Development and validation of a deep learning model for identifying high-quality laryngoscopic images

초록

Laryngoscopy is essential for evaluating laryngeal pathology, particularly vocal fold lesions, but large endoscopic datasets often contain low-quality or irrelevant frames that hinder use. We developed and validated a deep learning model to automatically identify high-quality laryngoscopic images that clearly show the vocal folds. This retrospective study included 4711 images from 125 patients. Expert reviewers labeled images as low (3099; 65.8%), mid (698; 14.8%), or high quality (914; 19.4%). High-quality images were defined as those in which the entire vocal fold region was clearly visible, allowing definitive assessment of lesions. Eight pretrained networks-AlexNet, ResNet-50, MobileNetV2, ConvNeXt-tiny, Vision Transformer (ViT-B/16), DaViT, Swin Transformer V1-tiny, and V2-base-were fine-tuned using transfer learning. Performance was evaluated by accuracy, precision, recall, F1-score, AUROC, and AUPRC. Swin Transformer V1-tiny showed the best binary classification performance (high vs. non-high quality): 95.1% accuracy, 84.9% precision, 91.3% recall, 87.9% F1, AUROC 0.979, and AUPRC 0.927. Grad-CAM was used for interpretability and confirmed focus on vocal folds and anterior commissure. External validation using the Laryngoscope8 dataset yielded 91.9% accuracy. The model enables efficient selection of high-quality images for clinical and research use. Its strong performance and simple architecture allow application to other endoscopic domains.

키워드