Surgery and chemotherapy, as the main treatments for liver cancer, require accurate extraction for the liver lesion area. Therefore, to solve the problems of the current segmentation methods for liver tumors, such as the loss of small-sized tumors, fuzzy segmentation of tumor boundaries, and severe missegmentation, a new method for liver tumor segmentation based on the attention mechanism and deformable residual convolution is proposed. U-Net was used as the backbone network, and a residual path with deconvolution and activation function was added at the end of the encoding convolution layer to connect with the upper layer, thereby solving the problem of missing small target segmentation and fuzzy boundaries caused by information loss in pooling and deconvolution operations. Furthermore, a deformable convolution was used to enhance the model for extracting features of tumor boundaries. Several convolution layers were added to the skip connection layer to compensate for the semantic gaps caused by simple skip connections in feature fusion. The model pays more attention to tumor characteristics through the dual-attention mechanism. The mixed loss function was used to address the problem of segmentation performance degradation caused by a class imbalance under the condition of ensuring the stability of training. The experiment was carried out using the Liver Tumor Segmentation Challenge (LITS) dataset. The experimental results show that the Dice coefficient of tumor segmentation of the proposed method reaches 85.2%. Moreover, the proposed method has a better segmentation performance than other comparison networks, meeting the requirements of auxiliary medical diagnosis.