基于CTA图像的两阶段U-Net冠状动脉分割

基于CT血管造影(computed tomography angiography，CTA)图像的冠状动脉自动分割的挑战在于冠状动脉结构复杂、前背景分布严重不平衡，分割时易受冠状静脉和其他组织的干扰.提出了一种两阶段的冠状动脉分割算法，第一阶段采用具有密集特征提取和残差特征修正能力的3D DRU-Net进行分割，保证分割的召回率;在第二阶段提出2D双编码多特征融合U-Net(2D DEMFU-Net)进行细分割，先对原始图像和第一阶段分割结果分别进行特征提取，再采用密集跳跃连接融合两个分支上的多层次语义特征，进一步提高分割准确性.实验结果表明，提出的两阶段分割算法在CortArt2020数据集上的Dice相似系数、召回率和精确度分别优于3D U-Net网络3.83%，5.31%和2.23%.

Two-Stage U-Net Coronary Artery Segmentation Based on CTA Images

There are many challenges in coronary artery segmentation based on CT angiography(computed tomography angiography)images， such as the complex structure of the coronary artery， the severely unbalanced distribution of the coronary artery and the background， and the susceptibility to interference from the coronary veins and other tissues during segmentation. A two-stage coronary artery segmentation method was proposed. In the first stage， the 3D DRU-Net with dense feature extraction and residual feature correction capabilities was used to ensure the segmentation recall rate. In the second stage， a 2D dual encoding multi-feature fusion U-Net(2D DEMFU-Net)was developed to perform fine segmentation， in which features were extracted from the original image slices and the segmentation results produced by the first stage respectively， and the multi-level semantic features from these two branches were fused using dense skip connections， further improving the accuracy of segmentation. Experimental results show that the proposed two-stage segmentation algorithm outperforms 3D U-Net by 3.83%， 5.31% and 2.23% in terms of Dice similarity coefficient， recall rate and precision， respectively.