Accurate segmentation of cerebrovascular structures is crucial for scientific research and clinical applications. However, manual labeling of the whole brain’s sophisticated and complex vasculature network is costly and limited, and could potentially compromise the performance and generalizability of supervised model which solely relies on high-quality labels. Semi-supervised strategies have been investigated to effectively take advantage of abundant unlabeled data. In this study, we propose a novel confident learning-based mean-teacher framework (CL-MT), which integrates noisy label refinement to alleviate the adverse effects of label noise and consistency regularization tailored for noisy labeled regions to learn useful representations from unlabeled data. In addition, we propose a backbone model UST-Net, which incorporates convolution and Transformer in both the encoder and decoder. This architecture enables the model to capture long-range dependencies at various scales. Comprehensive experiments demonstrated that our model outperformed state-of-the-art supervised and semi-supervised methods and can be generalized to diverse human and non-human primate datasets.