Lung cancer, the second most common type of cancer worldwide, is primarily treated through surgery. During the operations, preserving pulmonary arteries and veins is a crucial problem. In recent years, 3D visualization techniques like virtual reality and 3D printing have been increasingly used in clinical practice for lung cancer surgery planning. Under the success of these techniques, automatic segmentation of pulmonary arteries and veins plays a key role. Particularly, the state-of-art approaches rely on two techniques, i.e. the deep neural networks (DNNs) or the traditional machine learning (ML) method, and both techniques have respective shortages. Basically, the ML-based methods generally demonstrate a limited performance, while the DNN-based methods lack sufficient annotation for accurate segmentation. In response to such a dilemma, this paper proposes a fusion method to combine the DNN-based and ML-based methods to segment pulmonary arteries and veins for lung cancer surgery planning. Particularly, the anatomy prior mask corresponding to pulmonary arteries and veins are identified using the marching cubes algorithm and Attention U-Net. Subsequently, an enhanced attention U-Net, is used to integrate the original CT scans with the anatomy prior mask to generate the refined segmentation results. Following this, an anatomy structure enhancement module is used to refine the segmentation further by refining disconnected vessel segments and correcting misclassified vessels based on anatomy prior masks. We experimented the proposed approach on a private dataset of 95 CT scans collected from patients after surgery, and then annotated by lung cancer experts. The results demonstrate that our approach outperforms the existing methods with an improvement of 5.1% to 16.2% in Dice score. The dataset and code have also been published [1] to facilitate further research in this field.