Abstract: To address the issues of fuzzy boundaries, diverse tumor types, low contrast with surrounding tissues in computerized tomography (CT) images, as well as insufficient utilization of texture information in existing networks for medical images, an automatic liver tumor segmentation method that combines multi-branch texture feature extraction and attention mechanism is proposed. Firstly, a parallel convolutional encoder is designed to replace the dual convolutional modules in the U-Net baseline network, aiming to extract superficial features under two different receptive fields. Secondly, a texture feature extraction network is introduced in the skip-connection part of the U-Net to extract deep texture information of liver tumors at multiple scales. Finally, a channel attention module with a residual path is incorporated in the decoding stage to effectively capture inter-channel dependencies and enhance the relevant features for liver tumor segmentation tasks. The proposed method is evaluated on the LiTS2017 and 3DIRDCADb-01 liver tumor segmentation datasets. Experimental results demonstrate superiority of the proposed method in terms of evaluation metrics and visualizations in comparison with the baseline methods, which shows advantages in segmenting small-sized and blurry boundary tumors, providing promising insights for liver tumor screening.