This paper proposes a novel UNet model called DCR-UNet, which combines Dilated Convolution (DC), Convolutional Block Attention Module (CBAM), and Residual Module (ResNet), and adds a three-layer UNet model to the decoder of the original UNetl. The DCR-UNet is applied to lung CT image segmentation. In order to improve the accuracy and stability of segmentation, based on the traditional UNet, the traditional convolution operation is first replaced with dilated convolution to expand the receptive field and enhance the UNet’s ability to capture global information of the image. Secondly, a hybrid attention mechanism was introduced, which improves the attention of UNet to important regions by simultaneously focusing on global and local information, thereby enhancing the accuracy of segmentation. In addition, to alleviate the problem of gradient vanishing during UNet training, a residual module was added to improve the convergence speed and stability of UNet. Through experimental verification, the DCR-UNet proposed in this paper has improved the Intersection over Union (Iou), Dice Similarity Coefficient (DSC), and Accuracy of segmentation results by 6.7%, 7.2%, and 2.73%, respectively, compared to traditional UNet on publicly available lung CT datasets. This indicates that the DCR-UNet has higher accuracy and robustness, and can provide more reliable support for the diagnosis and treatment of lung diseases.