Abstract: To enhance hepatic vessel structure segmentation in the absence of large amount of hepatic vessel label information, this paper proposed a local-global-adjacent information fusion module and embeded it into the encoding phase of U-Net, thus obtaining a structure-aware hepatic vessel segmentation network. The module has three branches, including residual convolution module extracting local information of feature maps, long distance information extraction module using the self-attention mechanism to extract global information of feature map and adjacent information extraction module using adjacent slices to supplement contextual information. By fusing the output feature maps of the above three branch modules, the vessel structure perception capability of the network can be effectively improved, and the problems of the 2D network unable to represent the vessel's three-dimensional direction and insufficient training data of the 3D network can be effectively alleviated. In this paper, extensive comparative experiments were conducted on the hepatic vessel and tumor datasets in the Medical Segmentation Decathlon dataset and the hepatic vessel dataset collected and labeled by us in a third-class hospital. Results show that our method achieves the optimal performance compared with several mainstream U-Net segmentation algorithms. The proposed method's Dice reaches 64.04% on the MICCAI dataset and 72.07% on the dataset gathered by us.