IMG-131. Radiomics-based Risk-stratification of Brain Metastases: Preliminary Findings

Abstract PURPOSE Brain metastasis (BM) is known for poor survival with ~10% of all cancer patients developing BM at some point in the treatment course. Reliably predicting patient survival is essential for tailoring treatments. Unfortunately, existing survival nomograms, based on clinical factors (Karnofsky performance score, #lesions, volume of largest brain metastasis, and presence/absence of extracranial metastases) are suboptimal in reliably predicting BM survival. We present a radiomics-based nomogram (RadSurv)BM on FLAIR-MRI to risk-stratify BM patients into low- and high-risk survival groups based on overall-survival. (RadSurv)BM is based on the hypothesis that radiomic features comprising shape and texture-based attributes may capture intra-lesional heterogeneity across the low-and high-risk groups, providing prognostic insights for BM patients. METHODS N=200 real world BM studies treated with stereotactic radiosurgery were obtained from the publicly available Pretreat-MetsToBrain-Masks dataset and were split into training (70%, N=140) and hold-out test sets (30%, N=60). A total of 1478 radiomic features including 38 shape (e.g., volume, angularity, roundness) - capturing lesion morphology; and 1440 GLCM (gray-level co-occurrence matrix) features’ statistics (median, kurtosis, skewness, variance) that capture spatial voxel interactions and tumor heterogeneity were extracted from FLAIR-MRI. These features were fed into a Cox proportional hazards model with Elastic-Net regularization. The top features, identified within cross-validation on the training-set, were used to develop (RadSurv)BM, followed by independent validation on the hold-out set. RESULTS Significant differences were observed between two risk groups (low, high) on the training-set (p=8.25×e-12, concordance-index (CI)=0.80). On the test set, subjects were stratified into high- and low-risk groups with p=0.04 and CI=0.65, demonstrating statistically significant differences between the two risk groups. CONCLUSION Our preliminary analysis shows the promise of (RadSurv)BM in BM risk-stratification. Following larger multi-institutional validation studies, (RadSurv)BM may complement existing risk assessment towards building reliable, prognostic biomarkers for BM patients.