Research On Computer Aided Diagnosis Technology Of COVID-19 Based On Deep Learning And CT Images

With the high infectivity of COVID-19,early diagnosis and treatment are crucial factors in reducing the losses caused by the epidemic.At the beginning of the epidemic,reverse transcription polymerase chain reaction(RT-PCR)was used as the "gold standard" for diagnosing COVID-19 infection.Still,there were problems,such as long detection cycle and insufficient sensitivity.Experts proposed the use of CT imaging detection combined with nucleic acid testing for mutual assistance in detection for the people who has gotten some COVID-19-liking symptoms.However,using CT imaging detection will greatly increase the number of medical images,and there is a need for more radiologists.Long-term overloading will enable doctors to make more missed diagnosis and misdiagnosis.Medical auxiliary diagnosis technology based on deep learning can help doctors complete disease screening,improve work efficiency,and reduce misdiagnosis and missed diagnosis rates.In order to solve the problems such as the fuzzy edge information of COVID-19 image features,and the U-Net network tends to ignore local edge features and small targets,a Multi-Scale Attention Network is designed.Through global pooling layers and set the sampling rate of the atrous convolution,it can be increased the network’s perceptual field and the ability to capture multi-scale information for effective segmentation of large targets.Utilizing channel attention and spatial attention can select features in the channel dimension and model the correlation of spatial lesion features in the spatial dimension.With the effective flow of information,the network’s representational ability will be improved.Dice loss function is introduced to solve the problem of slow convergence from a global perspective.Leaky Re LU activation function is used to solve the zero-gradient problem of the original network.It is shown by the experiment that MSANet has improved the segmentation accuracy to a certain extent,reducing the semantic gap between the encoder and decoder.Compared with U-Net,The MSANet has improved the Mean Intersection Over Union of segmentation by 2.9%,the Mean Pixel Accuracy by 2.11%,and the accuracy by1.79%.There are many problems in classification such as the complex texture features of COVID-19,lots of spatial redundancy in the Res Net50 that caused many convolution layers are not effectively used,and the perceptual information capability of the network did not match the number of layers of the model.According to those disadvantages,a Fusion Coordinate Attention Network is designed.Embedding location information into channel attention on the one hand avoids high computational overhead,on the other hand,it makes the network obtain accurate location information and long-term dependencies to enhance network’s ability to express features and recognize target regions.Si LU activation function is adopted to further reduce the zero-gradient problem of the standard Res Net50 network.The experiment has shown that FCANet improves classification accuracy to a certain extent,and can effectively utilize more information.Compared with the standard Res Net50,the FCANet has improved the average accuracy of classification by 2.66%,accuracy by 2.37%,precision by 4.59%,F1-Score by 2.44%,and false positive rate by 4.4%.In summary,to solve the difficulties and challenges faced by the current COVID-19 auxiliary diagnosis technology,this paper focuses on the key technologies in medical auxiliary diagnosis,which has certain clinical significance.