Automatic segmentation of retinal vessels plays a crucial role in the diagnosis of fundus diseases. Despite significant advancements made by U-shaped networks, they face two main limitations. First, they struggle to capture fine details and contextual information in raw images, especially in regions with poor image quality and contrast, leading to challenges in differentiating between vessels and background. Second, their dependence on traditional convolution with fixed sizes and shapes limits their adaptability to vessels with varying forms, particularly slender and tortuous local features. To address these limitations, we introduce 3VNet, comprising a 3V-shaped framework driven by skeleton topology. This architecture employs a skeleton supervision mechanism to capture and map vascular structures, refining features progressively through a cascaded model with skip connections. To comprehensively extract intricate vessel topological details such as bends and bifurcations, we developed the Multi-View Deformable Vessel-shape Extract Unit (MDVEU). This unit is seamlessly integrated into both the encoding and decoding stages of 3VNet, dynamically capturing vessel topology from multiple perspectives and employing adaptive fusion techniques for precise and detailed feature utilization. Extensive experiments on DRIVE, STARE, and CHASE_DB1 reveals that our method achieves higher accuracy than current approaches, surpassing the state-of-the-art (SOTA) on all datasets. Sensitivity exceeds SOTA by 1.71% and 1.30% on the DRIVE and CHASE_DB1 datasets, respectively. The source code is released at https://github.com/wangwxr/3VNET .