Research On The Detection Method Of COVID-19 Chest X-Ray Based On CNN

The control of Corona Virus Disease 2019(COVID-19)is urgent because it interferes with normal human life and threatens human health.COVID-19 is highly transmissible,and rapid isolation,confirmation and treatment are important tasks to control the epidemic,and Chest X-Ray(CXR)images as a method for detecting COVID-19.With the continuous development of deep learning,CNN is used in the medical imaging field,and image classification method can identify COVID-19 CXR images and achieve automatic computer-aided detection of COVID-19,which can reduce the costs of time and many resources of society.The research in this thesis is as follows:(1)To address the problem of improving the classification accuracy of image classification networks and realizing automatic assisted detection of COVID-19,a COVID-19 CXR image detection method based on high-level semantic weighted fusion is proposed to construct a feature pyramid fusion network in image classification.The feature maps extracted from the shallow network and the deep network are combined by elemental cascading.and then the high-level semantic information in the deep network is processed and weighted into the branches of the fusion network to reinforce the semantic features and reuse the feature maps in both deep and shallow network,and the Trapezoid Pyramid Fusion-VGG(TP-VGG)moving one layer of convolution of the deep network in VGG into the trapezoid pyramid fusion branches to aggregate the weighted information of the higher-level features,and replacing the fully connected layer with global average pooling in the TP-VGG.It is able to reduce the parameters.Compared to VGG,the TP-VGG has a smaller number of parameters and significantly improved classification accuracy.The final application is to the COVID-19(two classes、three classes)dataset.(2)In order to solve the problem of recognizing COVID-19 CXR images using image classification networks.The Attention Steered Trapezoid Pyramid Fusion Network(ASTPNet),which can be attached to different CNNs.Firstly,by parallel spatial attention and channel attention,we emphasize the effective semantic information in space and channel,weaken the influence of interference information in space and channel,and construct the ASTPNet structure,and secondly,we propose the Attention Steered Block(AS Block)to aggregate the weighted information efficiently.It was shown experimentally that the performance of ASTPNet was improved by attaching ASTPNet to VGG16/19,Res Net34/50 and Res Ne Xt,and applied to recognize COVID-19 CXR images,the binary and triple classification accuracies reached 98.40% and 97.10%,respectively,ASTPNet has a better recognition results.(3)The parameters and Floating Point Operations(FLOPs)were used as one of the evaluation method for the problem of optimizing the model structure for the recognition of COVID-19 CXR images.Need to meet the mobile and embedded devices.A lightweight network-based method for COVID-19 CXR image detection is proposed.The Lightweight Attention Steered-Shuffle Net(LWAS-Shuffle Net)model is constructed by guiding the fusion branches with Lightweight Attention(LWA).It can perform feature extraction efficiently.The weights on the source datasets are obtained during the training process using migration learning,and finally LWAS-Shuffle Net0.5and LWAS-Shuffle1.0 are obtained.Experiments have shown the effectiveness of the LWA module and the better performance of the LWAS-Shuffle Net 0.5,the recognition accuracy reaches 98.40% and 97.46% on binary and triple classification datasets,and the parameters and FLOPs amount reaches 1.6282 M and 65.1892 M.The parameters and FLOPs are lower while ensuring accuracy,optimizing model structure.Figure[39]Table[16]Reference[81]...