Larynx Object Segmentation and Indicators Generation Based on 3D VOSNet and Laryngeal Endoscopy Successive Images

Abstract

Clinically, the laryngoscopy videos are often used to observe vocal folds movement and analysis larynx-related lesions preliminarily. However, there is a lack of objective larynx indicators in medicine currently. Thus, the 3D VOSNet architecture is used to extract spatial features and classify the larynx object in the sequence images of laryngoscopy. The results represent that the 3D VOSNet can accurately segment the left vocal fold, right vocal fold, and glottal, and the accuracy is 93.48%, 94.63%, and 89.91%, respectively. Finally, the self-built algorithm is utilized to calculate six measured indicators including the length, area, curvature, deviation of length and area of vocal folds, area of glottal, and symmetry of the vocal folds. Improve the effectiveness and quality of vocal fold examination by providing immediate and objective information to otolaryngologists.