Since the introduction of UNet in 2015,U-shaped architecture has become an important paradigm in the field of medical image segmentation. However, due to the inherent local limitations of convolutions, a fully convolutional segmentation networks with U-shaped architecture cannot effectively utilize global context information. Although the combination of transformers and CNNs can address these issues, it also brings larger computational cost. At the same time, Due to the specificity of medical scene, scarce medical resources will be an obstacle to the application of transformers in the medical field, which is well adapted by the inductive bias in lightweight networks. For this reason we propose CS-UNet, a lightweight fully convolutional model that enables fast image segmentation in real medical scenarios. In this model, we reconstruct the encoder and decoder to generate a dense receptive domain to extract contextual information. We evaluate CS-UNet on five different types of medical image datasets, and found that CS-UNet excelled in segmentation performance. While ensuring high accuracy, CS-UNet has faster computation speed and smaller model parameters, which achieves a better balance between performance and computation cost, and is less constrained by the environment as well as the equipment in practical application scenarios.