Risk Prediction Of Brain Metastasis In EGFR-Mutated Non-Small Cell Lung Cancer Patients Based On Thoracic CT Radiomics

PurposeNon-small cell lung cancer(NSCLC)has a high mutation rate of epidermal growth factor receptor(EGFR)gene.Brain metastasis(BM)is a common treatment failure pattern in EGFR-mutated NSCLC patients,but there was few study predicting the risk of BM in these patients.The purpose of this study was to develop a comprehensive model to predict brain-metastasis free survival(BM-FS)based on clinical factors and primary CT radiomics method in stage Ⅲ/Ⅳ EGFR-mutated NSCLC patients.MethodsA total of 318 patients admitted to Shandong Cancer Hospital with first diagnosed EGFR mutation,stage Ⅲ/Ⅳ NSCLC were collected from 5 August 2014 to 24 November 2020.According to the enrollment and exclusion criteria,142 patients were selected to construct a follow-up cohort for a retrospective cohort study.Clinical information was collected and follow-up,and patients’ BM-FS was calculated.The pretreatment thoracic CT images were collected and the radiomics signatures were extracted using the 3D-slicer software.The LASSO regression and 5-fold cross-validation were used for radiomics data to identify optimal features and the Radscores were calculated according to the feature-weighted regression coefficients.Patients were classified into high risk or low risk group according to the Radscores cut-off value.The Kaplan-Meier survival curve and receiver operating characteristic(ROC)curve were drawn and the area under the ROC curve(AUC)was calculated to evaluate the efficacy of radiomics model.The clinical model was established by the univariate Cox regression model,then including clinical features and Radscores into the multivariate Cox regression model to construct a comprehensive model.Comparing the C-index of clinical,radiomics and comprehensive models to evaluate the predictive efficacy.The predictive performance of comprehensive model were evaluated based on model discrimination ability,calibration performance,and clinical utility.ResultsA total of 851 radiomics features were extracted and 6 optimal features were selected to constitute Radscores.Patients were classified into high risk group and low risk group according to the Radscores cut-off value and there was significant difference in BM-FS between the two groups in total cohort(P<0.0001).The 1-year,2-year,3-year BM-FS AUC of the radiomics model in training group were 0.774,0.809,0.704,respectively and the 1-year,2-year,3-year BM-FS AUC in validation group were 0.815,0.896,0.763,respectively.First treatment response was selected as the clinical predictive factor for brain metastasis(P<0.001).The 1-year,2-year,3-year BM-FS AUC of the comprehensive model in training group were 0.858,0.834,0.722,respectively and the 1-year,2-year,3-year BM-FS AUC in validation group were 0.906,0.866,0.831,respectively,showing a well predictive performance.The C-index of the comprehensive model in training and validation groups were 0.865(95%CI:0.810-0.921)and 0.824(95%CI:0.716-0.932),respectively.The comprehensive model which combining clinical factor and Radscores showed better predictive performance compared to the clinical model(P<0.001 in the training group and in the validation group).Compared with the radiomics model,the predictive ability of the comprehensive model was not significantly improved(P=0.083 in the training group,P=0.647 in the validation group).The calibration curve and decision curve analysis of the comprehensive model indicated a good model calibration and clinical utility.ConclusionsThe comprehensive model combining clinical factor and Radscores showed good results in predicting BM-FS in stage Ⅲ/Ⅳ EGFR-mutated NSCLC patients,which contributes to the stratification and prediction of the risk of brain metastasis and provides a reference for the development of clinical treatment plan.