Reserch On COVID-19 X/CT Image Recognition Method Based On Deep Learning

Image recognition based on computer vision is one of the important research directions of computer-assisted medical diagnosis.In particular,since the outbreak of Coronavirus 2019(COVID-19),the research on medical image classification and recognition algorithm using artificial intelligence and deep learning technology has set off an upsurge of research over the world.Therefore,this paper studies the classification and recognition algorithms for COVID-19 by deep learning techniques.The dataset studied in this paper iscollected from public data with image enhancement,including more than20,000 X images and 17,000 CT images.The main research contents are as follows:(1)Based on capsule neural network,an improved capsule neural network model iCaps Net(improved-Caps Net)is proposed.Firstly,the i-Caps Net model constructs a multidimensional data input layer from the three-dimensional data deep convolution layer to enhance the extraction ability of deep and shadow features.Secondly,C-squashing nonlinear function is designed as the activation function to improve the dynamic routing in the encoder,which can compress the module length quickly and accurately,so as to classify images.Finally,the model is trained,validated and tested by using the ten-fold crossvalidation method.The comparison results among the i-Caps Net,SVM,Lenet-5,Alex Net,VGGNet-16 and Caps Net models show that the training loss value of the i-Caps Net model proposed in this paper converges to 0.1584 on the X-ray image data set,and the test accuracy reaches 93.15%.The loss value of i-Caps Net model training on CT image data set converges to 0.1285,the detection accuracy reaches 95.62%,and the classification effect is satisfied,which verifies the effectiveness and feasibility of the proposed i-Caps Net.(2)Based on the attention mechanism and depthwise convolution,a lightweight CNNbased multi-scale gated multi-head fusion Attention network(MMA-Net)is proposed.Firstly,the framwork of MMA-Net network model is MGMA,a multi-scale gated multiattention mechanism.Secondly,the depthwise convolution is used to reduce the number of calculation parameters of the model to obtain a lightweight model.Then,the space and channel are fused through a fusion module “M” to pay more attention on the lost small target information.Finally,the training,validating and testing of the ten-fold crossvalidation are carried out on the data set.The comparisons of Goog Le Net,Res Net-50,VGGNet-16,i-Caps Net and three MMA-Net models show that the test accuracy of MMANet(MMA-3)reaches 96.92% for X-ray images and 98.45% for CT images.The specificity and sensitivity of X-ray images are97.17% and 96.45%,and that of CT images are 98.26% and 98.05%,respectively.Experiments have proved that MMA-3 can detect and classify COVID-19 with higher accuracy and efficiency.(3)A hybrid RMT-Net network model is designed based on transformer mechanism and Res Net-50 network.The RMT-Net network model uses transformers to capture feature information over a long distance,the CNN and depth-wise convolution are used to obtain local features while reducing the amount of computation,and 4 stage blocks are used to enhance the ability of receptive field and migration during downsampling.Stage 1 and Stage 2 modules in RMT-Net focus on important characteristic information regions to build the relationship between tokens.Stage 3 uses the global self-attention mechanism to calculate the feature graphs of different tokens,and the Layer Norm(LN)layer and residual layer is realize the connection.Stage 4 carries out residual convolution.The global average pooling layer and full connection(FC)layer are used to perform classification tasks.RMTNet model is compared with Res Net-50,VGGNet-16,i-Caps Net and three MMA-Net models.Results for X-ray images,the Train＿acc value of RMT-Net model is 99.64%,and the specificity,sensitivity and accuracy are 98.26%,98.08% and 97.65%,respectively,which were significantly higher than the other models.For CT images,the Train＿acc value of RMT-Net model is 99.87%,and the specificity,sensitivity and test accuracy are 99.34%,98.76% and 99.12%,respectively,which are slightly better than other models.The size of RMT-Net model is only 38.5M,and the detection speed of X-ray images and CT images is5.46 ms and 4.12 ms respectively,which has improved the detection speed to a new level and verified the model classify and identify COVID-19 more quickly and accurately.In conclusion,this paper proposes three new image classification and recognition algorithms based on deep learning technology.By comparing with existing deep learning models,the classification and recognition effects in X-ray and CT images have reached ideal results,especially RMT-Net model has significant advantages in detection speed.