MGASM-Net: morphology-guided multi-task learning network with anatomic spatial mamba for 3D airway segmentation

Airway segmentation in computerized tomography (CT) images is a prerequisite for the diagnosis of respiratory diseases and bronchoscopic navigation. Severe class imbalance and the low intensity contrast between the bronchial lumen and wall pose significant challenges in segmenting complete airway structures, especially in peripheral bronchi. In addition, the low intensity contrast can also lead to leakage at blurred boundaries. To address these challenges, we proposed a novel multi-task learning network based on Mamba and CNN for airway segmentation, utilizing boundary segmentation as an auxiliary task to enhance attention to the airway region. Specifically, to tackle class imbalance, the anatomical spatial mamba module is designed to capture local and global features within and between slices, which can effectively detect the bronchi with varying diameters across the sparse airway distribution. Meanwhile, given the severe inter-class imbalance that impedes data-driven models from learning airway features, a skeleton point-based data sampling strategy is proposed to localize the specific airway regions. Inspired by deep supervision, a morphological feature guidance module introduces additional structural knowledge to direct attention of model to the airway regions, which helps alleviate the low intensity contrast problem. We conducted extensive experimental validation on two public airway datasets. Compared to state-of-the-art methods, MGASM-Net was capable of extracting much more bronchi and more complete airway structures while preventing over-segmentation, indicating its accuracy, robustness, and generalization. Furthermore, our work can be extended to other tubular structure segmentation tasks. Codes are available at https://github.com/cactusgithub/MGASM-Net .