Research On The Segmentation Method Of COVID-19 CT Images Based On Deep Learning

In the clinical diagnosis of patients with new coronary pneumonia,it is the primary task of medical workers to use chest CT images to segment lesions and obtain the status of lung infection.Due to the large number of patients in our country and the shortage of medical resources,in recent years,the use of deep learning to segment lesions to assist manual diagnosis has shown high application value in the field of medical images.However,different stages of COVID-19 have different lesions on CT images,and it is not enough to qualitatively diagnose patients with complex lesions in middle and advanced stages only by segmenting the total lesion area to determine whether they are infected with COVID-19.Based on this,this thesis designs a two-stage new coronary pneumonia lesion segmentation method by studying the relevant knowledge of deep learning.work to help.The main research is as follows:(1)In the segmentation task of the CT image lesion area of the new coronary pneumonia,in order to realize the fine segmentation of the edge of the area,the EG-Net segmentation model is designed.The model uses U-Net as the basic structure.On the encoding side,Res Ne Xt50 is used as the feature extraction network,and an edge feature extractor is introduced after the first and second Res Ne Xt50 residual blocks to collect local edge information;the downsampling and upsampling operations in the last layer of the model are replaced by The dilated convolution expands the receptive field of the model without losing the details of the feature map;the crossentropy(CE)loss function is used during the experiment to guide the model to distinguish the lesion area from the non-lesion area.Compared with other models,the EG-Net segmented lesion area is more accurate and the segmentation performance is better.(2)In the segmentation task for multiple types of lesions,in order to avoid the influence of background noise and make the model focus on learning the characteristics of multiple types of lesions in the infected area,first obtain the lesion area of the CT image of new coronary pneumonia through EG-Net,and then the lesion The region is jointly input as a Mask and the original CT image to the multi-class lesion segmentation model PA-Net.In order to achieve multi-scale feature extraction for different types of lesions,PA-Net improves the skip connection method of the original U-Net,introduces pyramidal convolution blocks,and enhances the ability of the model to obtain multi-scale features by setting convolution kernels of different sizes;Through dense layer-skip connections,the feature maps extracted by the multi-layer encoder are integrated as the input of the pyramid convolution block to supplement the input with richer semantic information;the decoding end introduces and improves the double attention module,and embeds the decoder unit to strengthen The ability of the model to learn the characteristics of small lesions.During the experiment,aiming at the problem of category imbalance between samples,the focal loss function was sampled to further guide the model to segment small target categories.Compared with other segmentation models,the method in this chapter has higher accuracy in multi-class lesion segmentation.