Research On Medical Image Segmentation Method Based On U-Net Network

Medical image segmentation is one of the hot research directions in the field of computer vision.Many scholars try to use deep learning to deal with related tasks in this field.In recent years,convolutional neural network is undoubtedly one of the main neural networks to promote the development of deep learning.Its significant advantage is to use local operations to extract features hierarchically.However,with the deepening of the study,researchers found that the size of the convolution kernel in the convolution layer of traditional convolutional neural networks would limit the scope of extracting target features,resulting in the lack of long distance feature information from the global scope of the network segmentation results.To solve this problem,this paper explores whether the introduction of Transformer structure in natural language processing tasks in the field of medical image segmentation can alleviate the limitations of convolutional neural networks.The details are as follows:(1)Swin E-UNet3+ network is proposed to solve the problem that convolution neural network cannot capture long distance feature dependence due to the limitation of receptive field in medical image segmentation.In this thesis,the Transformer variant model is introduced into the U-Net3+ network,in the encoder layer two continuous Swin Transformer blocks are used to replace the original encoder layer structure.The Swin Transformer blocks can learn the advantages of long-distance features in the image to enhance the ability of the encoder to extract feature contour information,and Patch Merging is used between the encoder layers to replace the Max-pooling method to complete the downsampling operation.The encoder layer information is transmitted to the decoder layer through multi-scale skip connection,and this information and the previous decoder layer after upsampling output feature information are aggregated in the decoder layer to complete the segmentation of the target feature.Swin E-UNet3+ is evaluted on the Tip DM Cup rectal cancer dataset and ISIC2017 dataset.Experimental results show that Swin E-UNet3+ model outperforms U-Net,U-Net++ and U-Net3+models in Dice coefficient,IOU value and Precision evaluation metric.(2)CSE-Trans Net network is proposed to improve the performance of U-Net in cell nucleus segmentation.The CSE-Trans Net network structure consists of three parts:encoder,decoder and skip connection.The encoder and decoder structure uses the UCTrans Net network for feature extraction,and a CSE-Transformer method is proposed at the skip connection.This method improves q and k calculation methods,so that q can perform self-attention mechanism calculation on k in the global range,and enhance the ability of the Multi-Head Self-Attention to capture long-distance feature information;In the MLP network,Squeeze-and-Excitation block is added to enhance the channel attention mechanism after the last fully connected layer and deep separable convolution is added to the residual path to extract local neighborhood information.By using the CSE-Transformer method,the CSE-Trans Net network takes into account the local information of features on the basis of capturing long-distance feature information,and achieves excellent segmentation performance in the Mo Nu Seg nucleus experiment.