The automated segmentation of liver tumors plays an important role in the diagnosis and treatment of liver cancer. As most of the computed tomography (CT) images are 3D structures, we design a 3D-based liver tumor segmentation model based on the UNet architecture. This model introduces the attention mechanism and dynamic convolution method, which effectively improves the feature extraction ability. In the training process, transfer learning is used to transfer the information learned in the liver segmentation task to the tumor segmentation task, which effectively improves the fitting ability of the model. The Dice coefficients of the liver and tumor segmentation results using this model are 94.9% and 53.2%, respectively. Compared with the basic network framework, the segmentation performance can be improved by 4.4% on the tumor segmentation task on average.