Semi-Supervised Pulmonary Airway Segmentation with Two-Stage Feature Specialization Mechanism

Pulmonary airway segmentation plays an important role in the diagnosis and treatment of pulmonary diseases. Recently, deep learning segmentation methods have been able to outperform traditional methods by capturing richer complicated features from the airway tree relying on a large label set, which is hard to acquire in practice. Semi-supervised methods significantly reduce the requirement for labels but can hardly deal with the complexity of the tree structures, especially fine peripheral bronchi. We proposed a novel two-stage feature specialization mechanism (TFSM) that learns the airway tree in a semi-supervised manner, which significantly improves the segmentation quality. Comprehensive experimental results reveal that the proposed TFSM method improved the dice score of a semi-supervised baseline by 10% with only 20% of the labeled CT images that stay competitive with the fully supervised method.