【摘要】 计算机断层扫描(CT)常用于肝脏消融手术的术前规划阶段辅助医生制定手术方案，其中医生比较关注的血管信息是从腹部CT图像中分割出肝脏血管。由于肝脏血管结构复杂，血管与周围组织对比度低，从CT图像中准确分割出肝脏血管难度大，而目前多数分割方法只关注血管像素分布，忽略了血管结构信息，分割结果常存在断裂和孔洞。针对以上问题，本研究提出一种基于血管骨架特征约束的肝脏血管分割方法，具体方法为：1)在数据预处理中加入血管增强步骤，并将增强后图像与原始图像联合作为网络输入，从而在分割网络的输入端引入血管空间注意力；2)在分割网络后加入基于欧拉特征的骨架化模块，将该模块提取的骨架特征加入损失函数中，设计了包含Dice损失和形态学骨架损失(MS＿Loss)的联合损失函数(Dice＿MS＿Loss),以提高分割网络的拓扑保持能力。本研究分别从公共数据集IRCAD中挑选了9例、数据集MSD8中挑选了29例，共38例数据用于算法实验评估。五折交叉实验结果表明，本研究方法在量化评价指标上优于其他SOTA方法，Dice系数达到了0.749、中心线Dice(clDice)达到了0.79、灵敏度(Sen)达到了0.754。实验的视觉效果表明，所提出方法分割出的血管结构断裂和缺失更少。更多还原

【Abstract】 Computed tomography(CT) is widely used in the preoperative planning stage of liver ablation surgery to assist doctors in making surgical plans. It is important to segment liver vessels from abdominal CT images. Due to the complex structure of liver blood vessels and the low contrast between blood vessels and surrounding tissues, it is difficult to accurately segment liver blood vessels from CT images. Most of the current segmentation methods only focus on the pixel distribution of blood vessels and ignore the structure information of blood vessels, so that the segmentation results often have fractures and holes. To solve the above problems, this paper proposed a liver vessel segmentation method based on vascular skeleton feature constraint and applies vascular skeleton features to the segmentation neural network. The detail process of the method included following steps: 1) A blood vessel enhancement step was added to the data preprocessing, then the blood vessel enhanced image and the original image were used as the input of the neural network, so as to introduce vascular spatial attention into the network input. 2) A skeletonization module based on Euler features was added to the post-processing procedure of the segmentation network and the output vascular skeleton feature was added to the loss function to design a joint loss function(Dice＿MS＿Loss) including Dice loss and morphological skeleton loss(MS＿Loss), which serves as a constraint to promote the network’s topology preserve ability. In this study, 9 cases were selected from the public dataset IRCAD, and 29 cases were selected from the dataset MSD8. A total of 38 cases were used for method evaluation. The results of five-fold cross experiment showed that the proposed method was superior to other SOTA methods in terms of quantitative evaluation indicators, with the Dice coefficient of 0.749, the centerline Dice(clDice) of 0.79 and the sensitivity(Sen) of 0.754. The visual effects of the experiments showed that the proposed method was able to effectively segment liver blood vessel structures with less fractures and deficiencies.更多还原