DSTUNet: UNet with Efficient Dense SWIN Transformer Pathway for Medical Image Segmentation

Abstract

Automatic medical image segmentation has achieved impressive results with the development of Deep Learning. However, although convolutional neural network, especially the U-shape network, has shown the superiority of method in many segmentation tasks, it can not model long-range dependency well and will be limited by the information recession due to the downsampling operation. Some recent Transformer-based works only used multi-head self attention mechanism in the main autoencoder architecture to enhance the long-range dependency on the single scale, and it failed to compensate for the information loss. In this paper, we propose a novel UNet with densely connected Swin Transformer blocks as efficient skip pathway, namely DSTUNet, for medical image segmentation. Specifically, each Dense Swin Transformer Block is composed of several Swin Transformer layers to make better use of the shift-window self attention mechanism at different scales to enhance the multi-scale long-range dependency. Moreover, the dense connection among Swin Transformer layers is introduced to boost the flow of feature information and minimize the information recession. Experiments have been conducted on multi-organ and cardiac segmentation tasks, and the results demonstrate that our method is able to achieve superior segmentation compared to the existing state-of-the-art approaches.