Segmentation of coronary arteries is essential for subsequent diagnostic efforts. State-of-the-art segmentation methods commonly result in disconnected arterial branch prediction due to the complex nature of these segmentation tasks driven by complex anatomy and acquisition challenges. However, the coronary artery tree contains critical arterial shape information via its centrelines which offers great potential for the improvement of automated segmentations works using 3D computed tomography coronary angiograms (CTCA). In this paper, we propose a deep learning architecture using 3D CTCA from the open source dataset ASOCA that combines the centreline of the coronary artery with existing segmentation methods to improve overall segmentation performance. This architecture contains three novel modules. First, the bridging centreline extraction module searches for missing centrelines that connect two disconnected components from an initial segmentation. Second, the centreline expansion module expands the bridging centrelines into coronary artery segments. Third, the centreline fusion module further combines the centreline and the initial segmentation to remove background noise. Experiments show that the proposed architecture can consistently boost the segmentation performance of various segmentation methods.