Automatic and accurate delineation of cerebrovascular structures from Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) images is crucial for diagnosing and treating cerebrovascular diseases. While most deep learning approaches have presented encouraging capabilities for vessel delineation, their results still suffer from disconnected and incomplete segments due to the insufficient exploration of cerebrovascular structures and topologies. To tackle this issue, this paper proposes a connectivity-reinforced deep learning approach to protect the topological information of the vessels. In detail, the cerebral vessels are tracked and highlighted with emphasis on both the central and edge voxels by the specially designed CONnectivity Attention Module (CONAM). Furthermore, a specially designed adaptive connectivity loss ( \(\mathcal {L}_{AC}\) ) is introduced to reinforce the network training by balancing the global penalty and the auto-adjusted regional penalty, subsequently optimizing the overall connectivity of the vessel predictions. Extensive experiments have been conducted on the well-known IXI dataset, and the proposed method has been compared to seven state-of-the-art approaches. The results demonstrate that our method outperforms these approaches both quantitatively and qualitatively. Code will be available at https://github.com/Yusjlalala/CONA-Net .