Weakly Supervised Learning Applied for Airway Patency Identification in Drug-Induced Sleep Endoscopy Videos

Obstructive Sleep Apnea (OSA) is a prevalent sleep disorder. Drug-induced sleep endoscopy (DISE) is a common technique for evaluating upper airway obstruction in OSA patients. Due to the time-consuming process of interpreting DISE videos, recent studies have utilized deep learning models to assist interpretation. However, these methods can only roughly distinguish severe obstruction and cannot objectively quantify airway patency. In this study, we address these challenges by employing a segmentation model to assess the airway area. Here we first create a quality assessment model to filter out low-quality frames, enhancing dataset reliability. Subsequently, we build a segmentation model using the weakly supervised learning strategy to minimize annotation efforts and improve efficiency. Finally, by deploying both a quality assessment model and a well-trained segmentation model, we evaluate airway patency based on the segmented airway area. The results demonstrate that our proposed strategy enhanced the ability of the segmentation model. By leveraging both the segmentation model and the quality assessment model, we achieve a more objective and accurate measurement of the airway area, providing valuable insights into the degree of obstruction.