Behavioral Analysis of Airway Deformation during Drug Induced Sleep Endoscopy

摘要

Drug Induced Sleep Endoscopy (DISE) is a procedure that simulates natural sleep which allows medical experts to obtain a dynamic assessment of the upper airway to monitor and diagnose obstructed breathing. DISE diagnosis is based on visual examination of the obtained endoscopic images that uses a subjective scoring system to diagnose reoccurring obstructive respiratory events. Due to this subjectivity, objective measurements of breathing patterns and obstructions are difficult to reliably obtain. Cine Magnetic Resonance Imaging and Computed Tomography (MRI/CT) scans are considered useful tools for both static and dynamic evaluation of OSA, but are both complex to perform and expensive. The lack of objective measurements causes difficulty for medical professionals for accurately understanding and examining the conditions of chronic airway conditions. In this work, we describe an integration of machine learning and DISE monocular video with Cine MRI data to construct dynamic 3D models, providing an objective measure of airway behaviors and conditions. Dynamic changes in airway dimensions that occur with the respiratory cycle during sleep can be reconstructed and represented by a 3D airway model for prognosis assistance. This approach for Upper Aerodigestive Tract (UAT) and laryngotracheal airway reconstruction is accomplished through image-to-surface modeling using a Generative Adversarial Network (GAN) to integrate Cine MRI and video endoscopy. Results show that we are able to reconstruct the airway into 3D point-cloud data as well as model complex dynamics of airway surfaces.