A Cascaded Network for Airway Tree Segmentation Incorporating Multiple Attention Mechanisms

Airway tree segmentation is crucial for diagnosing and treating lung diseases. A three-dimensional representation of the patient's airway tree serves as the foundation for bronchoscopic navigation, instrumental for preoperative planning and intraoperative visualization during intervention procedures. However, certain inherent characteristics such as complex structure, scale variation, and class imbalance often result in inadequate segmentation. To mitigate these limitations, we propose a cascaded airway tree segmentation network named Spatial-Channel-Deformable Cascade Network (SCD-CascadeNet). It consists of two main stages: coarse segmentation and fine segmentation. The coarse segmentation network swiftly localizes the major airways, while the fine segmentation network offers more precise segmentation of the peripheral minor airways. Specifically, the Spatial-Channel-Deformable Attention module (SCDA) is designed to capture crucial spatial regions and significant channel features within the skip connection of the fine segmentation network, thereby enhancing the network's ability to recognize tubular structures and peripheral airways. We evaluate our method using the challenging dataset named Airway Tree Modeling (ATM). Both qualitative and quantitative experimental results demonstrate that our method achieves competitive performance and holds promise for further clinical applications.