A Deep-Learning Model for Multi-class Audio Classification of Vocal Fold Pathologies in Office Stroboscopy.

Area of Science:

• Otolaryngology
• Artificial Intelligence in Medicine
• Speech Science

Background:

• Videolaryngostroboscopy is a key tool for visualizing vocal fold (VF) dynamics.
• Differentiating between healthy VFs, unilateral paralysis (UVFP), and VF lesions is crucial for diagnosis and treatment.
• Automated classification of VF conditions using voice analysis holds potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop and validate deep-learning classifiers for distinguishing three VF states: healthy (HVF), UVFP, and VF lesions.
• To assess the performance of binary (HVF vs. pathological) and multi-class (HVF, UVFP, lesions) classification models.
• To evaluate model generalizability on an independent external dataset.

Main Methods:

• Voice data were extracted from stroboscopic videos of 105 UVFP, 63 VF lesion, and 41 HVF patients.
• Audio samples were converted to Mel-spectrograms and used to train ResNet18 deep-learning models.
• Models were trained for binary and multi-class classification and validated internally and externally.

Main Results:

• The binary classifier achieved higher performance on the test set (accuracy 83%, F1-score 0.90) than the multi-class classifier (accuracy 40%, F1-score 0.36).
• External validation showed reduced performance for both models, with the binary classifier achieving 63% accuracy and 0.48 F1-score.
• The multi-class classifier performed poorly on the external dataset (accuracy 35%, F1-score 0.25).

Conclusions:

• Deep learning models can differentiate healthy and pathological vocal fold conditions from stroboscopic video voice data with moderate accuracy.
• Multi-class classification and external validation significantly reduced model performance, indicating challenges in generalization.
• Voice data from stroboscopic recordings may have limited standalone diagnostic value for complex VF conditions; further research is warranted.