Diagnostic Accuracies of Laryngeal Diseases Using a Convolutional Neural Network-Based Image Classification System

Area of Science:

• Otolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Interobserver variability exists in diagnosing laryngeal diseases from laryngoscopic images.
• Clinical experience influences diagnostic accuracy, necessitating objective assessment tools.

Purpose of the Study:

• To develop and evaluate deep learning-based models for computer-assisted diagnosis of common laryngeal diseases.
• To compare the diagnostic performance of AI models against human visual assessments.

Main Methods:

• Retrospective analysis of 4106 laryngeal images across nine disease categories.
• Stratified eightfold cross-validation using deep neural networks (DNNs).
• Performance evaluation based on precision, recall, accuracy, F1 score, PR curve, and PR-AUC, compared to four trainees' assessments.

Main Results:

• DNNs outperformed trainees in classifying most laryngeal diseases, showing higher PR-AUC and F1 scores.
• DNNs achieved a micro-average PR-AUC of 0.95 and macro-average PR-AUC of 0.91, surpassing all trainees.
• While DNNs showed slightly lower performance for Reinke's edema and nodules, they were comparable to expert trainees.

Conclusions:

• Deep neural network technology shows promise for application in laryngoscopy.
• AI can supplement clinical assessments by offering diagnostic clues and acting as a diagnostic reference.
• Computer-assisted diagnosis using DNNs can potentially reduce interobserver variation in laryngeal disease diagnosis.