Minimally invasive surgery is a specialized surgical technique that permits vascular interventions through very small incisions. It minimizes the patient's trauma and permits a faster recovery compared to traditional surgery. Although traditional invasive surgery training system can complete general training work, real-time performance and accuracy of most training system failed to meet the requirements of training work. Therefore, in this study, three parts, including 3D modeling, collision detection algorithm and application architecture were improved in the existing training system. Firstly, an improved Marching cubes algorithm was adopted to simplify the mathematical modeling of vessels by merging the related points of the mesh model. Secondly, a hybrid collision detection algorithm was proposed and implemented. Lastly, the CPU-GPU parallel computing architecture was adopted. Particularly, the design of the improved VR-based system and the experimental results were presented and analyzed. Moreover, experimental results showed that the proposed system was beneficial to improve the skill of surgeons in manipulating the catheter and guide wire. Thus, the simulators could be used for trial surgery training.