Abstract: Cerebrovascular segmentation is a crucial foundation for cerebrovascular disease screening, pathological progression tracking, and precise surgical treatment. Owing to patient privacy protection concerns and complex structure, existing public cerebrovascular datasets have limited samples and insufficient manual annotation, which hinders the research and development of high-performance segmentation. To address the issues, this paper proposes a coarse-to-fine segmentation Double U-Net collaborative optimization model (DU-Net), aiming to improve performance on small-sample cerebrovascular segmentation. The DU-Net model consists of two paths, both based on 3D U-Net as the backbone. The coarse segmentation path utilizes lightweight convolutions with fewer channels to reduce the collaborative training burden and achieve preliminary learning of cerebrovascular topology. These features are then passed to the fine segmentation path to compensate for generalization features. In the fine segmentation path, convolutions with more channels contribute to enhancing the model ability to express detailed features, boundary information, and small targets. This improves segmentation accuracy and stability, particularly when refining local features and optimizing small target and boundary regions. In addition, a Squeeze and Excitation (SE) module is introduced in the skip connections of the fine segmentation path to assign higher learning weights to the target areas, significantly enhancing the model representation ability of fine segmentation path to learn and segment complex regions and detailed features. Based on existing public resources, this paper conducts comparative experiments with 10 typical medical segmentation models. Experimental results exhibit that the DU-Net achieves a precision of 78.29%, recall of 67.91%, Jaccard coefficient of 57.81%, and Dice coefficient of 72.59% on the small-sample cerebrovascular dataset, demonstrating superior performance across all metrics with the best overall performance. Visualization results of diverse segmentation show that the DU-Net model has significant advantages in vascular continuity, branch detail preservation, and complete reconstruction of vascular morphology, allowing for more accurate identification of vascular regions and reducing the occurrence of artifacts and mis-segmentation. To validate the effectiveness of the Squeeze and Excitation module and the collaborative optimization framework, an ablation study was conducted on the small-sample cerebrovascular dataset . Ablation experiments on the DU-Net modules demonstrate that the fine segmentation path using collaborative optimization effectively improves segmentation accuracy, particularly in handling complex, blurry boundaries and fine structures, thus significantly enhancing the overall precision and stability in medical image segmentation tasks. Furthermore, DV-Net and DCSR-Net were deployed based on V-Net and CSR-Net to validate diverse backbones. The results present that the DU-Net outperforms DV-Net and DCSR-Net, highlighting the ease of use and applicability of DU-Net on small-sample dataset. In conclusion, DU-Net possesses stronger feature learning and recognition capabilities, enabling accurate cerebrovascular segmentation on small-sample datasets and broad applications in multiple scenarios.