Medical image segmentation is a crucial step in medical diagnosis. Incorrect segmentation could lead to misdiagnosis, thereby threatening the health of patients. Therefore, improving the accuracy of medical image segmentation has become an urgent problem to solve. The introduction of U-net has attracted the attention of many researchers and improvements have been made to the network. However, not all improvement strategies have an effective impact on the results. Therefore, we re-examined the optimization strategies and found that improvements based on skip connections and feature aggregation strategies can significantly enhance the segmentation performance of the model. We optimized U-net in these two aspects. The original direct skip operation might lead to the loss of important information, so we added a transformer mechanism at the skip connection to focus on different spatial representations, enhancing the ability to extract global and local texture detail information. At the same time, we used a residual network in the downsampling process to effectively solve the problem of insufficient non-linearity due to too little transformation of deep information. Compared with other mainstream methods, the results show that this method is competitive in terms of IOU and Dice coefficients.