Lung Segmentation Of CT Image Based On Convolutional Neural Network And Regional Segmentation Of Pneumonia

In the early stage of novel coronavirus(COVID-19)patients,small patches appeared in the lungs,and with the aggravation of the disease,vitreous structures appeared.Because the lung CT image is noisy and the early pneumonia area is too small to identify,the traditional image segmentation method is easy to lead to misdiagnosis and missed diagnosis.Using deep learning technology to segment the image can provide reliable diagnosis basis for practical clinical medical application,and improve the doctor’s work efficiency and the correct rate of disease diagnosis.The main content of this paper is the research of COVID-19 lesion segmentation in lung CT images based on convolutional neural network.The research contents of this paper mainly include the following:(1)lung image segmentation algorithm based on U-Net-SCB network is proposed.In U-Net-SCB network,SCB module replaces the original convolution layer and extracts more image information;The spatial hole pyramid structure is introduced into the improved network to obtain image features with different scales of context,and attention mechanism is introduced into each layer of decoding and encoder to focus on image feature information;It solves the problems of network degradation and poor segmentation effect of COVID-19 lesions in the segmentation of lung images by U-Net model.The model test shows that the accuracy of the proposed method reaches 97.3%.(2)A lung image segmentation algorithm based on U-Net network is proposed.In order to solve the problem of poor segmentation performance of pneumonia region in the above-mentioned method,the model introduces dense connection module on the basis of traditional U-Net,which strengthens the information transmission between feature maps of all layers;The spatial cavity pyramid structure retains more image features;Finally,attention monitoring mechanism is adopted when learning deep feature information,so as to extract the feature information of lung elements more effectively.After testing the data collected in Haikou People’s Hospital,the accuracy of the proposed method reaches 97.6%.This method can effectively segment the elements of lung and pneumonia accurately,and can be applied to the development and construction of image processing system.