Clinical Applications of Deep Learning for Glottal Area Segmentation and Glottal Area Waveform Feature Computation

Area of Science:

• Laryngology
• Medical Imaging
• Computational Biology

Background:

• Glottal area waveform analysis is crucial for understanding vocal fold vibration.
• Manual annotation for this analysis is time-consuming and prone to errors.

Purpose of the Study:

• To develop an automated system for glottis segmentation and glottal area waveform feature extraction.
• To improve the efficiency and accuracy of laryngeal assessment.

Main Methods:

• An automated system integrating YOLO and U-Net architectures was developed.
• The system performed glottis detection and segmentation on videostroboscopy recordings.
• Subject-independent 5-fold cross-validation was used on 5017 annotated frames.

Main Results:

• The system achieved high accuracy with an average Intersection over Union of 92.8% and Dice Similarity Coefficient of 95.8%.
• Computation time was reduced by 27.7-fold compared to manual methods.
• Consistent trends in glottal area dynamics were detected post-treatment in patients.

Conclusions:

• The automated system significantly enhances the efficiency and accuracy of glottal area waveform analysis.
• The system demonstrates clinical utility for objective laryngeal assessment.
• This technology aids in evaluating voice treatment outcomes.