Comparison of Deep Learning Models for Voice Disorder Classification Using Kymographic Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Speech Science

Background:

• Diagnosing voice disorders with high-speed video (HSV) endoscopy is challenging due to manual frame analysis.
• Kymography aids clinical decision-making by visualizing vocal fold vibrations.
• Automating analysis of kymographic images can improve diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automated classification of voice disorders from kymographic images.
• To assess the performance of various pretrained deep learning models in detecting pathological vibratory patterns.

Main Methods:

• Generated kymographic images from high-speed recordings in the Benchmark for Automatic Glottis Segmentation (BAGLS) dataset.
• Employed deep learning models (AlexNet, DenseNet121, Xception, Inceptionv3, ResNet50v2) for binary and tertiary classification.
• Trained models to identify subtle variations indicative of voice pathology.

Main Results:

• DenseNet121 demonstrated superior performance in classifying voice disorders compared to other evaluated models.
• The deep learning classifier achieved high accuracy and outperformed existing methods.
• The model effectively detected intricate variations in pathological vibratory patterns.

Conclusions:

• A deep learning-based kymographic image classifier, particularly DenseNet121, shows significant potential for automating voice disorder diagnosis.
• This automated approach can serve as a valuable diagnostic assistance tool for clinicians.
• Further research may lead to improved clinical decision support for voice pathology.