Laryngeal disease classification using voice data: Octave-band vs. mel-frequency filters

Area of Science:

• Otolaryngology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Laryngeal cancer diagnosis traditionally relies on specialist examinations.
• Non-invasive diagnostic methods using voice analysis are emerging with AI advancements.
• Mel Frequency Cepstral Coefficients (MFCCs) are standard for voice analysis, but may lack accuracy for subtle laryngeal changes.

Purpose of the Study:

• To compare the diagnostic effectiveness of MFCCs and Octave Frequency Spectrum Energy (OFSE) in classifying voice data.
• To differentiate between healthy voices, laryngeal cancer, benign mucosal disease, and vocal fold paralysis using AI.
• To evaluate OFSE with 1/3 octave band filters for improved laryngeal disease detection.

Main Methods:

• Analysis of voice samples from 363 patients using Convolutional Neural Network (CNN) models.
• Application of both MFCC and OFSE with 1/3 octave band filters for feature extraction.
• Utilized Grad-Class Activation Mapping (Grad-CAM) for visualizing critical voice features.

Main Results:

• OFSE with 1/3 octave band filters achieved significantly higher classification accuracy (0.9398 ± 0.0232) compared to MFCCs (0.7061 ± 0.0561).
• Grad-CAM identified distinct voice features for laryngeal cancer, including noise in the over-formant area and fundamental frequency changes.
• AI differentiation between benign conditions and laryngeal cancer remains challenging due to overlapping voice characteristics.

Conclusions:

• OFSE with 1/3 octave band filters demonstrates superior performance for diagnosing laryngeal diseases, including cancer.
• This AI-based voice analysis method shows potential for accurate, non-invasive early detection of laryngeal conditions.
• Further research is needed to address challenges in differentiating benign mucosal diseases from laryngeal cancer using voice data.