U-Net with Transformers: Advancing Medical Image Segmentation

The U-shaped network architecture, featuring an encoder-decoder structure, has quickly gained widespread adoption in medical image analysis. However, as previously noted, CNN-based encoder-decoder models encounter challenges in extracting global information from medical images, hindering their ability to effectively address intricate medical image segmentation tasks. Recent advancements have focused on integrating transformers with U-shaped networks. TransUNet incorporates transformer layers into the encoder portion of the U-Net, while Swin UNet designs a U-shaped network using 12 Swin Transformer blocks. U-Transform, on the other hand, augments the classic UNet architecture with MHSA and MHCA modules. These innovative approaches showcase the potential of transformers to enhance U-Nets' performance in medical image segmentation.