Deep learning for identification of velopharyngeal patency in drug-induced sleep endoscopy

• Main Authors: Chen, Chun-Ting, Chung, I-Fang, Hsu, Ying-Shuo, Wang, Chi-Tang, Lin, Yun-Hsuan, Huang, Sheng-Yao
• Format: Conference Proceeding
• Language: English
• Subjects: airway segmentation; Atmospheric modeling; attention mechanism; Data augmentation; Data models; Deep learning; Endoscopes; focal loss; Image segmentation; obstructive sleep apnea; Sleep apnea; sleep endoscopy; U-Net
• ISSN: 2325-5994
• DOI: 10.1109/iCAST57874.2023.10359292

Sleep endoscopy is a valuable diagnostic tool for assessing airway obstruction during sleep. Precise segmentation of airway structures is essential for clinical evaluation and treatment planning. In this study, we utilize a well-known deep learning method, U-Net++, for automatic velopharyngeal airway segmentation in sleep endoscopy images. Here U-Net++ is equipped with an attention mechanism to enhance the representational power of the model, the focus loss to handle the imbalanced situation of airway areas, and data augmentation techniques to introduce the data heterogeneity and then get better performance. Results demonstrate that using U-Net++ with data augmentation and focal loss significantly enhances segmentation performance, yielding a Dice Similarity Coefficient (DSC) of 0.912. Incorporating the attention mechanism leads to notable improvements in a challenging case, enhancing the DSC by 11.4%. Our findings highlight the potential of deep learning models in accurately segmenting velopharyngeal airway structures in sleep endoscopy images. Future research aims to further enhance the model's capabilities and address challenges posed by factors like saliva. In addition, we shall perform the segmentation task of oropharyngeal airway structures.