Research On CT Image Aided Diagnosis Of COVID-19 Based On Deep Learning

COVID-19 is an acute respiratory infection with high infectiousness,high variability and long incubation period.The global outbreak of the COVID-19 epidemic has overwhelmed the medical system,economic situation and people’s lives,so the rapid diagnosis of COVID-19 is of great significance for the follow-up treatment and control of the outbreak.However,the detection rate of nucleic acid testing is low and the waiting time is long,so a large number of CT imaging examinations are easy to cause the risk of misdiagnosis and missed diagnosis by doctors.To address the above pain points,this paper proposes a deep learning-based CT image assisted diagnosis system for COVID-19 based on existing research methods,through data processing,network design,experimental comparison,and model deployment.After the validation of multiple test sets and multiple evaluation metrics,the system can achieve lung parenchyma segmentation,lung slice classification and COVID-19 lesion segmentation quickly,accurately and end-to-end,which greatly improves the diagnostic efficiency of doctors.The main research and innovation points of the paper are as follows.To address the problem that the CT signs of COVID-19 and common pneumonia are highly similar leading to difficult recognition,a tri-classified Vi Linformer network for COVID-19,common pneumonia and normal slices is proposed,which introduces the Linformer self-attentive mechanism based on the Vi T network,uses serialized image blocks as the input of the self-attentive mechanism,and uses the matrix low-rank decomposition method to the time complexity and space complexity of the operation are reduced from O(n~2)to O(n).By comparing the performance with classical convolutional neural networks,Vi Linformer can achieve close to most complex models with very small number of parameters,and the AUC is 0.98.The SAUNet++network fusing squeeze excitation residual modules and the atrous spatial pyramid pooling module is proposed to address the problem that existing methods segment the lung parenchyma and COVID-19 lesions with low accuracy.The SER module assigns more weights to more important feature channels and alleviates the gradient disappearance problem,and the ASPP module captures multi-scale information by parallel atrous convolution with different sampling rates.The experimental results show that SAUNet++is better than the most advanced model in many segmentation metrics,and the sub-module has a significant improvement effect,and the Dice coefficient reaches 87.38%.To address the problems of micro lesions and fuzzy boundary segmentation in COVID-19,a SAUNet++network with dual outputs of mask map and level set distance map is proposed,and the level set generalized dice loss function LGDL can focus on both area and boundary information of the segmentation target when guiding the model training,which can effectively segment micro lesions and has certain noise immunity.The sensitivity achieved experimentally is 94.27%,and the HD distance is only 17.12 mm.