Abstract: The segmentation of coronary tomography (CTA) is the first step in judging vascular occlusion, and it is also a prerequisite for subsequent medical diagnosis such as three-dimensional reconstruction and qualitative analysis. This paper proposes a multi-stage method to complete the segmentation of coronary arteries from coarse to fine. In order to reduce the impact of non-cardiac tissue on neural network training, the adaptive threshold-based method is first used to pre-extract the heart region. Then, a deep full-convolution network with V-net as the basic network framework is proposed, which expands the third-dimensional channe of the convolution kernel at each layer , makes full use of the continuity of blood vessels, and increases the network learning ability. The heart region extracted in the first stage is combined with the corresponding label as the training data of the next-stage full convolution network to achieve accurate coronary vessel segmentation. Finally, the blood vessel edge contour is iteratively optimized by the level set function to obtain the segmentation result.The proposed method has an average Jaccard of 0.813 for vessel segmentation and a Dice of 0.903, which enables accurate three-dimensional segmentation of coronary CTA.