Deep Learning-Based Detection of Sessile Serrated Lesions in Colonoscopy Images: Optimization of a YOLOv3 Framework With Mucus Cap Feature Enhancement for Sessile Serrated Lesion Detection.

Background

AND

Aim

Sessile serrated lesions (SSLs) are difficult to detect during colonoscopy via standard white-light imaging (WLI) because they are flat, have indistinct borders, and are often hidden by mucus caps. Although narrowband imaging (NBI) can help, its effectiveness is inconsistent. This study evaluated the feasibility of an artificial intelligence (AI)-assisted approach incorporating mucus cap features and both WLI and NBI for SSL detection and diagnosis.

Methods

This study developed an AI model for SSL detection by training it on 1980 colonoscopic images (2023-2025). The model enhances standard detection frameworks by specifically integrating mucus cap features and optimizing WLI performance. Model performance was evaluated using detection rate, accuracy, and AUC.

Results

At the image level, the SSL detection model achieved an accuracy of 98.7% (sensitivity 100.0%, specificity 97.4%). Lesion-level detection improved from 77.4% with the conventional model to 95.0% after incorporation of mucus cap screening, representing an improvement of 17.6%, with detection rates of 95.9% for mucus cap-free lesions and 98.4% for mucus cap-containing lesions. For morphology-based typing, the WLI-based model achieved an accuracy of 85.8% (AUC 0.89), whereas the NBI-based model showed superior performance with an accuracy of 91.6% (AUC 0.94).

Conclusions

This study provides real-world clinical evidence validating AI-based models for SSL detection and diagnosis. Although representing an incremental step rather than a novel framework, such evidence is essential for improving model generalizability and enabling stable extrapolation across clinical settings. Further prospective and multicenter studies are required before reliable clinical deployment.