Quantification and Analysis of Laryngeal Closure From Endoscopic Videos

Objective: At present, there are no objective techniques to quantify and describe laryngeal obstruction, and the reproducibility of subjective manual quantification methods is insufficient, resulting in diagnostic inaccuracy and a poor signal-to-noise ratio in medical research. In this work, a workflow is proposed to quantify laryngeal movements from laryngoscopic videos, to facilitate the diagnosis procedure. Methods: The proposed method analyses laryngoscopic videos, and delineates glottic opening, vocal folds, and supraglottic structures, using a convolutional neural networks (CNNs) based algorithm. The segmentation is divided into two steps: A bounding box which indicates the region of interest (RoI) is found, followed by segmentation using fully convolutional networks (FCNs). The segmentation results are statistically quantified along the temporal dimension and processed using singular spectrum analysis (SSA), to extract clear objective information that can be used by the clinicians in diagnosis. Results: The segmentation was validated on 400 images from 20 videos acquired using different endoscopic systems from different patients. The results indicated significant improvements over using FCN only in terms of both processing speed (16 FPS vs. 8 FPS) and segmentation result statistics. Five clinical cases on patients have also been provided to showcase the quantitative analysis results using the proposed method. Conclusion: The proposed method guarantees a robust and fast processing of laryngoscopic videos. Measurements of glottic angles and supraglottic index showed distinctive patterns in the provided clinical cases. Significance: The proposed automated and objective method extracts important temporal laryngeal movement information, which can be used to aid laryngeal closure diagnosis.