Prediction Of EGFR Mutation Status And Prognostic Evaluation In Non-small Cell Lung Cancer Based On Radiomics

Part 1 The predictive value of conventional CT morphological features for EGFR mutation status in NSCLC patients 【Objective】To explore the value of conventional CT morphological features to predicting EGFR mutation status in non-small cell lung cancer patients and to construct predictive model of EGFR mutation combining with clinical features.【Methods】The imaging data of 243 patients that pathologically confirmed as NSCLC and underwent EGFR gene detection were analyzed retrospectively.Of 243 patients,112 were EGFR mutation types(46.1%);131 were EGFR wild types.Five clinical features(gender,age,clinical stage,CEA level and smoking status)and 17 conventional CT morphological features(maximum diameter,density,shape,lobulation,pleural indentation,etc.)between these two groups were compared.Multivariate Logistic regression analysis was performed by taking EGFR mutation status as the dependent variable,and statistically different features were taken as independent variables.The predicted probability of Logistic regression model and each individual feature were enrolled in ROC analysis.【Results】Univariate analysis showed that gender,clinical stage,maximum diameter,density,mediastinum or hilar lymph node enlargement,location,vacuole sign and CT bronchograms were significantly different between the two groups(P<0.05).Multivariate Logistic regression analysis showed that gender,density and lesion location were independent predictors of EGFR mutation.The ROC curve analysis showed that the AUC of Logistic regression model was 0.693,and the AUC of gender,density and lesion location were 0.608,0.631 and 0.579,respectively.【Conclusion】In NSCLC patients,gender,lesion density and lesion location are related to EGFR gene mutation.Female patients,subsolid density and peripheral lung cancer may suggest that EGFR mutation is positive.Part 2 The predictive value of radiomics signatures for EGFR mutation status in NSCLC patients 【Objective】To develop and validate EGFR mutation prediction model based on radiomics signatures and compare it with the morphological combined prediction model.【Methods】243 cases of non-small cell lung cancer(NSCLC)from April 2012 to February 2016 were included retrospectively as training cohort.161 cases of NSCLC from May 2015 to October 2016 were set as an independent validation cohort.The agreement of radiomics features was evaluated by using intraclass correlation coefficient.Unsupervised consensus clustering was used for radiomics signature building.All features and radiomics signatures were compared by Mann-Whitney U test and Chi-square test.Multivariable Logistic regression analysis was performed to develop EGFR mutation prediction models.ROC curve analysis was used to evaluate the predictive performances of individual features and models,and the different predictive performance between the two models were compared with Delong test.【Results】Of the 485 radiomics features,313 radiomics features of high repeatability were selected and 5 radiomics signatures were constructed by cluster analysis.The radiomics combined prediction model was developed with maximum diameter,RS1 and RS2.In the training cohort and validation cohort,the radiomics combined prediction model showed better predictive performance(AUC=0.789,0.797)compared with the morphological combined prediction model(AUC=0.693,0.616).Delong test showed that the difference between the two AUC was statistically significant(P<0.05).【Conclusion】Type Ⅰ RS1,type Ⅰ RS2 and smaller maximum diameter indicate EGFR gene mutation.The radiomics combined prediction model has a high predictive power,which can assist clinical targeted therapy.Part 3 A pilot study on the prediction of distant metastasis in patients with stage Ⅰ NSCLC by radiomics signatures 【Objective】To explore the related factors of postoperative distant metastasis in stage I non-small cell lung cancer(NSCLC),and to evaluate the feasibility of radiomics signatures in predicting the prognosis of patients with stage I NSCLC.【Methods】404 cases of NSCLC from April 2012 to October 2016 were included retrospectively.243 of them were used as training cohort to construct radiomics signatures,and the other 161 cases were used as validation cohort to test the reliability of radiomics signatures.According to inclusion criteria and exclusion criteria,194 patients with stage I NSCLC were selected and their distant metastasis were followed up.The hazard ratio(HR)of seven clinical features,16 conventional CT morphological features and 5 radiomics signatures were calculated.Univariate survival analysis was performed by log-rank test.On this basis,multivariate Cox regression was performed to select independent prognostic factors.【Results】The follow-up time ranged from 3 to 63 months(median,21 months).25 patients(12.9%)had distant metastasis and had distance metastasis free survival(DMFS)for 3~54 months(median,15 months).Univariate analysis showed that except for age,smoking,shape,lobulation sign,cusp sign,vacuolar sign,pleural thickening and RS4,the other 20 features were associated with distant metastasis in stage I NSCLC patients(P<0.05).Multivariate analysis showed histologic subtype,pleural indentation and RS1 were statistically different(P<0.05).【Conclusion】In patients with stage I NSCLC,histological subtype,pleural indentation and RS1 are independent prognostic factors affecting distant metastasis.Patients with non-adenocarcinoma,type II RS1 and pleural indentation tend to have distant metastasis after operation.