(Thursday, 8/2/2018) 7:30 AM - 9:30 AM
Room: Davidson Ballroom B
Purpose: To develop a model using MRI radiomics features to early predict local recurrence (LR) of brain metastases after gamma knife (GK) radiosurgery.
Methods: Eighty-one brain metastases undergoing GK radiosurgery from eighteen patients were analyzed in this study. Twelve of the tumors had recurred in one-year follow-up. The T1-weighted post-contrast brain MR images were acquired prior and 1~2 months after irradiation. Radiomics features of intensity statistics and imaging textures of each tumor were calculated on MR images using the treatment planning tumor volume. Radiomics features of the dose statistics for each tumor were also calculated on the 3D dose distribution. Unsupervised feature reduction based on variance and correlation was applied to the total 164 features for the initial feature selection. A support vector machines (SVM) prediction model with a radial basis function kernel was built on the remaining features to predict LR. Feature selections were implemented through modeling on every possible subset of the features with leave-one-out cross-validation (CV). Receiver operating characteristic (ROC) analysis on the CV predictions was used for feature selection and model evaluation.
Results: Initial feature reduction selected 20 radiomics features for SVM modeling. Feature selection showed a SVM model with 8 radiomics features achieving the best area under the CV ROC curve (AUC) for LR. These features included 3 from the dose distribution, 3 from the pre-treatment MRI and 2 from the post-treatment images. The eight-feature model showed predictive AUC of 0.84, with an overall prediction accuracy of 0.85.
Conclusion: The model using dose distribution, together with pre- and post-treatment MRI radiomics features provided confidence in predicting LR following brain metastases stereotactic radiosurgery.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
