In a meta-analysis, the performance of CADx systems was superior to endoscopists, with pooled sensitivity and specificity of 92.3% and 89.8% respectively. The proportion of incorrect predictions could be significantly reduced by 12% with CADx assistance. It was reported that autonomous AI prediction could achieve a 91.5% agreement with the standard pathology-based surveillance intervals. In a prospective study targeting diminutive rectosigmoid polyps, CADx achieved a negative predictive value of 97.6% compared to the histology as gold standard. 82% of lesions were amendable for a leave-in-situ strategy, suggesting a potential to reduce the burden of unnecessary polypectomies. It was estimated that the average colonoscopy cost and annual reimbursement could be reduced by 6.9-18.9% under the 'diagnose-and-leave' strategy.
