Performance of artificial intelligence for colonoscopy regarding adenoma and polyp detection: a meta-analysis

Abstract

BACKGROUND AND AIMS One fourth of colorectal neoplasia is missed at screening colonoscopy, representing the main cause of interval colorectal cancer (CRC). Deep learning systems with real-time computer-aided polyp detection (CADe) showed high accuracy in artificial settings, and preliminary randomized clinical trials (RCT) reported favourable outcomes in clinical setting. Aim of this meta-analysis was to summarise available RCTs on the performance of CADe systems in colorectal neoplasia detection. METHODS We searched MEDLINE, EMBASE and Cochrane Central databases until March 2020 for RCTs reporting diagnostic accuracy of CADe systems in detection of colorectal neoplasia. Primary outcome was pooled adenoma detection rate (ADR), Secondary outcomes were adenoma per colonoscopy (APC) according to size, morphology and location, advanced APC (AAPC), as well as polyp detection rate (PDR), Polyp-per-colonoscopy (PPC), and sessile serrated lesion per colonoscopy (SPC). We calculated risk ratios (RR), performed subgroup, and sensitivity analysis, assessed heterogeneity, and publication bias. RESULTS Overall, 5 randomized controlled trials (4354 patients), were included in the final analysis. Pooled ADR was significantly higher in the CADe groups than in the control group (791/2163, 36.6% vs 558/2191, 25.2%; RR, 1.44; 95% CI, 1.27-1.62; p<0.01; I 2:42%). APC was also higher in the CADe group compared with control (1249/2163, 0.58 vs 779/2191, 0.36; RR, 1.70; 95% CI, 1.53-1.89; p<0.01;I 2:33%). APC was higher for <5 mm (RR, 1.69; 95% CI, 1.48-1.84), 6-9 mm (RR, 1.44; 95% CI, 1.19-1.75), and >10 mm adenomas (RR, 1.46; 95% CI, 1.04-2.06), as well as for proximal (RR, 1.59; 95% CI, 1.34-1.88) and distal (RR, 1.68; 95% CI, 1.50-1.88), and for flat (RR: 1.78 95% CI 1.47-2.15) and polypoid morphology (RR, 1.54; 95% CI, 1.40-1.68). Regarding histology, CADe resulted in a higher SPC (RR, 1.52; 95% CI,1.14-2.02), whereas a nonsignificant trend for AADR was found (RR, 1.35; 95% CI, 0.74 – 2.47; p = 0.33; I 2:69%). Level of evidence for RCTs was graded moderate. CONCLUSIONS According to available evidence, the incorporation of Artificial Intelligence as aid for detection of colorectal neoplasia results in a significant increase of the detection of colorectal neoplasia, and such effect is independent from main adenoma characteristics.