2848
Diffusion Tensor and Conventional Imaging Radiomics Features to Differentiate the EGFR Mutation Status of Brain Metastases from Non-Small Cell Lung Cancer1Ewha Womans University College of Medicine, Seoul, Korea, Republic of, 2Ajou University School of Medicine, Suwon, Korea, Republic of, 3Korea University, Seoul, Korea, Republic of, 4Yonsei University College of Medicine, Seoul, Korea, Republic of, 5Gangnam Severance Hospital, Seoul, Korea, Republic of
Synopsis
We assessed whether radiomics features on diffusion tensor imaging and postcontrast T1-weighted (T1C) images differentiates the epidermal growth factor receptor (EGFR) status in brain metastases from non-small cell lung cancer (NSCLC). Radiomics features (n=5046) were extracted from 54 brain metastases patients with NSCLC (29 EGFR-wildtype, 25 EGFR-mutant). After feature selection, radiomics models were constructed by various machine learning algorithms. Diagnostic performances were compared between multiparametric and single MRI radiomics models. The best performing multiparametric radiomics model (AUC 0.97) showed better performance than any single radiomics model using ADC (AUC 0.79, p=0.007), FA (AUC 0.75, p=0.001), or T1C (AUC 0.96, p=0.678).
Purpose
Knowledge of the epidermal growth factor receptor (EGFR) mutation status in non-small cell lung cancer (NSCLC) with brain metastases is important for planning personalized treatment and predicting prognosis. 1, 2, 3 Moreover, previous studies have shown that there are discordances in the EGFR mutation status between the primary lung cancer and metastases. 4,5,6 Radiomics extracts high-throughput quantitative features within images and performs subsequent mining of the data for knowledge extraction and application. The purpose of this study was to assess whether radiomics features on diffusion tensor imaging (DTI) and conventional postcontrast T1-weighted (T1C) images can differentiate the EGFR molecular status in brain metastases from NSCLC.Methods
Fifty-four pathologically confirmed brain metastases patients with underlying NSCLC and known EGFR mutation status (29 EGFR wild type, 25 EGFR mutant) who had undergone preoperative DTI and T1C MRI were included. After preprocessing of the images, segmentation was separately performed by two neuroradiologists (S.S.A. and Y.W.P., with 14 years and 7 years of experience) to select the contrast-enhancing (CE) portion (on T1C images) and the total tumor portion (on T1C images, including the necrotic tumor volumes) using a semiautomatic method with an interactive level-set volume of interest using intensity-based algorithms. Radiomics feature extraction was performed from 1) the CE segmentation mask and 2) the total segmentation (CE + necrotic volume) mask in the apparent diffusion coefficient (ADC), fractional anisotropy (FA), and T1C images.
Radiomics features (n = 5046) were extracted from preoperative MRI including T1C and DTI of 54 pathologically confirmed brain metastases with underlying NSCLC and known EGFR mutation status (29 EGFR wild type, 25 EGFR mutant). A subset of 4317 features (85.6%) with high stability (intraclass correlation coefficient > 0.9) were selected for further analysis. After the feature selection by the least absolute shrinkage and selection operator (LASSO), the radiomics classifier were constructed by various machine learning algorithms including random forest (RF), generalized linear model (GLM), gradient boosting machine (GBM), k-nearest neighbors (KNN), and naïve Bayes (NB) classifiers with leave-one-out cross-validation (LOOCV). Diagnostic performance was compared between multiparametric MRI radiomics models and single imaging radiomics models using the area under the curve (AUC) from ROC analysis.
Results
The EGFR mutation status was also available in the primary lung cancer samples in 25 (46.3%) of the 54 patients. EGFR mutation status showed an overall discordance rate of 12% (McNemar’s test, p= .250) between primary tumors and corresponding brain metastases.
Thirty-seven significant radiomic features (6 from apparent diffusion coefficient [ADC], 6 from fractional anisotropy [FA], 25 from T1C) were selected for radiomics modeling. The AUCs of the multiparametric radiomics models ranged from 0.88 to 0.97. The best performing multiparametric radiomics was achieved by a combination of LASSO and GLM with an AUC of 0.97 (95% confidence interval [CI] 0.94-1). The accuracy, sensitivity, and specificity were 94.4%, 96.6%, and 92.0%, respectively. The best performing multiparametric radiomics model (AUC 0.97, 95% confidence interval [CI] 0.94-1) showed better performance than any single radiomics model using ADC (AUC 0.79, p =0.007), FA (AUC 0.75, p = 0.001), or T1C (AUC 0.96, p = 0.678).
Discussion
Development of noninvasive imaging biomarkers indicating the EGFR mutation status of brain metastases from lung cancer is important because they would help predict the response to TKI treatment and aid in decision making process for clinicians, towards a genetic-based and personalized medicine. Our study results show that radiomics model could predict the EGFR mutation status noninvasively with high AUC and accuracy, which may be of great importance because EGFR mutant brain metastases patients will receive benefits from TKI therapy.
In this study, the combination of DTI and conventional radiomics significantly improved the diagnostic performance of predicting EGFR mutation status compared to single radiomics models from DTI. In our study, among patients with available EGFR mutation status in both primary tumor and brain metastases, 12% showed discordant EGFR mutations in the primary lung cancer and brain metastases, which are similar to the recent study reporting a discordance rate of 13.9%. 6 The genetic heterogeneity between the primary lung cancer and metastases could partially account for the fact that some advanced NSCLC patients with wild type EGFR respond to EGFR TKI whereas some patients with EGFR mutations fail to respond to TKI therapy. 7 Noninvasive radiomics study may be useful in practical decision-making situations on TKI treatment to predict the EGFR mutation status in brain metastases.
Acknowledgements
None.References
1 Porta R, Sanchez-Torres J, Paz-Ares L, et al. Brain metastases from lung cancer responding to erlotinib: the importance of EGFR mutation. European Respiratory Journal 2011;37:624-31
2 Gow C-H, Chien C-R, Chang Y-L, et al. Radiotherapy in lung adenocarcinoma with brain metastases: effects of activating epidermal growth factor receptor mutations on clinical response. Clinical Cancer Research 2008;14:162-68
3 Mok TS, Wu Y-L, Thongprasert S, et al. Gefitinib or carboplatin–paclitaxel in pulmonary adenocarcinoma. New England Journal of Medicine 2009;361:947-57
4 Monaco SE, Nikiforova MN, Cieply K, et al. A comparison of EGFR and KRAS status in primary lung carcinoma and matched metastases. Human pathology 2010;41:94-102
5 Gow C-H, Chang Y-L, Hsu Y-C, et al. Comparison of epidermal growth factor receptor mutations between primary and corresponding metastatic tumors in tyrosine kinase inhibitor-naive non-small-cell lung cancer. Annals of Oncology 2008;20:696-702
6 Chen Z-Y, Zhong W-Z, Zhang X-C, et al. EGFR mutation heterogeneity and the mixed response to EGFR tyrosine kinase inhibitors of lung adenocarcinomas. The oncologist 2012;17:978-85
7 Han H-S, Eom D-W, Kim JH, et al. EGFR mutation status in primary lung adenocarcinomas and corresponding metastatic lesions: discordance in pleural metastases. Clinical lung cancer 2011;12:380-86
Figures
Figure 3. The heatmap of significant radiomics features from LASSO according to the EGFR mutation status. Each column corresponds to one patient, and each row corresponds to the normalized z-scores of the radiomics features.
ADC = apparent diffusion coefficient, FA = fractional anisotropy, T1C = postcontrast T1
