Research Of Brain Tumor Segmentation Algorithm Based On Transformer And CNN Fusion

Convolutional Neural Networks based on deep learning have been widely used in the field of medical images.Due to the limitations of convolution operations,Convolutional Neural Networks cannot well learn a global and long sequence of semantic information,which also limits its performance in medical image segmentation.Transformer processes images into a sequence of image blocks for prediction,using a self-attention mechanism within it to enhance the extraction of global information.But its low-level detail processing capabilities are limited.To solve the above problems,combining the global image information processed by Transformer with the low-level detail features processed by CNN can better and more effectively capture brain tumor image features.Swin Transformer builds Transformer in a hierarchical manner,and its internal multi-head self-attention mechanism based on moving windows(SW-MSA)can effectively enhance information exchange between image blocks.Therefore,this article intends to design a brain tumor segmentation algorithm based on improved Transformer and CNN fusion.The main work is as follows:(1)A new parallel fusion network is proposed,which combines the global semantic information extracted by Swin Transformer at different levels with the detail features extracted by Resnet 101 at different scales.It not only ensures the integrity of segmented region information,but also overcomes the limitations of traditional convolution operations.(2)A new fusion module CSM is proposed,which combines the channel attention information of Swin Transformer and the spatial attention information of CNN in a parallel manner.CSM makes the entire network pay more attention to the useful channel information and spatial information in brain tumor images,and suppresses irrelevant information in nontarget regions.Based on the integration of the above two works,compared to Trans Fuse last year,the mean Dice of brain tumor segmentation based on Swin Transformer and CNN fusion network proposed in this article has increased from 93.92% to 95.28%,and the mean Io U has increased from 88.65% to 91.05%.Compared to U-Net,the mean Dice of brain tumor segmentation has increased by 4.02%,and the mean Io U has increased by 6.31%.