Multi-Classification of Vocal Cord Masses via Local-Higher Order Graph Neural Network on Voice Data

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Bioacoustics

Background:

• Vocal fold tumors require accurate pathological classification for effective treatment.
• Current diagnostic methods can be invasive and may not always provide definitive results.
• Analyzing voice data offers a potential noninvasive avenue for tumor detection.

Purpose of the Study:

• To develop and validate an intelligent diagnostic model using a Local-Higher Order Graph Neural Network (LHGNN).
• To achieve noninvasive, multi-class pathological classification of vocal fold tumors.
• To analyze Mel-spectrograms derived from voice data for tumor classification.

Main Methods:

• A retrospective study included patients with vocal fold polyps, premalignant lesions, and malignant tumors.
• Audio data (vowels, continuous sentences) were converted into Mel-spectrograms and augmented.
• A convolutional backbone network-based LHGNN architecture was employed for classification.

Main Results:

• The LHGNN model achieved high performance on a dataset of 4745 audio samples.
• Validation metrics included an overall accuracy of 92.94%, F1-score of 0.9289, and AUC of 0.9877.
• Specific accuracies for polyps, premalignant lesions, and malignant tumors were 91.35%, 90.99%, and 94.05%, respectively.

Conclusions:

• The LHGNN model demonstrated high efficacy in multi-class classification of vocal fold tumors.
• Deep learning holds significant potential for advancing vocal pathological analysis.
• The developed model serves as a novel tool for noninvasive preliminary screening of vocal fold tumors.