FDA: Feature Decomposition and Aggregation for Robust Airway Segmentation

3D Convolutional Neural Networks (CNNs) have been widely adopted for airway segmentation. The performance of CNNs is greatly influenced by the dataset while public datasets are mainly clean CT scans with coarse annotation, thus difficult to be generalized noisy (e.g. COVID-19 scans). In this work, we proposed a new dual-stream network address variability between domain and domain, which utilizes small amount labeled We designed two different encoders extract transferable features unique separately, followed independent decoders. Further on, refined channel-wise feature recalibration Signed Distance Map (SDM) regression. module emphasizes critical SDM pays more attention bronchi, beneficial extracting topological robust labels. Extensive experimental results demonstrated obvious improvement brought our method. Compared other state-of-the-art transfer learning methods, method accurately segmented bronchi in scans.