimation of colorectal polyps during colonoscopy

BACKGROUND Accurate polyp size estimation during colonoscopy is crucial for clinical decision-making, follow-up, and cost-saving strategies. Objective sizing methods are lacking, and interobserver variability is high. This prospective, multicenter, study evaluated the accuracy of a novel artificial intelligence (AI) algorithm for polyp size estimation. METHODS Subjects aged ≥18 undergoing colonoscopy for CRC screening or surveillance were enrolled across three centers. Polyp size was initially assessed by operators using forceps/snare comparison (ground truth). Procedures were recorded, and AI-based polyp size estimates were obtained offline. The primary endpoint was AI accuracy in size class determination (diminutive ≤5 mm, small 6–9 mm, large ≥10 mm). Secondary endpoints included size estimation in mm and its impact on clinical management strategies. RESULTS Among 465 polyps (307 diminutive, 107 small, 51 large) from 225 patients (51.6% female, mean age 61.9 ± 10.4 years), AI accuracy for size class determination was 85.8% [95%CI: 82.6–88.8]. Accuracy for diminutive, small, and large polyps was 93.3%, 74.6%, and 55.1%, respectively. The AI tool assigned 90.8% of patients to correct surveillance intervals and achieved a MAE of 1.13 mm and RMSE of 1.40 mm for polyps ≤10 mm. CONCLUSIONS The AI model performs similarly to expert endoscopist in clinically relevant size-related endpoints, potentially improving the accuracy and efficiency of colorectal cancer screening.