Accuracy of artificial intelligence on histology prediction and detection of colorectal polyps: a systematic review and meta-analysis

Abstract

Background and aims: We performed a meta-analysis of all published studies to determine the diagnostic accuracy of artificial intelligence (AI) on histology prediction and detection of colorectal polyps.

Method: We searched Embase, PubMed, Medline, Web of Science, and Cochrane library databases to identify studies using AI for colorectal polyp histology prediction and detection. The quality of included studies was measured by the Quality Assessment of Diagnostic Accuracy Studies tool. We used a bivariate meta-analysis following a random-effects model to summarize the data and plotted hierarchical summary receiver operating characteristic curves. The area under the hierarchical summary receiver operating characteristic curve (AUC) served as an indicator of the diagnostic accuracy and during head-to-head comparisons.

Results: A total of 7680 images of colorectal polyps from 18 studies were included in the analysis of histology prediction. The accuracy of the AI (AUC) was .96 (95% confidence interval [CI], .95-.98), with a corresponding pooled sensitivity of 92.3% (95% CI, 88.8%-94.9%) and specificity of 89.8% (95% CI, 85.3%-93.0%). The AUC of AI using narrow-band imaging (NBI) was significantly higher than the AUC using non-NBI (.98 vs .84, P < .01). The performance of AI was superior to nonexpert endoscopists (.97 vs .90, P < .01). For characterization of diminutive polyps using a deep learning model with nonmagnifying NBI, the pooled negative predictive value was 95.1% (95% CI, 87.7%-98.1%). For polyp detection, the pooled AUC was .90 (95% CI, .67-1.00) with a sensitivity of 95.0% (95% CI, 91.0%-97.0%) and a specificity of 88.0% (95% CI, 58.0%-99.0%).

Conclusions: AI was accurate in histology prediction and detection of colorectal polyps, including diminutive polyps. The performance of AI was better under NBI and was superior to nonexpert endoscopists. Despite the difference in AI models and study designs, AI performances are rather consistent, which could serve as a reference for future AI studies.