DS transunet: Dual Swing transformer u-net for medical image segmentation

Thanks to the development of deep learning , Medical image automatic segmentation technology has made great progress . However , Most of the existing methods are based on Convolutional Neural Networks （CNN）, Due to the limitation of receptive field in convolution operation , Unable to establish long-term dependencies and global context connections . suffer Transformer The inspiration of success , Some researchers have spent a lot of energy designing based on Transformer Of U-Net A robust variant of ,Transformer The self attention mechanism of has a strong ability to model remote context information . Besides , Patch segmentation used in visual converter usually ignores the pixel level inherent structural features in each patch . To alleviate these problems , This paper presents a new framework for deep medical image segmentation , Called Double Swin-Transformer U-Net（DS-Transune）, This may be the first attempt to layer Swin Transformer The advantages of both into the standard Ushaped Encoder and decoder of the architecture , To enhance the semantic segmentation quality of various medical images . Different from many previous converter based solutions , What this article puts forward DS-transune First, based on Swin-Transformer To extract coarse-grained and fine-grained feature representations of different semantic scales . As DS Transune Core components , We propose a well-designed transformer interactive fusion （TIF） modular , Through the self attention mechanism, we can effectively establish the global dependence between different scale features , To make full use of these obtained multi-scale features . Besides , We will also Swin Transformer The block is introduced into the decoder , To further explore the remote context information in the upsampling process . A large number of experiments in four typical medical image segmentation tasks have proved DS-TransUNet The effectiveness of the , It shows that our method is obviously superior to the most advanced method .

Index words ; Medical image segmentation ; Remote context information ; classification Swin transformer ; Double scale ; Transformer interactive fusion module .

I. INTRODUCTION

Medical image segmentation is an important and challenging