User perceptions of an artificial intelligence-based computer-aided detection system in colonoscopy: a single-center exploratory survey study

Purpose: Artificial intelligence (AI)-based computer-aided detection (AI-CADe) systems can improve adenoma detection during colonoscopy. However, successful clinical implementation depends on diagnostic performance and user acceptance, usability, and workflow integration. This study evaluated user perceptions before and after the clinical implementation of AI-CADe. Methods: This single-center, repeated, cross-sectional descriptive survey evaluated the endoscopy unit staff perceptions of AI-CADe at the National Cancer Center. Two anonymous online surveys were administered before and 1 month after the system installation. The surveys assessed the perceived diagnostic benefits, workflow impact, user satisfaction, concerns regarding false-positive detection, dependence on AI, and factors influencing acceptance and continued use. The analyses were descriptive and exploratory. Results: Twenty-nine and 25 participants completed the pre- and post-installation surveys, respectively. Positive responses regarding the interest in AI-CADe, perceived adenoma detection rate improvement, expected lesion removal, patient satisfaction, procedural satisfaction, and workflow impact were reported in both survey phases. Concerns regarding overdetection, unnecessary biopsies, and increased dependence on AI were also reported. Accuracy and sensitivity were the most frequently selected adoption factors, followed by cost and false-positive rates. Open-ended feedback included positive comments regarding lesion recognition, procedural support, and limitations related to repeated alerts, false-positive alarms, delayed detection, and system responsiveness. Conclusion: Endoscopy unit staff members showed favorable perceptions of AI-CADe and recognized its potential value in colonoscopy practice. However, concerns regarding false-positive detection, workflow integration, and dependence on AI indicate the need for user-centered optimization and long-term real-world evaluation.