Abstract: Cerebrovascular segmentation is a crucial foundation for screening cerebrovascular diseases, monitoring pathological progression, and planning precise surgical interventions. Compared to other magnetic resonance angiography (MRA) modalities, phase-contrast MRA offers significant advantages due to its non-invasive nature, high spatial resolution, and ability to capture detailed flow information, making it particularly valuable for cerebrovascular analysis. However, due to concerns over patient privacy and the inherent complexity of cerebrovascular structures, publicly available cerebrovascular datasets are generally limited in terms of sample size and the extent of manual annotations, restricting the development of high-performance segmentation algorithms. To address these issues, this study proposes a coarse-to-fine segmentation Double U-Net collaborative optimization model (DU-Net), aiming to improve performance in small-sample cerebrovascular segmentation tasks. The DU-Net model consists of two paths, both built upon a three-dimensional U-Net backbone. The coarse segmentation path utilizes lightweight convolutions with fewer channels to minimize the collaborative training burden and achieve preliminary learning of cerebrovascular topology. The extracted features are subsequently transferred to the fine segmentation path, which enhances the ability of the model to generalize global features. In the fine segmentation path, convolutions with a higher number of channels enhance the ability of the model to capture detailed features, boundary information, and small targets. This improves segmentation accuracy and stability, particularly in refining local features and optimizing small target and boundary regions. Additionally, a Squeeze and Excitation (SE) module was incorporated into the skip connections of the fine segmentation path to assign higher learning weights to target areas, significantly enhancing the model representation ability of the fine segmentation path to learn and segment complex regions and detailed features. Based on publicly available resources, this study conducted comparative experiments with ten representative medical segmentation models. Experimental results revealed that the DU-Net achieved a precision of 78.29%, recall of 67.91%, Jaccard coefficient of 57.81%, and Dice coefficient of 72.59% on a small-sample cerebrovascular dataset PCA22, outperforming all baseline models across all metrics. Visualization results of diverse segmentation demonstrated that the DU-Net model offers significant advantages in maintaining vascular continuity, preserving branch details, and reconstructing complete vascular morphology, allowing for more accurate identification of vascular regions and reduction in artifacts and mis-segmentation. To verify the superior adaptability of the SE-Residual module and the effectiveness of the collaborative optimization framework, ablation experiments were conducted on PCA22. The results showed that the collaborative optimization mechanism enabled the fine segmentation path and significantly improved segmentation accuracy and model stability, particularly in handling complex, blurred boundaries and fine structures. Furthermore, double volumetric network (DV-Net) and Double Cross-scale residual network (DCSR-Net) models were constructed based on volumetric network (V-Net) and Cross-scale residual network (CSR-Net), respectively, to assess the performance gains of the collaborative optimization framework across different backbone networks. Results indicated that DU-Net consistently outperformed both DV-Net and DCSR-Net in terms of overall metrics, further underscoring its ease of deployment and strong adaptability to small-sample cerebrovascular datasets. Moreover, cross-dataset validation experiments showed that DU-Net, despite domain shifts in imaging modality and resolution, outperformed most comparative methods on the Time-of-flight MRA (TOF-MRA)-based subset of the public IXI dataset (IXI-Sub), validating its strong generalization capability. In conclusion, DU-Net exhibits enhanced feature learning and recognition capabilities, enabling accurate cerebrovascular segmentation on small-sample datasets and offering broad applicability across diverse scenarios.