CADe Systems Show Variable Improvements in Adenoma Detection During Colonoscopy in Meta-Analysis

This systematic review and network meta-analysis synthesized evidence from 48 randomized controlled trials (RCTs) involving a total of 38,986 patients undergoing colonoscopy. The specific clinical setting was not reported. The analysis aimed to compare the performance of different artificial intelligence-based computer-aided detection (CADe) systems for colorectal polyp detection. The population consisted of patients undergoing colonoscopy, though specific inclusion or exclusion criteria for the underlying trials were not detailed in the provided data.

The intervention was colonoscopy performed with one of three CADe systems: ENDO-AID, CADEYE, or GI Genius. The comparator was standard colonoscopy performed without any CADe assistance. The specific procedural protocols, including bowel preparation standards or endoscopist experience levels, were not reported. The analysis treated each CADe system as a distinct node in a network meta-analysis to allow for indirect comparisons between them, with all systems compared against the common control of standard colonoscopy.

The primary outcome was the adenoma detection rate (ADR). All three CADe systems showed improved ADR compared to controls. For the ENDO-AID system, the risk ratio (RR) was 1.26 (95% credible interval [CrI] 1.14 to 1.40). For the CADEYE system, the RR was 1.18 (95% CrI 1.10 to 1.26). For the GI Genius system, the RR was 1.15 (95% CrI 1.08 to 1.22). The credible intervals for all systems did not cross 1.0, indicating statistical significance. Absolute numbers for ADR in each group were not reported, limiting the interpretation of the clinical magnitude of benefit.

A key secondary outcome was the sessile serrated lesion detection rate (SSLDR). Data for this outcome were available for only two systems. The ENDO-AID system showed an RR of 1.36 (95% CrI 1.03 to 1.79) for SSLDR compared to control. The GI Genius system showed an RR of 1.25 (95% CrI 1.08 to 1.46). The CADEYE system's performance for SSLDR was not reported. As with the primary outcome, absolute detection rates were not provided.

No safety or tolerability findings were reported in this meta-analysis. Rates of adverse events, serious adverse events, procedural complications, or study discontinuations were not provided. The analysis focused solely on efficacy outcomes (detection rates), leaving a gap in the understanding of the risk profile associated with implementing these CADe systems in routine practice.

These results align with and extend the findings of prior individual RCTs and meta-analyses that have generally shown a benefit for CADe in improving ADR. The novel contribution of this network meta-analysis is the head-to-head comparison suggesting variable effect sizes between different commercially available systems, with ENDO-AID appearing to have the largest point estimate for ADR improvement. However, the overlapping credible intervals indicate that these differences between systems may not be statistically significant.

Key methodological limitations must be considered. The certainty of the evidence, as assessed by the CINeMA framework, ranged from low to moderate. The analysis did not report on several important limitations common to meta-analyses, such as heterogeneity between included trials, publication bias, or the quality of individual RCTs. The lack of absolute outcome numbers limits the ability to calculate the number needed to treat to find one additional adenoma. Furthermore, the analysis does not address whether improved detection translates into improved long-term clinical outcomes, such as reduced colorectal cancer incidence or mortality.

The clinical implications are that CADe systems represent a promising adjunct technology that may enhance the quality of colonoscopy. The finding of variable effect sizes suggests that not all AI systems may be equivalent, though direct comparative trials are needed to confirm this. For practices considering adoption, the choice may depend on factors beyond efficacy, including cost, integration with existing endoscopy platforms, and real-world usability, none of which were addressed here. The improvement in SSLDR for some systems is particularly noteworthy given the challenges in visually identifying these lesions.

Significant questions remain unanswered. The long-term impact of increased adenoma detection on cancer prevention is unknown. The cost-effectiveness of implementing these systems has not been established. The performance in specific high-risk subgroups or in the hands of endoscopists with varying baseline ADRs is unclear. Finally, whether the use of CADe leads to increased resection of non-neoplastic lesions (increased false positives) or unnecessarily prolongs procedure time was not reported and requires investigation.