NIMG-08. AN INTEGRATED INFORMATICS MODEL COMBINING CLINICAL FACTORS, RADIOMICS AND A NOVEL CONNECTOMICS FRAMEWORK TO DISTINGUISH PATHOLOGICALLY-PROVEN RADIONECROSIS FROM PROGRESSION IN TREATED BRAIN METASTASES

Abstract

Abstract PURPOSE/OBJECTIVE(S) To distinguish radionecrosis (RN) from true progression (TP) in brain metastases treated with stereotactic radiosurgery (SRS), we apply machine learning to create a multi-domain model that incorporates clinical factors, multiparametric radiomics(mpRads), and tumor connectomics, a novel MRI-based complex graph theory framework that describes the intricate network of relationships within the tumor and surrounding tissue. MATERIALS/METHODS Metastases treated with SRS that had pathologic confirmation of RN vs. TP after imaging progression were included from a single institution. Regions of interest were manually segmented using the single largest diameter of the T1 post-contrast(T1C) lesion plus the corresponding area of T2 FLAIR hyperintensity. We developed an Integrated Radiomics Informatics System (IRIS) based on an isomap support vector machine (IsoSVM) model to classify TP from RN using leave-one-out cross-validation (LOOCV). Class imbalance was resolved using differential misclassification weighting during model training using IRIS. Area under the receiver operating characteristic (AUC-ROC) and AUC-PR (precision recall) analysis were performed. RESULTS We analyzed 135 lesions in 110 patients. There were 43 cases (31.9%) of RN and 92 cases (68.1%) of TP. The top-performing connectomics features were degree centrality (increased with RN) and average path length (decreased with RN), suggesting greater “connectivity” and increased similarity in intralesional features between the T1C and FLAIR signal regions in RN cases. The top-performing radiomics feature was multidimensional entropy (increased in TP), demonstrating greater heterogeneity in TP cases. Finally, the top-performing clinical features were prior RT before SRS, histology, and treated lesion size. The LOOCV IsoSVM model successfully classified TP from RN with an AUC-ROC of 0.84 (95% CI: 0.77-0.90) and AUC-PR of 0.90 (95% CI: 0.82-0.95). The F1 score was 0.89. CONCLUSION Our novel machine-learning framework was able to efficiently combine features from multiple domains (i.e., radiomics, connectomics, and clinical factors) to distinguish pathologically-proven TP from RN with excellent discrimination.