Automatic Segmentation of Membranous Glottal Gap Area with U-Net-Based Architecture

Area of Science:

• Medical imaging analysis
• Vocal fold dynamics
• Artificial intelligence in healthcare

Background:

• Videostroboscopy is the standard for vocal fold function assessment.
• Quantitative analysis of vocal fold parameters, like membranous glottal gap area, is time-consuming in clinical settings.

Purpose of the Study:

• To develop and validate deep learning models for automatic segmentation of the membranous glottal gap area.
• To assess the efficiency and accuracy of U-Net based architectures for this task.

Main Methods:

• Five U-Net based deep learning models were developed using 2507 videostroboscopy images.
• Models were validated on an independent dataset of 410 images.
• Performance metrics included intersection over union, Dice coefficient, Hausdorff distance, mean squared error, and mean absolute error.

Main Results:

• Efficient U-Net achieved the highest intersection over union (0.8455) and Dice coefficient (0.9163).
• Four models demonstrated practical inference times (16-138 ms), with Efficient U-Net being a strong performer.
• Accurate calculation of the normalized membranous glottal gap area index was validated.

Conclusions:

• U-Net based architectures enable automatic segmentation of the membranous glottal gap area.
• Efficient U-Net presents a viable option for clinical application due to its segmentation quality and speed.
• The developed models facilitate quantitative analysis of vocal fold parameters and glottal area waveform.