Classification of Tumor Progression and Radiation Necrosis Following Gamma Knife for Brain Metastases Using Radiomic Features and Clinical Variables
D Mitchell*, S Buszek, B Tran, H Liu, D Luo, S Ferguson, C Chung, UT MD Anderson Cancer Center, Houston, TX
Presentations
WE-IePD-TRACK 2-6 (Wednesday, 7/28/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: To develop a classification tool to differentiate between tumor progression and necrosis following Gamma Knife treatment for brain metastases based on both radiomic features and clinical variables.
Methods: Under IRB approval, this analysis includes a retrospective cohort of 154 patients treated with Gamma Knife for brain metastases who had contrast-enhanced T1-weighted MR images acquired using 3-D spoiled gradient echo (SPGR) sequences and surgical resection of the suspected lesion. Lesions were segmented using semi-automated methods, and tumor/necrosis was determined pathologically (100% necrosis vs. tumor present). 3-D radiomic features (3×3×3-mm spatial normalization, 64-bin intensity discretization) were extracted using PyRadiomics and clinical variables via chart review. Z-scores were used to normalize feature values. Feature reduction was performed using principal component analysis (components represented 95% explained variance). Fifteen classification models and ensembles were trained, and performance was assessed using 5-fold cross-validation.
Results: 100 radiomic features were selected a priori and extracted. Ten clinical variables were implemented, including gender, race, age at diagnosis, time from Gamma Knife treatment, histology, time between primary and metastasis diagnoses, metastasis location, Gamma Knife dose, and Gamma Knife treatment volume. A RUSBoosted trees ensemble yielded a classification accuracy of 66.2% (AUC = 0.71). 12 out of 17 necrosis cases were correctly predicted, and 90 out of 137 tumor progression cases were correctly predicted.
Conclusion: These results are promising for the development of a reliable classification tool for tumor progression and radiation necrosis using data that is readily clinically available. Such a tool could have significant value by assisting with clinical decision making. Further investigation of classifiers and parameter optimization may yield higher performance.
Funding Support, Disclosures, and Conflict of Interest: This work funded in part by MD Anderson - CCSG Radiation Oncology and Cancer Imaging Program Grant.
ePosters
Keywords
MRI, Brain, Feature Extraction
Taxonomy
Contact Email