Application of Artificial Intelligence in Developing Calculation System for Airway Obstruction

Introduction: Drug-induced sleep endoscopy (DISE) is important tool to evaluate sites and degrees of airway obstruction in patient with obstructive sleep apnea (OSA). Traditional manual testing cannot accurately calculate the degree and duration of epiglottic obstruction, and previous grading mainly identified only three degrees of obstruction. In order to accurately calculate the obstruction ratio, this study developed a system for calculating the epiglottic obstruction ratio.

Methods: Pictures of epiglottic obstruction are acquired via DISE. The detection for airway collapse range uses the YOLO model to locate the epiglottic cartilage in an image and then employs the median-cut color quantization method to transform the image into color regions. Proposed region puzzle algorithm is used to obtain the image regions of the epiglottic cartilage and the airway range. The obstruction ratio is calculated accordingly then. The algorithm’s effectiveness and practicality were verified by comparing its results with clinical judgment results.

Results: The YOLO model’s training demonstrated good performance, with an overall high recognition rate (F1 score: 91%). The accuracy of this system in calculating the degree of airway collapse is 0.1%, and the average calculation time for each single image is only 0.71 second.

Conclusions: The system can perform in-depth image analysis of DISE videos and accurately calculate the obstruction rate and duration of epiglottis.