Cardiovascular diseases have long been the leading cause of death globally, even surpassing cancer. According to the 2019 statistics from the World Health Organization (WHO), over 8 million people die from heart-related diseases. Cardiovascular imaging segmentation techniques have become increasingly important as they can assist doctors in the treatment of cardiovascular diseases, making diagnostics more efficient. Although continuous Coronary Angiograms (CAG) are commonly used in cardiac catheterization, the high costs associated with collecting and labeling large quantities of continuous X-ray images have led most vascular segmentation techniques to be developed for single images only. Segmentation using single images alone may overlook important temporal information between images. Existing video segmentation methods, while achieving high accuracy for everyday life videos, cannot be directly applied to continuous CAG as they are not trained to capture the specific structures of coronary artery images. In this study, we explore effective training strategies for continuous CAG segmentation in the absence of labeled video data. Through extensive experiments, we confirm that generic video segmentation networks are indeed suitable for continuous CAG segmentation; we successfully train models that yield satisfactory results even in the absence of explicit labeled video data; additionally, we leverage the characteristics of CAG data to further enhance training and inference outcomes. Experimental results demonstrate that with proper training strategies, we can use generic video segmentation networks for CAG segmentation without the need for explicit labeled CAG video data. In optimal conditions, our trained network achieves an average F1 score of 85.32 and an average region similarity of 74.40, approximately 0.57 and 1.11 higher respectively compared to state-of-the-art techniques. These improvements contribute to enhancing the overall efficiency and accuracy of vessel segmentation in CAG, highlighting the potential of our approach in improving coronary artery disease diagnosis capabilities. Furthermore, our final experiment verifies that our model does not heavily rely on the quality of the first frame, indicating its promising applications in medical practice.