| Medical diagnosticians have used machine learning to assist them in the classification and detection tasks of lung diseases.The characteristics of visual signals in lung images are complex and inconspicuous.The classification algorithm and detection algorithm for lung disease medical images of COVID-19,infectious lung disease,and non-infectious lung disease are designed and improved in this thesis,which is based on the advantages of deep learning.And fast and accurate automatic classification and detection of lung diseases has been realized,also.The main work of this thesis is as follows:1.Constructed lung medical imaging dataset LUNGS.The data set in this article contains:200 images(CT)of the new type of coronary pneumonia(COVID＿19)with a label collected from the First People’s Hospital of Lin’an District,Second Affiliated Hospital of Hangzhou Medical College,200 CTs for and infectious lung diseases(Infectious),200 CTs for noninfectious lung diseases,and 200 CTs for normal lungs(Normal).The data enhancement method is used in the detection task,so that each type of data set reaches 2000.Labeling is carried out by Label Img software,and the target labeling frame is obtained,and the resolution is 256×256.2.The N-Densenet100 network is designed on the basis of the Dense Convolutional Network(Dense Net),which combines the characteristics of lung medical imaging.The dense block(Densenet Block)structure is improved,the first layer of convolution kernel of the dense block is set to 5,and the BN layer in the dense block is removed to realize the activation reconvolution method.A layer of 3×3 convolution is added after each dense block,which can reduce the amount of parameters.The experimental results show that compared with the previous network,the classification error rate of the improved network in the four data sets of COVID-19,non-infectious lung disease,infectious lung disease,and normal lung has decreased by 0.41%,0.53%,0.55%,1.26%.The amount of parameters is reduced by nearly half.Compared with classic classification networks such as VGG and Res Net under the same data set,the error rates of the N-Densenet100 network are 9.91%,4.11%,10.09%,and 12.08%,respectively,which are the best results.3.The D-Efficient Det algorithm is designed on the basis of the Efficient Det algorithm.An NES with the number of layers set to 3 is used as the new FPN search strategy.Its singlechannel feature map node is 5.A six-step process is planned for searching.The searched NFPN replaced the artificially designed FPN.Moreover,the N-Densenet100 network designed in this thesis is used to replace the original Efficient Det backbone network.The last 5 dense blocks of N-Densenet100 are used as backbone network connections.The extracted features are input into N-FPN for feature fusion.Experimental results prove that the AP values of DEfficient Det in the detection of COVID-19 CT image,non-infectious lung disease CT image,and infectious lung disease CT image are 55.90%,56.90%,and 55.90%,respectively.Increases by 15.3%,12.3%,and 13.3% respectively,compared to the original network. |
