Cloud-based Artificial Intelligence for Detection of Colorectal Neoplasia – A Randomized Controlled Trial (EAGLE Trial)

Aims In recent years, several Randomized Controlled Trials (RCTs) have evaluated the role of artificial intelligence (AI) Computer-Aided Detection (CADe) systems to enhance polyp detection during colonoscopy. However, these studies have focused exclusively on locally deployed hardware-based AI systems, whereas cloud-based systems may improve the accessibility and upgrading of CADe technology. ‘CADDIE’ (Odin Vision, an Olympus company) is a cloud-based CADe system that receives the live endoscopy image stream via the hospital’s internet network, where the AI algorithm processes the images and outputs in real-time overlay bounding boxes on the endoscopy screen to highlight suspected polyps. This prospective RCT aimed to evaluate the efficacy and safety of the novel cloud-based CADe-system (‘CADDIE’) in improving polyp detection.

Methods This international, multi-centre, parallel-group RCT was conducted across eight centres in four European countries. Participating endoscopists had a baseline of>1000 colonoscopy procedures and an adenoma detection rate (ADR) of≥25%. Patients aged≥40 years and at average risk for colorectal cancer (CRC) undergoing screening or surveillance indications for colonoscopy were recruited. Patients were randomised in a 1:1 ratio to standard colonoscopy (SoC) or CADDIE-assisted (CADe-arm) colonoscopy. The co-primary endpoints were superiority in adenomas per colonoscopy (APC) as the efficacy endpoint, and non-inferiority for the Positive Percent Agreement (PPA), defined as the percentage of resections that are histologically confirmed adenomas, SSLs and large (> 10mm) hyperplastic polyps of the proximal colon, as the safety endpoint.

Results Overall, 22 endoscopists and 841 participants were included in the analysis. The CADe-arm significantly increased APC by 33% (0.62 vs. 0.82, Ratio 1.33 [95%CI 1.06-1.67], p=0.01) and ADR (35.9% vs. 43.2%, difference of 7.3% [95%CI 0.7%-13.9%], p=0.03) compared to the SoC-arm. The mean detection of sessile serrated lesions (SSLs) (0.03 vs. 0.08, Ratio 3.30 [95%CI 1.41-7.57], p<0.01), large adenomas (0.04 vs. 0.09 vs; Ratio 1.93 [95%CI 1.03-3.62], p=0.04), and large polyps (0.05 vs. 0.12 vs; Ratio 2.36 [95%CI 1.33-4.17], p<0.01) also significantly improved. A higher proportion of large polyps (27.3% vs. 37.0%) and large adenomas (15.8% vs. 24.2%) were non-polypoid ‘flat’ morphology in the CADe-arm compared to the SoC-arm. Other detection metrics, including neoplastic polyps per colonoscopy (0.66 vs. 0.91; Ratio 1.39 [95%CI 1.12-1.73], p<0.01 and polyps per colonoscopy (1.06 vs. 1.41 (Ratio 1.35 [95%CI 1.13-1.61], p<0.01), were significantly higher in the CADe-arm compared to the SoC-arm. Non-inferiority in PPA was achieved (53.4% vs 53.9%; 0.5% (-5.0%,∞)), with no increase in unnecessary resections.

Conclusions This trial demonstrates the safety and efficacy of a cloud-based CADe-system (CADDIE) to increase endoscopists' polyp detection across multiple metrics, most notably for large adenomas, large polyps, and SSLs.

Publication History

Article published online:
27 March 2025