Automatic liver segmentation is a crucial prerequisite yet challenging task for computer-aided hepatic disease diagnosis and treatment. In this paper, we implemented an improved 3D U-net[1] architecture, which achieves a more precise segmentation effect. The proposed 3D U-net takes advantage of dilated convolution [2] that extracts multi-scale feature information and separable convolution[3] that achieve separation of cross-channel correlation and spatial correlation. In addition to the skip concatenation of the down-sampling feature and the up-sampling feature, we add skip concatenation at intervals of two convolution layers during the down-sampling process. The improved 3D U-net produces high-quality segmentation result of liver in CT scans. We also used a post-processing based on liver feature information in CT to optimize the segmentation.