TU-Net: U-shaped Structure Based on Transformers for Medical Image Segmentation.

Recently, the development of deep learning technology in medical image segmentation has become increasingly mature, and the symmetric U-Net has made breakthrough progress. However, because of the inherent limitations of convolution operations, U-Net has some shortcomings in the interaction of global context information. For this reason, this paper proposes TU-Net based on transformers. TU-Net can strengthen the modeling of global context information, enhance the extraction of detailed information and reduce the computational complexity of the algorithm. In patch embedding, successive convolutional layers with small convolutional kernels are proposed to extract features. Cross Attention-Skip is proposed to complete the fusion of shallow and deep features during the skip connection process. TU-Net is performed on the Synapse dataset to segment eight abdominal organs. The experimental results show that TU-Net is superior to ViT, V-Net, U-Net and Swin-Unet.