Liver CT imaging is the main imaging tool for liver tumor examination. Due to the large individual variability of liver tumors in terms of size, shape, and location, and the complexity of grayscale features in the tissue region of the lesion, it is difficult to make image segmentation of liver tumors. U-Net based on deep learning has made a lot of progress in medical image segmentation, so this study intends to establish a method for liver tumor segmentation by improving the characteristics of liver tumor segmentation on the basis of U-Net. Aiming at the problems in liver tumor segmentation, the U-Net segmentation network is improved by introducing the residual module and the attention mechanism, and the group normalization strategy is used to process the feature map along channels for feature learning, so as to improve the learning ability of the network and the generalization ability of the model. In this study, the segmentation network was tested on the LiTS2017 liver tumor image segmentation challenge dataset provided by CodaLab, and the results show that the average Dice coefficient reaches 0.773, which is higher than that of other image segmentation networks, the precision index reaches 0.983, and the recall index is 0.637, which indicates that the RA-UNet network model is more effective in segmenting images and provides a reliable basis for the clinical diagnosis of liver tumors.