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

OBJECTIVE: To develop an intelligent diagnostic model based on a Local-Higher Order Graph Neural Network (LHGNN) to achieve noninvasive auxiliary multi-class pathological classification of vocal fold tumors by analyzing Mel-spectrograms of voice data. METHODS: Patients with pathologically confirmed vocal fold polyps (n = 1272), premalignant lesions (n = 180), and malignant tumors (n = 197) were retrospectively identified. Patients were divided into training and validation sets in a ratio of 8:2. The patient's vowel and continuous sentence audio were collected, and the audio was segmented and converted into Mel-spectrograms. The data was enhanced by random occlusion, adding noise and time shifting. We used a convolutional backbone network-based LHGNN architecture, whose core consists of multiple cascaded LHG modules. The model was evaluated using metrics including accuracy, recall, F1-score, and the area under the curve (AUC) value. RESULTS: The final dataset comprised 4745 audio samples after balanced sampling (1603 vocal fold polyps, 1598 premalignant lesions, 1544 vocal fold carcinomas). The validation set achieved optimal performance after 461 training epochs: overall accuracy of 92.94%, F1-score of 0.9289, recall of 0.9281, and AUC of 0.9877. The multi-class confusion matrix demonstrated identification accuracies of 91.35% for vocal fold polyps, 90.99% for premalignant lesions, and 94.05% for malignant tumors. CONCLUSION: This study successfully validated the high efficacy of the LHGNN model in the multi-class classification of vocal fold tumors. Its excellent performance highlights the significant potential of deep learning in vocal pathological analysis. The model provides a new tool for the noninvasive preliminary screening of vocal fold tumors.