Abstract: Objective    To predict epidermal growth factor receptor (EGFR) mutation status and human epidermal growth factor receptor 2 (HER2) expression status of brain metastasis based on different radiomics interpolator and to explore the optimal interpolator for prediction.    Methods   Data from 100 patients with brain metastasis from adenocarcinoma (56 with mutant EGFR/HER2+, 44 with wild-type EGFR/HER2-) from 2 institutions were retrospectively reviewed and analyzed. Contrast-enhanced T1-weighted imaging (T1-CE) sequence was selected for radiomics features extraction by using 3 different interpolators (sitkNearestNeighbor, sitkLinear and sitkBSplines). A total of 1409 radiomics features were extracted from each MR interpolator. The patients were randomly divided into training coherent and independent testing coherent according to 7:3. The least absolute shrinkage selection operator (LASSO) was used to select informative features, a radiomics signature was built with the support vector machine (SVM) model of the training cohort, and the radiomics signature performance was evaluated by using an independent testing data set. The accuracy, sensitivity and specificity of the model was calculated by the ROC curve.  The predictive performance of the model was assessed by the ROC area under curve (AUC).   Results   Nineteen selected radiomics features based on sitkBSplines interpolator showed good discrimination in both the training and independent test cohorts. The radiomics signature yielded an AUC of 0.99, a classification accuracy of 0.95, sensitivity of 0.92, and specificity of 0.97 in the training cohort, and an AUC of 0.86, a classification accuracy of 0.8, sensitivity of 0.82, and specificity of 0.78 in the independent testing data set. SitkBSplines showed better discrimination performance than other interpolations in both Center 1 and Center 2 modeling. Nine selected radiomics signature features based on sitkLinear interpolator yielded an AUC of 0.74 in training cohort and an AUC of 0.53 in the independent testing cohort.   Conclusions   Radiomics signature model has certain application value in predicting EGFR mutation /HER2 status in adenocarcinoma brain metastases, among which radiomics features based on sitkBSplines interpolation is the most effective in discrimination performance. 

Key words: Brain metastasis; ,  , Radiomics; ,  , EGFR; ,  , HER2