Abstract
Accurate segmentation of pulmonary structures from computed tomography (CT) is critical for lung disease management, yet progress is hampered by a lack of large-scale, public datasets with comprehensive multi-structure annotations. To address this, we present the Airway and Pulmonary Vessel Structural Representation in CT (AirRC) dataset, comprising 254 CT scans from the LUNA16 dataset meticulously annotated with 3D masks for pulmonary veins, arteries, airway lumen, and airway wall. Technical validation was performed via 5-fold cross-validation using a custom MONAI-based deep learning pipeline. The model achieved high mean Dice Similarity Coefficients (DSC) for Pulmonary Veins (0.953), Pulmonary Arteries (0.950), and Airway Lumen (0.941), with strong performance on the challenging Airway Wall (0.866). A two-stage refinement strategy further improved small airway branch segmentation. External validation on public benchmarks (ATM'22, Parse2022, HiPas) confirmed the utility and generalizability of models trained on AirRC, establishing it as a robust resource for developing and evaluating advanced pulmonary segmentation algorithms.