Tumor Connectomics: A Novel MRI-Based Framework to Distinguish Pathologically-Proven Tumor Progression and Radiation Necrosis in Brain Metastases Treated with Radiosurgery

Abstract

<h3>Purpose/Objective(s)</h3> Distinguishing true progression (TP) from radionecrosis (RN) in brain metastases after stereotactic radiosurgery (SRS) remains difficult. While traditional radiomics focuses on characterizing individual voxels, we applied tumor connectomics, a novel MRI-based complex graph framework that describes the intricate network of relationships within the tumor and surrounding tissue, to identifying TP vs. RN after SRS. <h3>Materials/Methods</h3> Brain metastases treated with SRS that underwent biopsy or resection for pathologic assessment of TP vs. RN were included from a single institution. The regions of interest were manually segmented based on the single largest diameter of the T1 post-contrast (T1C) enhancing lesion plus the corresponding area of T2 FLAIR hyperintensity, on the MRI images preceding surgical assessment typically within 5 days. The data was randomly allocated into training (80%) and test (20%) sets. The training data was standardized to zero mean and unit standard deviation. The standardization parameters were then applied to the test set to prevent any data leakage. We trained a linear support vector machine classifier using the connectomics graph theoretic metrics (consisting of node strength, degree centrality, betweenness centrality, average path length, and eigenvector centrality) in combination with biologically effective dose (BED10) to classify TP from RN. Class imbalance was resolved using the Synthetic Minority Over-sampling Technique (SMOTE). Area Under the Curve (AUC) analysis was performed. Statistical significance was set as p<0.05. <h3>Results</h3> We analyzed 135 SRS-treated lesions in 110 patients. The most common histologies were NSCLC (35.6%), breast (20.7%), and melanoma (20.0%). The median BED10 for SRS was 43.2 Gy (range: 28-60). There were 43 cases (31.9%) of pathologically proven RN and 92 cases (68.1%) of TP after SRS. The highest-performing connectomics features were degree centrality and average path length. Degree centrality was increased while average path length was decreased in RN cases compared to TP cases. This suggests a higher degree of connectivity in RN, with greater similarity in intralesional features between regions represented by the T1C and FLAIR signal. The model built using these two connectomics metrics had a sensitivity, specificity, and AUC-ROC of 0.85, 0.57, and 0.81 (95% CI: 0.61-0.93), respectively. Incorporating BED10 into the model improved performance to a sensitivity of 0.8 and specificity of 0.85, with AUC-ROC of 0.89 (95% CI: 0.7-0.98). <h3>Conclusion</h3> Our novel tumor connectomics machine-learning framework was able to distinguish pathologically-proven TP from RN with excellent discrimination, which may ultimately serve as a useful advanced imaging metric to guide clinical decision-making for radiographically progressive brain metastases.