Deep learning screening for severe nasal cavity structural abnormalities using endoscopic imaging

To develop and evaluate a deep learning–based decision-support model using nasal endoscopy videos to identify severe nasal cavity structural abnormality (S-NCSA), an image-based morphological category defined by a visible septum–inferior turbinate contact point, and to explore its potential for future triage applications in non-specialist settings. We retrospectively analyzed 628 nasal cavities from 317 patients; frames were assigned by board-certified otolaryngologists to the image-based morphological categories S-NCSA, nasal cavity structural abnormality (NCSA), or normal (NL), with S-NCSA defined by a visible septum-inferior-turbinate contact point. ImageNet-pretrained convolutional neural networks (CNNs)-Visual Geometry Group 16 (VGG16), 50-layer residual network (ResNet50), 18-layer three-dimensional residual network (3DResNet18), and 34-layer residual network (ResNet34)-were fine-tuned and evaluated with patient-level stratified four-fold cross-validation. Two ResNet34 strategies were compared: an Average-Diagnosis Model, for which the case-level score was the mean logit difference across frames and a positive mean logit difference indicated S-NCSA; and a Severe-Priority Model, for which a case was classified as S-NCSA when at least one frame was predicted as S-NCSA. In the bilateral-nasal ResNet34 analysis, the Average-Diagnosis strategy achieved 85.5% accuracy (95% confidence interval [CI], 82.6–88.1), 66.2% sensitivity (57.8–73.9), and 90.6% specificity (87.8–93.2). The Severe-Priority strategy yielded 79.5% accuracy (76.3–82.7), 84.6% sensitivity (78.1–90.2), and 78.1% specificity (74.2–81.8), indicating a sensitivity–specificity trade-off. Artificial intelligence (AI)-assisted nasal endoscopy can flag S-NCSA morphology, defined by a visible septum–inferior turbinate contact point, as decision support rather than autonomous clinical diagnosis. Because symptom severity and functional nasal patency were not evaluated, these findings should not be interpreted as validation of clinically severe nasal obstruction. The potential role of the Severe-Priority strategy in primary-care triage requires prospective clinical evaluation, while the Average-Diagnosis strategy may support image-based morphological assessment in specialist settings.