Research On COVID-19 Ct Image Segmentation Technology Based On Improved U-Net

The COVID-19 first appeared on the scene in late 2019,rapidly sweeping the world in just a few years.The number of confirmed cases is rising rapidly due to the spread of the epidemic,and different strains of the COVID-19 are derived in the development.Although the COVID-19 has been classified as group B and B tube,how to diagnose and treat the COVID-19 effectively is still a research hotspot.With the development of society,image segmentation based on deep learning has been widely used in the medical field and has been highly recognized by the industry.The diseased tissues in the Computed Tomography(CT)images of COVID-19 infected patients are called infected areas.The task of this thesis is to separate the COVID-19 infected areas in the CT images.However,the COVID-19 infected areas are widely distributed in CT images,with uneven density,blurred boundaries,different sizes,and difficult to identify when they adhere to the pleura.In order to solve these problems,this thesis adjusts the window width and window level of COVID-19 CT image.Based on U-Net network,this thesis improves dense connection module in the codec layer to enhance feature extraction and reuse,obtain more features,and at the same time adds attention mechanism to improve the model’s attention to effective features.Experimental results show that the improved algorithm has better segmentation effect.The specific work of this thesis is as follows:(1)To solve the problems such as small target and inconsistent size of COVID-19 infected areas in CT images and further improve the U-Net segmentation effect and generality,this thesis first adjusted the window width and window position of the COVID-19 CT image data set,and improved the CT value of the image to distinguish normal and infected areas.Meanwhile,based on the U-Net model,improved attention modules are used in the model.For the feature image after up-sampling,through the channel attention module and the feature image from skip connection through the spatial attention module are used to obtain the classification information and location information of the feature images.Secondly,Batch Normalization(BN)was added to the model and cosine annealing strategy was introduced in the training to accelerate the model convergence.Finally,the segmentation performance of the model was verified by comparing with other networks.Compared with U-Net network,the Dice Coefficient(Dice)of the designed model increased by 1.8%,the Intersection over Union(Io U)increased by 2.9%,the Recall increased by 4.1%,and the Precision increased by 2.5%.(2)In order to further improve the segmentation effect of COVID-19 micro-infected areas under CT images,this thesis designs the COVID-19 CT image segmentation model based on dense connection and attention fusion.The model mainly does two aspects of work.On the one hand,an improved dense connection module is introduced to replace the multi-convolution series operation of the traditional U-Net;on the other hand,a dual attention fusion module is designed to realize the re-calibration of features and the atrous convolution at the bottom of the network coding layer is used to improve the receptive field and realize the multi-scale fusion of features.The designed model was trained and compared with other models on the data set,and the segmentation performance of the model was verified by evaluation indexes and infected areas segmentation results.Experimental results show that,compared with the model designed in Chapter 3,the COVID-19 CT image segmentation model designed in Chapter 4 of this thesis has increased Dice by 2.7%,Io U by0.6%,Recall by 3.1%,and Precision by 3.9%.The overall segmentation effect is better,the segmentation performance of micro-infected areas is more prominent,and the model has better generalization ability.