Abstract: To address the current challenges in liver CT image segmentation, such as blurred boundaries and the under-segmentation of small target regions, this paper proposes MSDA-Net (Multi-scale Dilated Attention Network), an improved liver segmentation algorithm based on the U-Net model. First, pyramid convolution replaces traditional convolution in both the encoder and decoder, enabling the precise capture of multi-scale features. Second, a multi-scale attention aggregation module is introduced into the skip connections to effectively aggregate features at different scales. Furthermore, a densely connected attentive dilated spatial pyramid pooling module is employed at the bottleneck layer to construct a denser feature pyramid structure and expand the receptive field, thereby fully mining and leveraging contextual semantic information. Finally, the proposed algorithm is validated on the public LiTS17 dataset. The results show that the Dice coefficient, IoU coefficient, recall, precision, and VOE coefficient are 92.77%, 89.19%, 93.47%, 93.46%, and 0.1079, respectively. Compared with current mainstream liver segmentation algorithms, the proposed MSDA-Net algorithm demonstrates improved performance in the fine segmentation of local liver structures by accurately capturing subtle features along liver edges.