STC-Net: Fusing Swin Transformer and Convolution Neural Network for 2D Medical Image Segmentation

Abstract

Convolutional Neural Network (CNN) has made significant development in intelligent medical image analysis. However, due to the limitations of convolution, it cannot well learn a global and long sequence of semantic information, which also limits its performance in clinical medical segmentation. To solve the above problems, STC-Net proposed in this paper combines the global image information processed by Swin Transformer with the low-level detail features processed by CNN. To fuse the two kinds of information, a fusion module with channel attention mechanism and spatial attention mechanism (CSM) is proposed in the information fusion part of STC-Net. In the channel attention part, local channel information interaction is realized by the Fast 1D convolution method, while the spatial attention part focuses on important spatial location, scale and other information. The CSM can focus on this important information and suppress the non-target segmentation information. The results show that STC-Net achieves good performance in brain tumor segmentation, polyp segmentation and skin lesion segmentation.