Accurate image segmentation plays a crucial role in the development of computer-aided diagnosis. The U-Net architecture based on Convolutional Neural Network (CNN) is widely adopted, but the locality of convolutional operations limits its ability to process global information, and Transformer can generate feature representations with more contextual relevance. To further explore the impact of the fusion of CNN and Transformer on medical image segmentation, we propose a network called SBC-UNet, which integrates the SBCFormer into the U-Net architecture. Specifically, we use SBCFormer as the core of the encoder, and the improved hourglass attention in SBCFormer reduces the impact of local images detail loss caused by low resolution, which is conducive to learning complete contextual information. We introduce Convolutional Block Attention Module (CBAM) and Multi-Path Feature Extraction Module (MFEM) on the basis of the primordial decoder, enhancing the comprehensiveness of features. Extensive experiments on Synapse multi-organ segmentation dataset and ISIC2018 Challenge dataset show the effectiveness of our model compared to the classical methods.