Radiomic signatures for the non-invasive prediction of EGFR mutation status in brain metastases of lung adenocarcinoma

Abstract

Abstract The epidermal growth factor receptor (EGFR) mutation exists in approximately 50% of patients with lung adenocarcinoma and is crucial for predicting response to targeted therapies. An increasing number of patients with lung adenocarcinoma have brain metastases (BMs) at diagnosis or later develop BMs. The study aimed to establish a non-invasive radiomics model for distinguishing EGFR mutation status in BMs and investigating the predictive performance of four MR sequences. 122 patients diagnosed with BMs of lung adenocarcinoma (57 mutant EGFR patients and 65 wild-type EGFR patients) were enrolled in the study. 960 features were extracted from contrast-enhanced T1-weighted imaging (CE-T1WI), fluid-attenuated inversion recovery (FLAIR), Diffusion Weighted Imaging (DWI), and contrast-enhanced susceptibility-weighted imaging (CE-SWI) sequences separately. 27 key radiomics features were selected after feature selection. The prediction performance of different machine learning models was evaluated and the model of four MR sequences was constructed using the SVM classifier. Accuracy, sensitivity, specificity, and AUC were used to evaluate our model performance. Our CE-T1WI + FLAIR + DWI + CE-SWI sequence model achieved the best performance with ACC reaching 0.9167, AUC reaching 0.9720, Sensitivity reaching 0.9167, and Specificity reaching 0.9015. It was significantly higher than the CE-T1WI model (ACC:0.7917, AUC:0.8631), CE-T1WI + FLAIR model (ACC:0.9167, AUC:0.9231) and CE-T1WI + FLAIR + DWI model (ACC:0.8333, AUC:0.9371) in the testing set. Our CE-T1WI + FLAIR + DWI + CE-SWI model can serve as an effective tool to predict the EGFR mutation status in BMs of lung adenocarcinoma and be conducive to guiding patient treatment strategies.