Fully automatic airway segmentation in CT images via dual-stage neural network with topology-constrained centerline completion.

Background: Precise segmentation of airways from computed tomography (CT) images is critical for pre-procedural planning in navigational bronchoscopy; however, existing methods often suffer from distal branch fractures and morphological losses due to slender diameters, low contrast, and noise. To address these limitations and achieve accurate, topologically complete airway segmentation, this study proposed a novel dual-stage neural network with topology-constrained centerline completion. Methods: A two-stage centerline-guided network (TSC-Net) was developed that innovatively combines centerline structural similarity (CSS) loss and centerline overlap rate (COR) loss. CSS loss maintains the continuity of airway tree structures by strengthening centerline morphological constraints, while COR loss enhances global centerline overlap to ensure branch integrity. Additionally, a topology-difference-driven fracture-region completion mechanism was adopted to adaptively repair the fracturedand misclassified areas. The model employs a residual UNet (ResUNet) backbone integrated with an atrous spatial pyramid pooling (ASPP) module to capture multi-scale contextual features and adopts a two-stage training strategy in which the encoder is frozen during stage 2 while the decoder is fine-tuned using dynamically adjusted loss weights. This strategy enables targeted optimization of discontinuous regions without compromising globally well-predicted areas. Results: Experiments on the Airway Tree Modeling Challenge 2022 (ATM'22), Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI), Extraction of Airways from CT (EXACT'09), and private datasets showed that TSC-Net achieved a Dice similarity coefficient (DSC) of 93.2%±2.1%, significantly outperforming SwinUNETR (92.9%±3.6%, P<0.05), nnUNet (91.2%±1.2%, P<0.05), and UNet (75.6%±4.3%, P<0.05). The tree-length detection (TD) rate reached 95.5%±4.2%, and the branch detection (BD) rate reached 92.6%±3.6%, both of which were significantly higher than those of all baseline methods (P<0.01). Moreover, the segmented distal branches exhibited continuous and smooth morphology. Conclusions: Overall, our method significantly enhanced both segmentation accuracy and distal branch integrity in automated CT-based airway segmentation through multi-dimensional feature enhancement, topological constraint optimization, and a dynamic completion mechanism. This method holds great promise for pre-procedural planning in robotic-assisted bronchoscopy by enabling accurate visualization of peripheral airways and facilitating navigation to distant pulmonary targets.