Computed tomography perfusion (CTP) imaging is a commonly used tool for stroke diagnosis, and accurate segmentation of cerebral vessels based on CTP images is of great significance for the diagnosis and treatment of stroke patients. However, the complex tree-like structure of cerebral vessels and inhomogeneous diffusion of contrast agents during dynamic imaging pose significant challenges to manual segmentation or the establishment of supervised automatic segmentation models. This research presents a multi-statistical features-based 4D level set model (MSF-4DLSM), which requires no labeled training samples while achieving segmentation accuracy comparable to fully supervised models. First, the temporal dimension information of the CTP image is introduced into the level set through the variance map to construct a 4D level set combining temporal and spatial features. Then, in the evolution term of the model, a mixed image of the fixed-point enhanced temporal variance map and the spatiotemporal mean map is used to replace the basic variance map, which improves the model’s adaptability to the uneven diffusion of contrast agents. Finally, to improve the continuity of the vessel segmentation results, we added optical flow information to the model evolution term. The proposed method was validated on 50 CTP images, and achieved an encouraging performance, with Dice, Sensitivity, Specificity, Accuracy and Jaccard of 0.90, 0.92, 0.94, 0.99 and 0.85, respectively. The results are close to those of the supervised model U-Net (Dice: 0.91, Sensitivity: 0.93; Specificity: 0.95; Accuracy: 0.99; Jaccard: 0.87).