JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) systems designed to enhance polyp detection during colonoscopy have shown promise in clinical trials, but the extent to which various systems improve detection of polyps of different sizes remains unclear.

Objective:

The objective of this article is to compare efficacy of various AI-assisted systems during colonoscopy.

Methods:

We searched PubMed, Cochrane Library and Web of Science for randomized controlled trials (RCTs) comparing AI-assisted colonoscopy with white-light colonoscopy, then literature screening was conducted and network meta-analysis (NMA). NMA was conducted using StataMP 14 and Review Manager 5.3 was used for the quality assessment of included studies.

Results:

NMA of 13 RCTs (4,156 participants) revealed marked size-dependent efficacy of AI-assisted systems for increasing mean polyp detection. For diminutive polyps, EndoScreener ranked first [SUCRA 77.7%, probability of being best 33.3%] and significantly increased mean polyp count versus standard colonoscopy (SMD 1.22, 95%CI 0.08 to 2.36). For small polyps (6-9mm), standard colonoscopy unexpectedly achieved highest ranking (SUCRA 91.5%), with AI systems showing minimal incremental value. For large polyps (≥10mm), CAD EYE and standard colonoscopy performed comparably (SUCRA 78.9% vs 80.0%), with no AI system demonstrating significant improvement. No head-to-head trials directly compared AI systems, indirect comparisons revealed no differences between AI platforms.

Conclusions:

Current evidence supports size-dependent efficacy of AI-assisted colonoscopy, with maximal benefit for diminutive polyps (≤5mm) and diminishing returns as polyp size increases. EndoScreener demonstrates the most consistent evidence for improving diminutive lesion detection. These findings support targeted AI implementation for screening populations where small lesion detection impacts surveillance intervals. Direct comparative trials between AI systems are needed to establish optimal platform selection. Clinical Trial: The protocol was registered with PROSPERO, registration number: CRD420251266932.