Assessing the Role of CADx in Polyp Characterization in Clinical Practice: A Multicenter, Prospective Randomized Controlled Trial

Aims Artificial Intelligence (AI) has shown promising results as a tool in endoscopic practices, particularly through computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems. While CADe has been proven to enhance polyp detection, evidence regarding the effectiveness and clinical utility of CADx systems in real-world practice remains limited. This study aimed to evaluate the comparative effectiveness of AI-assisted optical assessment versus conventional methods for polyp characterisation in real-world settings.

Methods We conducted a 4-arm multicentre randomized controlled trial (RCT) with 1:1 matched patient groups to balance polyp characteristics across the participating hospitals. Patients aged 40 to 90 years undergoing colonoscopy in the Region of Västra Götaland, Sweden, were included. Procedures were performed at university hospitals (Sahlgrenska University Hospital, Sahlgrenska Östra, Sahlgrenska Mölndal) and community hospitals (Angereds Närsjukhus, Alingsås Sjukhus, Falköpings Sjukhus, Sjukhuset i Lidköping, and Skövdes Kärnsjukhus). Exclusion criteria included a history of inflammatory bowel disease (IBD), contraindications for polypectomy, or known polyps.

Results A total of 1,096 patients with 1,654 polyps were included. Patients were divided into three groups: CADx (328 patients, 542 polyps) CADe (248 patients, 369 polyps), and Conventional Colonoscopy (CC) (520 patients, 674 polyps). Before analysis, 27 polyps were excluded due to specific histological characteristics. Additionally, 17% (278/1,654) of polyps lacked histology due to resect-and-discard strategy. For per-polyp analysis, 408 polyps were included in the CADx group, 332 in the CADe group, and 609 in the CC group. The overall accuracy of optical diagnosis was 77.3% for CADx, 71.4% for CADe, and 80.7% for CC. Cohen’s kappa values were 0.66 (CADx), 0.604 (CADe), and 0.764 (CC). Sensitivity was 78.7% (CADx), 84.3% (CADe), and 90.7 (CC), while specificity was 43.8% (CADx), 46.9% (CADe) and 58.3% (CC). PPV were 0.888 (CADx), 0.750 (CADe), and 0.830 (CC), while NPV were 0.567 (CADx), 0.613 (CADe), and 0.736 (CC). No statistically significant improvement was observed between the CADx and CADe systems compared to their respective control groups.

Conclusions This large multicentre RCT found that CADx systems do not improve the accuracy of polyp characterization in real-life settings. Further development is needed to enhance the performance of CADx systems before can be fully integrated into everyday clinical practice.

Publication History

Article published online:
27 March 2025