Deep Learning-Based Objective Quantification of Nasopharyngeal Endoscopic Findings for Standardized Assessment of...

Background/Objectives: Nasopharyngeal inflammation is commonly evaluated through visual inspection of endoscopic findings, which remains subjective and prone to interobserver variability. This study aimed to develop and validate a deep learning-based system for objective quantification of key nasopharyngeal endoscopic findings. Methods: A total of 200 endoscopic videos were retrospectively analyzed as an independent evaluation dataset, while a separate annotated dataset of 279 cases was used for model training. Four findings-mucosal color tone, swelling, mucus or crust adhesion, and bleeding after abrasion-were scored by expert otolaryngologists using a three-point scale, and their sum was used as a composite reference severity score (Y8, range 0-8). A convolutional neural network generated continuous probability outputs for each finding, which were aggregated into a composite score (S8). Results: For the primary threshold (Y8 ≥ 3), the AI-derived score demonstrated strong agreement with expert consensus (AUC 0.874). A predefined rule-based diagnostic criterion also showed comparable discriminative performance (AUC 0.851). Conclusions: Deep learning-based quantification provides an objective and reproducible method for evaluating nasopharyngeal endoscopic findings. This approach may enable standardized assessment of inflammation and support more consistent clinical decision-making, particularly for identifying clinically relevant inflammation, while its ability to stratify higher severity levels is more limited.