Vascular segmentation is a crucial task in medical image analysis. The task has three challenges: (1) Low intensity contrast; (2) Complex geometry of blood vessels; (3) Sensitive and thin local structures. Although fully or weakly supervised learning methods have been intensively developed to address these challenges, they lack either a specific design for topological structures or a metric to evaluate the accuracy of topological structures. In this work, we define a type of topological structures, topological shape points (TSP or TS-points), for blood vessels. Based on these TS-point structures, we propose a metric called Topological Shape Point Dice (tspDice), which can be utilized to evaluate the segmentation accuracy of shape structures and their topological connectivity. Furthermore, in order to integrate TS-points into the loss function for preserving geometric and topological structures, we propose a differentiable algorithm to extract a TS-points map from vessel prediction probabilities. Lastly, we introduce TSP-Warp-X losses, which utilize the TS-points map to warp various X losses, applicable in both fully supervised and weakly supervised learning scenarios. These losses enhance the neural network's capacity to preserve shape structures and topology for vessel segmentation. Experiments conducted on three datasets, under both fully supervised and weakly supervised learning conditions, show that our proposed TSP-Warp-X losses lead vessel segmentation with better accuracy in preserving geometric and topological structures. Code is publicly available at https: //github.com/orangeqqq/KeypLoss.