Localization of liver tumor lesions is a prerequisite for the treatment of liver cancer, such as radiation therapy, surgical resection and arterial embolization chemotherapy, so accurate segmentation of liver tumors is very necessary. Network models based on U-Net and Transformer have achieved great success in medical image segmentation tasks. However, the backbone network based on convolutional neural network (CNN) in U-Net has some shortcomings. Due to the limitations of convolutional operation, the network is difficult to capture the relationship between distant elements in the feature map. The Transformer structure itself emphasizes the global relationship, and its processing capacity of local information is relatively limited. The above problems lead to missing segmentation and inaccurate localization of the edge ability of the region of interest in the small volume tumor segmentation task Therefore, we propose a new segmentation framework, FUF-TransUnet. Specifically, the framework includes two submodules, Displaceable Module (DS Module) and Skip Connection Feature Fusion Module (SFF Module). Among them, the DS Module increases the reusability and richness of features by changing the position of elements in the feature map, so that the module can explore the relationship between features at different distances. In addition, we rethink the design of skip links in U-shaped networks, and propose another submodule, SFF Module, which balances the semantic complexity between codec output features, so that the network can better integrate the details of different scale feature maps at skip links. We conducted experiments on LiTS data set and 3DIRCADb data set, and proved that adding DS Module or SFF Module separately can improve the segmentation task index. Further, better tumor segmentation results can be obtained by using both modules simultaneously. In LiTS dataset and 3DIRCADb dataset, the Dice index of liver tumors could reach 83.41% and 89.77%, respectively.