Optimal Deep Learning-Based Vocal Fold Disorder Detection and Classification Model on High-Speed Video Endoscopy.

Area of Science:

• Laryngology and Speech Science
• Biomedical Imaging
• Artificial Intelligence in Medicine

Background:

• Accurate vocal fold boundary identification is crucial for studying speech phonation.
• High-speed video-endoscopy (HSV) offers high temporal resolution for capturing vocal fold vibrations during speech.
• Existing methods may lack precision in analyzing complex vocal fold dynamics in running speech.

Purpose of the Study:

• To develop an automated deep learning-based method for precise vocal fold boundary identification using HSV.
• To enhance the diagnosis and categorization of vocal fold abnormalities during connected speech.
• To improve the temporal resolution and accuracy of analyzing vocal fold vibratory characteristics.

Main Methods:

• Utilized high-speed video-endoscopy (HSV) for laryngeal imaging.
• Developed an automated algorithm (ODL-VFDDC) involving temporal segmentation and motion correction.
• Employed a deep belief network (DBN) model optimized with an agricultural fertility algorithm (AFA) for classification.
• Applied a farmland fertility algorithm (FFA) for accurate glottal limit determination.

Main Results:

• The ODL-VFDDC technique demonstrated superior performance in vocal fold disorder classification compared to existing methods.
• Successfully tracked vocal fold boundaries across frames with high accuracy and resilience to noise.
• Achieved precise identification of glottal limits in vibrating vocal folds.
• Showcased the potential for automated analysis of vocal fold movement during connected speech.

Conclusions:

• The proposed ODL-VFDDC method offers an effective and robust approach for analyzing vocal fold dynamics during speech.
• This technique significantly advances the automated diagnosis and understanding of vocal fold disorders.
• Provides a novel, self-sufficient pathway for studying vocal fold motion in connected speech.