Optimizing Colorectal Polyp Screening: A Novel Artificial Intelligence-Assisted Colonoscopy Diagnostic System Based on NICE Classification

Abstract

Background: This study aims to evaluate the diagnostic performance of an enhanced artificial intelligence-assisted colonoscopy system, CAD-N-Pro, based on the NICE (Narrow-band Imaging International Colorectal Endoscopic) classification. Methods: Compared to the previous CAD-N system, this study optimized the algorithm into a segmentation network to comprehensively assess the diagnostic performance of the CAD-N-Pro model. A total of 14 675 images from 5 hospitals were classified using the NICE classification for training, internal and external validation. The model's performance was also compared with the previous CAD-N model. To validate the clinical applicability, 200 colonoscopy videos were prospectively collected and analyzed, with comparisons made among endoscopists of different seniority. Results: In external image validation, CAD-N-Pro demonstrated excellent diagnostic accuracy across polyp types, with an overall AUC of 0.979. The system achieved accuracies of 0.966 for type 1 polyps and type 2 polyps (95% CI 0.956–0.975), and 0.997 for type 3 polyps (95% CI 0.993–0.999), 0.994 for normal background (95% CI 0.990–0.997). In the video validation, the performance of CAD-N-Pro was demonstrated to be superior to that of endoscopists with different years of experience, particularly in the diagnosis of type 1 and type 2 polyps. Moreover, CAD-N-Pro exhibited superior performance to endoscopists in detecting colorectal polyps of different sizes, especially those < 10 mm. For polyps larger than 10 mm, its performance was comparable to that of endoscopists with > 3 years of experience. Conclusion: The optimized CAD-N-Pro model enhances optical diagnostic accuracy for colorectal polyps, providing a robust tool for clinical decision-making in real-time colonoscopy examinations.