Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation

Swin-Unet adapts the U-Net design to a fully Transformer-based architecture for medical image segmentation, motivated by the inability of convolutional networks to model global and long-range dependencies. Input images are split into tokenized patches and fed through a U-shaped encoder-decoder with skip connections. The encoder is a hierarchical Swin Transformer using shifted windows to capture context features, while a symmetric Swin Transformer decoder equipped with patch-expanding layers performs up-sampling to recover the original spatial resolution.

With inputs and outputs down- and up-sampled by a factor of four, the pure-Transformer network was tested on multi-organ and cardiac segmentation tasks. It outperformed competing approaches based on full convolution or on combinations of transformers and convolutions, demonstrating that a convolution-free U-shaped Transformer can learn stronger local-global semantic features. The authors released code and trained models publicly, and the design became a widely referenced template for Transformer-based medical segmentation.