The Value Of CT-based Radiomics In The Diagnosis And Differential Diagnosis Of Silicosis And Tuberculosis

Part ? The value of CT-based radiomics in diagnosis of silicosis,tuberculosis and normal lung tissuesObjective By retrospectively analyzing the CT images of silicosis,tuberculosis patients and normal lung tissues in control group with consistent scanning protocol,segmentation and merge,fusion,calculation,feature selection and machine learning of the lesions and normal lung tissues that met the conditions of enrollment were used for distinguishing silicosis,tuberculosis lesions and normal lung tissues.We established a machine learning model that based on radiomics for diagnosis of silicosis and tuberculosis.Materials and methods All enrolled objects include:from January to August 2018,53 patients who met the diagnostic criteria for silicosis,89 patients who met the diagnostic criteria for pulmonary tuberculosis,and 100 patients with normal physical examination were included.According to the application requirements of AK software,139 silicosis shadows,119 shadows of tuberculosis,and 100 ROI of normal random control group in the lung field with a diameter or width greater than 5 mm were used for researching.Regularly scanned CT images were scanned with US GE Discovery CT750 HD.The tube voltage was 120kVp,the rotation time was 0.5s/r,the pitch was 1.375:1,the bed speed was 55mm/r,the tube current was automatic mA,the layer thickness and spacing were 5mm,FOV Medium Body,detector collimation was 64×0.625mm.Reconstruction using the chest algorithm after the scanning was completed.The reconstruction layer thickness was 0.625mm,the reconstruction interval was 0.625mm,the ASIR was 30%.All images were passed to the GE ADW4.6 workstation for thin layer data copying.The recombinant thin-layer images incorporating the study criteria were imported into AK software.The lesions were manually segmented and merged to generate the corresponding NRRD file and saved it.Calculation:Imported the original thin layer image data and NRRD file data of the three groups into the AK software,perform the calculation,then selected the Histogram,Formfactor,Harlick,GLCM,RLM parameters.Selected the displacement vectors as 1,4 and 7 in the relevant window,calculated 396 features for each lesion,and added the appropriate label(1 or 0).Feature Section:the preprocessing were divided into three steps:the first step was the outlier processing which replaced with the average;the second step was to set the data to Training data and test data for 7:3;the third step was to standardize the pre-processed training data and test data.Using the feature selection combination 1:(1)ANOVA+MV,Correlation Analysis,Lasso and(2)ANOVA+MV,Correlation Analysis,Lasso,PCA two sets of feature selection methods to reduce the analytical indicators number and get the required feature data.Machine Learning:Randomly divide the above feature data into training and test groups in a 7:3 ratio,and perform machine learning for SVM,BAYRat ES,KNN,Logistic Regression,RT Tree,Boosting,and ER Tree classifiers.Result 1.The silicosis and normal lung tissues which adopted the feature selection combination 1 method,the accuracy,the sensitive and the specificity of the SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were all 1.00;RT Tree,ER Tree and Logistic Regression AUC values were all 1.00.The silicosis and normal lung tissues which adopted the feature selection combination 2 method,the accuracy of the SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were:1.00,1.00,1.00,1.00,0.93,1.00,1.00;the sensitivity were 1.00;the specificity were:1.00,1.00,1.00,1.00,0.83,1.00,and 1.00;the RT Tree,ER Tree,and Logistic Regression AUC values were all 1.00.2.The tuberculosis and normal lung tissues which adopted the feature selection combination 1 method,the accuracy,the sensitive and the specificity of the SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were all 1.00;RT Tree,ER Tree,Logistic Regression AUC values were all 1.00.The tuberculosis and normal lung tissues which adopted the feature selection combination 2 method,the accuracy of the SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were:0.98?1.00?1.00?1.00?1.00?1.00?1.00;the sensitive were 1.00;the specificity were:0.97?1.00?1.00?1.00?1.00?1.00?1.00;the RT Tree,ER Tree,and Logistic Regression AUC values were all 1.00.Conclusion CT-based radiomics can accurately distinguish normal lung tissue and silicosis and tuberculosis nodules larger than 5 mm.Part ? The value of CT-based radiomics in the differential diagnosis of silicosis and tuberculosisObjective By retrospectively analyzing the CT images of patients with silicosis and tuberculosis consistent with the scanning protocol,segmentation and merge,fusion,calculation,feature selection and machine learning of 0.625mmCT thin-layer images conforming to the conditions of enrollment were used for selecting the best classifier for machine learning based on radiomics for diagnosing silicosis and tuberculosis.Materials and methods All enrolled objects include:from January to August 2018,53 patients who met the diagnostic criteria for silicosis,89 patients who met the diagnostic criteria for pulmonary tuberculosis were included.According to the application requirements of AK software,139 silicosis shadows,119 shadows of tuberculosis with a diameter or width greater than 5 mm were used for researching.Equipment,scanning methods,segmentation and merge,calculation,feature selection,and machine Learning were the same as the Part ?.Result The silicosis and tuberculosis which adopted the feature selection combination 1 method.The accuracy of SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were:0.85,0.73,0.78,0.85,0.86,0.77,0.80;the sensitive were:0.81,0.64,0.81,0.81,0,81,0.76,0.83;the specificity were:0.86,0.83,0.75,0.89,0.92,0.78,0.75;the RT Tree,ER Tree,and Logistic Regression AUC values were 0.88,0.89,0.89.The silicosis and tuberculosis which adopted the feature selection combination 2 method,the accuracy of SVM,BAYES,KNN,Logistic Regression,RT Tree,Boosting and ER Tree classifiers were:0.84?0.81?0.86?0.81?0.86?0.70?0.86;the sensitive were:0.81?0.74?0.79?0.76?0.81?0.70?0.79;the specificity were:0.86?0.89?0.94?0.86?0.92?0.70?0.94;the RT Tree,ER Tree,and Logistic Regression AUC values were 0.90?0.90?0.88.Conclusion The machine learning model of radiomics can be used as a means to differentially diagnose silicosis and tuberculosis.The best classifier for identifying silicosis and tuberculosis was the RT Tree.