Vision-Based Assistance for Vocal Fold Identification in Laryngoscopy with Knowledge Distillation

Area of Science:

• Otorhinolaryngology
• Computer Vision
• Medical Imaging

Background:

• Laryngoscopy images are crucial for otorhinolaryngology research.
• A scarcity of laryngeal image datasets limits comparative studies.
• Advancements in computer vision can enhance diagnostic tools.

Purpose of the Study:

• To introduce a novel dataset of vocal fold images.
• To develop a lightweight AI model for laryngeal image analysis.
• To create an AI-assisted smartphone solution for portable laryngoscopy.

Main Methods:

• Creation of a new laryngeal image dataset.
• Implementation of knowledge distillation for model compression.
• Development of a lightweight neural network (student model).
• Integration of the model into a smartphone application.

Main Results:

• The proposed lightweight model achieved ~98.4% accuracy.
• Model weights were reduced by up to 88% compared to EfficientNetB1.
• An AI-assisted smartphone system for laryngoscopy was developed.

Conclusions:

• The study provides a valuable laryngeal image dataset.
• A compressed, efficient AI model is suitable for handheld laryngoscopy devices.
• The AI-assisted system facilitates portable and intelligent vocal fold observation and diagnosis.