(043) Applying Knowledge Distillation for Vocal Fold Identification in Laryngoscopy

Introduction: Laryngeal endoscopic images are crucial for the development of research combining computer vision and otorhinolaryngology. However, there have only been a few studies that provide a laryngeal dataset, especially vocal folds, for a comparative performance evaluation.

Methods: The data is retrospectively collected from 4,624 images of 876 different patients who are outpatient or inpatient having indicated laryngoscopy at the Department of Otorhinolaryngology, Cho Ray Hospital, Ho Chi Minh City, Vietnam from January 2020 to November 2021. All images of laryngoscopic examinations are captured at a resolution of 480 x 360 pixels and saved in JPEG format. The endoscopy process is taken using a flexible endoscopic system (Olympus Medical Systems Corp., Tokyo, Japan) by two laryngologists. In evaluating various backbones, three factors are utilized to measure the models: accuracy, recall, and precision.

Results: EfficientNetB1 shows extremely high accuracy (98.7%), recall (99.2%), and precision (98.5%) in assessing images of vocal fold appearance. EfficientNetB1 exhibits good Grad-CAM results by focusing heavily on critical anatomical landmarks for vocal fold exploration. Furthermore, our distilled model also performs well in feature extraction in both visible and invisible vocal fold images compared to the scratch model. Distillation helps the student to approximately match the teacher’s performance and also helps faster convergence and require minimal computing resource.

Conclusions: In this research, we introduce a new dataset about the laryngeal endoscopic images for detecting vocal folds. Moreover, we also evaluate some backbones to choose a suitable model for applying knowledge distillation. Finally, we propose a solution of AI assistance on smartphones to create a portable and smart laryngoscope system.