To address the low accuracy of traditional deep learning algorithms in medical image segmentation tasks and the unclear boundaries in the dataset of lung nodule images, an improved Unet network, named DM-Unet, was proposed. The model was enhanced through various optimization strategies to improve segmentation performance.First, a multi-scale residual convolution layer was employed to replace standard convolutional layers, enabling the network to capture feature information at different scales more effectively by increasing its depth. Second, in the decoder stage, a multi-level boundary enhancement modules was introduced. This module utilized upper-layer masks and an attention mechanism to enhance boundary recognition capabilities. Finally, during the skip connection stage, a detailed feature cascade module was embedded to strengthen the perception of multi-scale contextual features, and the final result was obtained through upsampling.Experimental results demonstrate that DM-Unet achieved superior performance on public pulmonary nodule image datasets, with IoU of 79.84%, Sensitivity of 87.33%, dice similarity coefficient of 88.62%, and positive predictive value of 90.19%, thereby fully validating the effectiveness of the proposed method.