Argus-Net: Anatomy-Guided Hierarchical Learning for Pulmonary Vasculature Segmentation in Contrast and Non-Contrast CT.

Abstract

Automatic segmentation of the pulmonary vasculature from thoracic CT is essential for pulmonary embolism assessment, surgical planning, and disease monitoring. However, accurate segmentation remains challenging due to low contrast, anatomical noise, and multi-scale vascular complexity, particularly in non-contrast-enhanced CT. To address these challenges, we propose an anatomically informed hierarchical deep learning framework that offers a practical balance between segmentation performance and computational efficiency. In the three-stage pipeline, MVP-U-Net extracts the anatomical region of interest, Argus-V-Net learns vascular representations using local, cross-sectional, and multi-scale attention, and Argus-A-Net performs artery segmentation via transfer learning. The framework was evaluated on the HIPAS and PARSE datasets. For vessel segmentation, Argus-V-Net achieved Dice scores of 0.9073 on contrast-enhanced CT and 0.9227 on NCCT. For artery segmentation, Argus-A-Net achieved a Dice score of 0.9020 on NCCT. Pulmonary artery segmentation in NCCT without explicit intensity-based artery-vein separation remains a challenging task and has received limited attention in the literature. Despite the limited intensity contrast between arteries and veins in NCCT, the framework achieves competitive performance by leveraging anatomical priors and hierarchical feature representations rather than explicit intensity-based separation. Compared with the state-of-the-art nnU-Net, Argus-Net achieves competitive Dice scores (0.9125 vs. 0.9237 for vessel segmentation, p = 0.112) while requiring 9.5 times fewer parameters and providing approximately six times faster inference, offering a favorable balance between segmentation performance and computational efficiency. Overall, the proposed framework provides a contrast-adaptive approach for pulmonary vessel and artery segmentation and demonstrates consistent performance across both contrast-enhanced and non-contrast CT datasets.