Prediction Of Invasive Subsolid Puimonary Nodules Based On Radiomics And Machine Learning

Objective: To select the best prediction model based on radiomic features to identify histopathological subtypes of invasive adenocarcinoma and noninvasive pulmonary nodules appearing as subsolid nodules(SSN).Method:1.Statistical models for predicting the invasiveness of SSNFrom January 2015 to September 2019,the patients in Zhongda Hospital Affiliated to Southeast University and Eastern Theater General Hospital underwent chest High-resolution computed tomography(HRCT)in which subsolid nodules were detected,were treated by surgical resection and those whose pathological results were atypical adenomatous hyperplasia(AAH),carcinoma in situ(AIS),microinvasive carcinoma(MIA),invasive adenocarcinoma(IA)were divided into non invasive group and invasive group.The cases in Zhongda Hospital Affiliated to Southeast University were randomly divided into training cohort and internal test corhort,meanwhile a certain number of cases in Eastern Theater General Hospital were randomly selected as independent external test cohort.Two quantitative parameters of nodules(diameter,average density),radiomic features,morphological characteristics,patient clinical data(gender,age),and serum tumor markers were recorded.In which the radiomic features are automatically identified and extracted by 12 sigmaσ-Discover/ Lung nodule detection system.LASSO regression was used to construct the radiomic label.The morphological model,CT model and comprehensive model were respectively constructed by binary logistic regression(LR).All models were verified in internal and external test cohort.2.Machine learning models for predicting the invasiveness of SSNFrom January 2015 to September 2019,the patients in Zhongda Hospital Affiliated to Southeast University and Eastern Theater General Hospital underwent chest HRCT and showed subsolid nodules,who were treated by surgical resection and whose pathological results were AAH,AIS,MIA,IA were divided into noninvasive group and invasive group.The cases in this hospital were selected as training cohort and were balanced using SMOTE method.The cases in Eastern Theater General Hospital were randomly selected as independent test cohort.Logistic regression(LR)algorithm,XGBoost algorithm,support vector machine(SVM)algorithm were used to build the machine learning model using correlation features of the statistical optimal model,and the grid search method was used to optimize the super parameters.The prediction efficiency of each model was compared.Results:1.215 cases were included in the training cohort(60 cases in invasive group and 155 cases in noninvasive group),69 cases were included in the internal test cohort(17 cases in invasive group and 52 cases in group noninvasive group),and 68 cases were included in the external test cohort(42cases in invasive group and 26 cases in noninvasive group).Two radiomic features,Gradient＿Shape＿Minor Axis and LBP＿Glszm＿Zone Entropy,were filtered to build the radiomic label.Comparing the prediction efficiency of all models in each data cohort,the CT model was much better than other models and selected as the optimal model,which was constructed by radiomic label,pleural depression,and quantitative parameters(diameter,average density),with the respective AUC of CT models in training cohort,internal test cohort and external test cohort was 0.954(95%CI:0.927-0.981),0.865(95%CI:0.764-0.966),0.940(95%CI:0.889-0.991).2.207 cases were included in each group after balancing training cohort,and 68 cases were included in the independent test cohort.With all machine learning models compared,LR model(Accuracy =0.853,MCC=0.684,F1 Score=0.884)and SVM model(Accuracy =0.882,MCC=0.751,F1 Score=0.876)were better than XGBoost model in test cohort(accuracy =0.768,MCC=0.670,F1Score=0.744).Conclusion: The CT model composed of radiomic label,pleural depression,average density and diameter has the better prediction efficiency on identifying invasive/noninvasive nodules appearing as subsolid nodules.The LR model and SVM model constructed by machine learning algorithm perform better than XGBoost model.