In order to solve the problem of low accuracy of segmentation model caused by the complex shape of liver tumor and blurred boundary with surrounding normal tissues in the liver tumor image, this paper proposes a novel liver tumor image segmentation model HFU-Net based on hybrid dilated convolutions and high-level feature fusion. In this model, a high-level feature fusion recalibration module is added to enrich the skip connection part of U-Net, so that it can calibrate the feature information by using feature fusion and squeeze and attention module to enhance the ability of network encoder to obtain feature information. And, in order to further improve the feature extraction effect of each layer of the network, the conventional convolution module in the original model’s encoding network is replaced by the hybrid dilated convolution to obtain dense tumor feature information and expand the network’s receptive field. The experimental results show that Dice coefficient, volumetric overlap error (VOE), sensitivity and precision are improved by 3.3%, 4.59%, 4.39% and 2.04% respectively compared with the U-Net algorithm. The proposed model significantly improves the segmentation precision of liver tumor images, and provides a reliable basis for the diagnosis and treatment of liver cancer.