A Deep Learning Approach for Quantifying Vocal Fold Dynamics During Connected Speech Using Laryngeal High-Speed...

Area of Science:

• Laryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Voice disorders necessitate examining vocal fold dynamics during connected speech.
• Flexible laryngeal high-speed videoendoscopy (HSV) provides high temporal detail for studying vocal fold mechanics.
• Accurate segmentation of vocal fold edges is crucial for analyzing HSV data.

Purpose of the Study:

• To present an automated deep-learning scheme for segmenting the glottal area in HSV during connected speech.
• To derive glottal edges from the segmented glottal area for further analysis.
• To overcome limitations in previous methods for analyzing vocal fold vibration.

Main Methods:

• A custom HSV system acquired data from a healthy participant reciting the "Rainbow Passage."
• A deep neural network was designed for glottal area segmentation.
• An automated labeling tool trained the network on HSV frames, annotating the glottis region during vocal fold vibrations.

Main Results:

• The deep-learning network achieved high performance with a mean Intersection over Union (IoU) of 0.82 and a Boundary F1 (BF) score of 0.96.
• Accurate segmentation of glottal edges/area was demonstrated, even during non-stationary phonatory events and when vocal folds were not vibrating.
• The automated scheme successfully trained without manual labeling and outperformed previous hybrid approaches.

Conclusions:

• The automated scheme ensures precise glottis representation in challenging HSV data, including low-quality images and laryngeal maneuvers.
• This facilitates the development of HSV-based measures for assessing vocal fold vibratory characteristics in connected speech.
• The method supports the assessment of individuals with and without voice disorders.