Objective: Adenoma detection rate (ADR) is the strongest quality indicator of colonoscopy and is inversely associated with interval colorectal cancer risk. Although computer-aided detection (CADe) improves ADR in randomized trials, evidence from routine practice and its impact on concurrent endoscopist-performed optical biopsy remains limited. This study evaluated whether real-time deep learning–based CADe improves ADR in routine colonoscopy and assessed the diagnostic performance of endoscopist-performed optical biopsy.
Materials and
Methods: This non-randomized prospective pragmatic study was conducted at a tertiary referral center in Türkiye between October 2025 and February 2026. A total of 2,122 consecutive colonoscopies were allocated by weekly alternation to AI-assisted detection (CADe-on, n=1,061) or standard colonoscopy (CADe-off, n=1,061). The CADe system was used exclusively for lesion detection, whereas optical diagnosis was performed by the endoscopist. The primary outcome was ADR; secondary outcomes included polyp detection rate (PDR), adenomas per colonoscopy (APC), lesion characteristics, withdrawal time, and optical biopsy accuracy using histopathology as the reference standard.

Results: Baseline characteristics were comparable between groups. ADR increased from 13.0% to 19.0% (absolute increase, 6% points; p<0.001), and AI assistance remained independently associated with adenoma detection (adjusted odds ratio 1.56; 95% confidence interval [CI] 1.22–1.99). PDR and APC were also higher in the AI-assisted group (35.4% vs. 28.7%, p=0.002; 0.32±0.71 vs. 0.21±0.54, p=0.004). Detection gains were mainly driven by diminutive and non-polypoid lesions, while withdrawal time did not differ significantly (p=0.11). Endoscopist-performed optical biopsy (428 polyps) showed sensitivity 88.4%, specificity 81.2%, positive predictive value 84.7%, negative predictive value (NPV) 85.4%, and overall accuracy 85.1% (95% CI 81.4–88.3). For diminutive rectosigmoid polyps, NPV was 91.2%, meeting the preservation and ıncorporation of valuable endoscopic ınnovations threshold.

Conclusion: Real-time CADe was associated with higher adenoma detection without prolonging withdrawal time. Endoscopist-performed optical characterization during AI-assisted colonoscopy showed reliable diagnostic performance, suggesting that this combined approach may be feasible in routine clinical practice.