Artificial Intelligence for Real-Time Prediction of the Histology of Colorectal Polyps by General Endoscopists

Background: Real-time prediction of histologic features of small colorectal polyps may prevent resection and/or pathologic evaluation and therefore decrease colonoscopy costs. Previous studies showed that computer-aided diagnosis (CADx) was highly accurate, though it did not outperform expert endoscopists. Objective: To assess the diagnostic performance of histologic predictions by general endoscopists before and after assistance from CADx in a real-life setting. Design: Prospective, multicenter, single-group study. (ClinicalTrials.gov: NCT04437615) Setting: 6 centers across the United States. Participants: 1252 consecutive patients undergoing colonoscopy and 49 general endoscopists with variable experience in real-time prediction of polyp histologic features. Intervention: Real-time use of CADx during routine colonoscopy. Measurements: The primary end points were the sensitivity and specificity of CADx-unassisted and CADx-assisted histologic predictions for adenomas measuring 5 mm or less. For clinical purposes, additional estimates according to location and confidence level were provided. Results: The CADx device made a diagnosis for 2695 polyps measuring 5 mm or less (96%) in 1252 patients. There was no difference in sensitivity between the unassisted and assisted groups (90.7% vs. 90.8%; P = 0.52). Specificity was higher in the CADx-assisted group (59.5% vs. 64.7%; P < 0.001). Among all 2695 polyps measuring 5 mm or less, 88.2% and 86.1% (P < 0.001) in the CADx-assisted and unassisted groups, respectively, could be resected and discarded without pathologic evaluation. Among 743 rectosigmoid polyps measuring 5 mm or less, 49.5% and 47.9% (P < 0.001) in the CADx-assisted and unassisted groups, respectively, could be left in situ without resection. Limitation: Decision making based on CADx might differ outside a clinical trial. Conclusion: CADx assistance did not result in increased sensitivity of optical diagnosis. Despite a slight increase, the specificity of CADx-assisted diagnosis remained suboptimal.