Predicting The Molecular Classification Of Lung Adenocarcinoma EGFR Based On CT Radiomics

Objective: Patients with lung cancer who have epidermal growth factor receptor(EGFR)mutations can benefit from targeted therapy.However,determining the EGFR mutation status non-invasively prior to treatment initiation is still challenging.In this study,a comprehensive model based on CT radiomics features was constructed to predict the mutation status of EGFR and its molecular subtypes.Materials and methods: 1.We enrolled 420 patients with histopathology confirmed lung adenocarcinoma in the development(n=294)and validation(n=126)cohorts.The Wilcoxon test and least absolute shrinkage and selection operator(LASSO)were used for feature selection.Decision tree(DT),logistic regression(LR),and support vector machine(SVM)classifiers were used for radiomics model building.Used the clinical and radiological features establish clinical-radiology(C-R)model.The C-R model with the best radiomics model to establish clinical–radiological–radiomics(C-R-R)model.The predictive performance of the model was evaluated by ROC and calibration curve analyses,and the clinical usefulness was assessed by a decision curve analysis.2.A total of 512 patients with lung adenocarcinoma confirmed by pathology in our hospital were collected retrospectively.All patients were detected for EGFR mutation status.According to the 7:3 stratified sampling method,these patients were divided into training set(n=359)and test set(n=153).Different methods(including Wilcoxon test,Lasso regression,LR and correlation analysis)were used for feature screening of EGFR molecular subtypes.LR was used to establish models based on clinical factors and radiomics features to predict the mutation status of EGFR molecular subtypes.Combine the clinical factors and radiomics features to build a comprehensive model.ROC curve analysis was used to evaluate the predictive performance of each model.Results: One clinical factor(smoking history,OR=0.373,95% CI,0.182-0.765,P=0.007)and three radiological features(bubble-like lucency,OR=3.669,95% CI,1.975-6.816,P<0.007;pleural attachment,OR = 0.296,95% CI,0.148-0.594,P=0.001;pleural retractoin,OR=2.207,95% CI,1.188-4.100,P=0.012)were independently associated with EGFR mutation by multiple logistic regression analysis.Twelve radiomics features were significantly correlated with EGFR mutation,and the best radiomics model was obtained using the SVM classifier.Of the three models(clinical model,C-R model,and C-R-R model),the C-R-R model had the best discriminative ability in predicting EGFR mutation status and was stable in the validation cohort,with AUC of 0.849(95% CI,0.805-0.893)and 0.835(95% CI,0.761-0.909)in the development and validation cohorts,respectively.2.Multivariate logistic regression analysis showed that there were two clinical factors(age,OR=1.040,95%CI,1.008-1.072,P=0.012;smoking history,OR=4.636,95% CI,2.620-8.205,P<0.001),1 clinical factor(smoking history,OR=3.657,95% CI,2.095-6.384,P<0.001),and 1 clinical factor(age,OR=0.951,95% CI,0.922-0.980,P=0.001)that were significantly related to molecular subtypes,respectively.The EGFR subtype group 1(Del 19 vs.Wild),2(L858R vs.Wild),and 3(Del 19 vs.L858R)had 10,8,and 1 radiomics features that were significantly correlated with EGFR.The comprehensive model combining the clinical factors and the radiomics features had the highest diagnostic performance among the molecular subtype groups.The AUC in the training set were 0.838(95% CI,0.788-0.888),0.766(95% CI,0.706-0.827),and0.668(95% CI,0.598-0.738);and the AUC in the test set were 0.818(95% CI,0.740-0.896),0.751(95%CI,0.655-0.847),0.643(95% CI,0.533-0.753),respectively.Conclusions: 1.C-R-R model has the best discriminative ability in predicting EGFR mutation status preoperatively,and its performance is stable in the validation cohort.This model can provide some imaging basis for non-invasive preoperatively in patients with candidate targeted therapy.2.The comprehensive model based on CT radiomics features has the highest diagnostic performance in the prediction of EGFR molecular subtypes,and also has a good performance in the test set.This comprehensive model has potential clinical value in the formulation of personalized targeted clinical therapy.