Application Of Enhanced CT-based Radiomics To Distinguish Between Renal Clear Cell Carcinoma And Renal Angiomyolipoma Without Visible Fat

Part I: Study of machine learning-based radiomics features and models in identifying benign and malignant renal tumorsObjective To apply machine learning radiomics features and models to identify benign and malignant renal tumors,and to evaluate the diagnostic efficiency of the models.Methods A total of 188 patients with renal malignant tumors and 39 patients with benign tumors were randomly divided into training group and testing group according to the ratio of 7: 3(training: 158 cases,testing: 69 cases).Take the enhanced CT(cortical phase)image data of the patient to manually outline the tumor’s region of interest,and extract 396 radiomics features,and use correlation analysis to remove features with correlation coefficients greater than 0.9,and single factor logistic regression retention P<0.05 Feature,Lasso algorithm(Least Absolute Shrinkage and Selection Operator,LASSO),and multi-factor logistic regression retained features with P<0.05.Four methods were used to perform feature dimension reduction in order,and 4 radiomics features with the largest differences were screened out,which are Min Intensity,One Voxel Volume,Small Area Emphasis,and created three machine learning models: support vector machine model,random forest model,and logistic regression model.The area under the receiver operating curve(AUC)value,accuracy,sensitivity,specificity,positive prediction,and negative prediction were used to assess the diagnostic efficacy of the model.Results There were 188 cases of renal malignant tumors and 39 cases of benign tumors,including 147 cases of renal clear cell carcinoma,20 cases of papillary cell carcinoma,21 cases of chromocytoma,30 cases of renal vascular smooth muscle lipoma,and 9 cases of eosinophil adenoma;On the training group,the logistic regression model was used to identify the benign and malignant renal tumors with AUC value of 0.89,sensitivity of 0.95,specificity of 0.59,positive prediction of 0.92,negative prediction of 0.73,AUC value of support vector machine model of 0.90,sensitivity of 0.97,and specificity 0.56,positive prediction 0.91,negative prediction 0.79,random forest model AUC value 0.95,sensitivity 1.0,specificity 0.63,positive prediction 0.93,negative prediction 1.0;logistic regression model on the testing group to identify benign and malignant renal tumors AUC value of 0.96,sensitivity of 0.97,specificity of 0.83,positive prediction of 0.97,negative prediction of 0.83,support vector machine model AUC value of 0.96,sensitivity of 0.98,specificity of 0.83,positive prediction of 0.97,negative prediction of 0.91,The random forest model had an AUC value of 0.95,a sensitivity of 1.0,a specificity of 0.42,a positive prediction of 0.89,and a negative prediction of 1.0.Conclusion The three radiomics models all show good diagnostic efficacy.The radiomics method has the value of accurate diagnosis of renal tumors.The support vector machine model has higher AUC value and accuracy in the testing group.Part Ⅱ: Application of enhanced CT-based imaging radiomics in the differential diagnosis of clear cell renal cell carcinoma and renal angiomyolipoma without visible fatObjective To distinguish between clear cell renal cell carcinoma and renal angiomyolipoma without visible fat by imaging radiomics model,and to evaluate the effectiveness of the model diagnosis.Methods 147 patients with clear cell renal cell carcinoma and 39 patients with renal angiomyolipoma without visible fat were randomly divided into training group and testing group according to the ratio of 7: 3(training: 129 cases,testing: 57 cases).Take the enhanced CT(cortical phase)image data of the patient to manually delineate the tumor’s region of interest,and extract 396 imaging radiomics features,use correlation analysis to remove features with correlation coefficient greater than 0.7,and single factor logistic regression retention P <0.05 Feature,Lasso algorithm(Least Absolute Shrinkage and Selection Operator,LASSO),and multi-factor logistic regression retained features with P<0.05.The four methods performed feature dimension reduction in order,and screened out the five most significant radiomics features,which are Quantile0.025,Surface Volume Ratio,One Voxel Volume,Spherical Disproportion,Low Intensity Large Area Emphasis,and establish an imageradiomics machine learning model: a logistic regression model.The area under the receiver operating curve(AUC)value,accuracy,sensitivity,specificity,positive prediction,and negative prediction were used to assess the diagnostic efficacy of the model.Results On the training group,the logistic regression model was used to distinguish between renal clear cell carcinoma and renal angiomyolipoma without visible fat with AUC values of 0.92,sensitivity of 0.98,specificity of 0.74,positive prediction of 0.94,and negative prediction of 0.91.Logistic regression model distinguished clear cell carcinoma of the renal from renal angiomyolipoma without visible fat with an AUC value of 0.93,a sensitivity of 0.96,a specificity of 0.75,a positive prediction of 0.94,and a negative prediction of 0.82 on the testing group.Conclusion The cortical phase-based imaging radiomics model has achieved good diagnostic efficacy in differentiating renal clear cell carcinoma from renal angiomyolipoma without visible fat.