Automated segmentation of central airway obstruction from endoscopic video stream with deep learning

Central airway obstruction (CAO) can disrupt normal breathing and pose significant risks, with treatments often involving surgical removal. Accurate image segmentation is crucial for identifying the target region to assist surgeons or robotic surgery system during operations. Although manual annotation is the gold standard, it is time-consuming and subjective, making a robust automated segmentation algorithm desirable. In this paper, we propose an automated deep learning framework for CAO segmentation. As a benchmark study, we build a custom CAO phantom model and acquire endoscopy videos. We then establish inter-rater variability. To assess the effectiveness of the proposed method, a 4-fold cross-validation is performed with The Dice score as an evaluation metric. The proposed CAO segmentation method yields mean binary Dice scores of 0.95 and 0.89 against annotations from two human raters, respectively, compared to an inter-rater variability Dice score of 0.88. This indicates that the proposed framework can provide robust CAO segmentation from endoscopic video.