NHPVS: a nodule-aware deep learning method for high-abundance pulmonary vessel segmentation on chest CT.
Authors
Affiliations (5)
Affiliations (5)
- Radiology, Peking Union Medical College Hospital, No.1 Shuaifuyuan Wangfujing Dongcheng District, Beijing, 100730, China, Beijing, Beijing, 100730, China.
- Research and Development Center, Canon Medical Systems China Co Ltd, 23rd Floor, Tower A, Ping An International Financial Center No.3, Xinyuan South Road, Chaoyang District, Beijing 100027, China, Beijing, Beijing, 100027, China.
- Research and Development Center, Canon Medical Systems China Co Ltd, 23rd Floor, Tower A, Ping An International Financial Center No.3, Xinyuan South Road, Chaoyang District, Beijing, 100027, China, Beijing, Beijing, 100027, China.
- Research and Development Center, Canon Medical Systems China Co Ltd, 23rd Floor, Tower A, Ping An International Financial Center No.3, Xinyuan South Road, Chaoyang District, Beijing, 100027, China, Beijing, 100027, China.
- Radiology, Peking Union Medical College Hospital, No.1 Shuaifuyuan Wangfujing Dongcheng District, Beijing, 100730, China, Beijing, 100730, China.
Abstract
Deep learning-based segmentation of pulmonary vessels is challenged by vessel-nodule confusion and insufficient detection of small vessels. To address these issues, we propose a Nodule-aware High-abundance Pulmonary Vessel Segmentation (NHPVS) framework that enhances nodule differentiation and improves small vessel segmentation. This multi-center study retrospectively collected 477 chest CT scans comprising non-contrast and contrast scans. A 3D U-Net-based model was trained on 370 scans for pulmonary vessel segmentation. Performance metrics included Dice similarity coefficient (DSC), sensitivity, centerline DSC (cl DSC), 95% Hausdorff distance (HD95), and nodule misclassification rate (NMR). Visual and quantitative assessments were conducted on the external test set. NHPVS outperformed existing methods, achieving a DSC of 89.2%, a sensitivity of 89.2%, a cl DSC of 93.2%, an HD95 of 1.4 mm, and an NMR of 1.8%. Compared with the state-of-the-art method (nnUNet-v2), NHPVS showed superior segmentation accuracy, vascular continuity, and branch abundance, with improvements of 31.1% in vessel length and 60.5% in branch counts. The volume of small pulmonary vessels (diameter < 3 mm) segmented by NHPVS was significantly greater than that segmented by nnUNet-v2 (p < 0.001). By enabling the precise visualization of pulmonary arteries and veins on chest CT, the proposed NHPVS framework addresses the critical need for accurate vascular characterization in disease diagnosis and surgical planning.