Leveraging Quantitative Imaging and Machine Learning to Differentiate Radionecrosis from Disease Recurrence in Patients with Brain Metastases

Purpose/Objective(s)Radiation necrosis can be difficult to non-invasively discern from tumor progression after stereotactic radiosurgery (SRS). In this work, we investigate the utility of radiomics (computerized features) and machine learning to capture per-voxel lesion heterogeneity on routine MRI scans, to differentiate radionecrosis from tumor recurrence in patients with brain metastases treated with SRS.Materials/MethodsA retrospective analysis was conducted of patients with brain metastases treated with SRS. Eighty-three lesions (n = 56 intact; n = 27 surgical cavity) from 69 patients were identified with median age 68.8 years (range 40.2 – 91.0), of whom 53.6% were male and 33.3% received prior whole-brain radiotherapy (WBRT). Lesion histology included lung (60.2%), renal cell (15.7%), melanoma (10.8%), breast (9.6%), and other (3.6%). Pathologic confirmation was available in 73.5% of lesions. Both intact and resection cavity lesions were included and individually segmented. Image preprocessing and radiomic feature extraction were done using ANTsPy and open-source software. A total of 210 features were extracted from post-contrast T1-weighted (T1w) and T2/FLAIR MRIs. Highly correlated features were removed. Univariate logistic regression was conducted on the remaining T1w and T2/FLAIR features as well as on clinical variables. Multivariate analysis was implemented with various classifiers (Random Forest, Ridge, Lasso, Support Vector Machine [SVM]) on the top-performing features found on univariate logistic regression. Models were assessed using cross-validation to select the best model by area under ROC curve (AUC). Specificity and sensitivity were calculated.ResultsOn univariate analysis, the top 10 radiomics features consisted of 6 T1w features and 4 T2/FLAIR features (4 GLCM, 3 first order, 1 GLSZM, 1 GLRLM, and 1 shape feature). Age, gender, disease site, prior WBRT, prior fractionated SRS, planning tumor volume, brain-GTV V12 Gy, and immunotherapy before or after SRS were not predictive (AUC less than 62.0%) on univariate analysis compared to radiomic features. Multivariate analysis of top performing radiomic features on both intact and surgical cavities yielded an AUC of 72.0% (standard deviation [SD] ±8.8%). Multivariate analysis of top features on intact lesions alone improved the AUC to 80.5% (SD ±10.8%), with sensitivity of 77.8%, specificity of 72.4%, and positive likelihood ratio of 2.82 in differentiating radionecrosis from recurrence.ConclusionRadiomics and machine learning tools may improve diagnostic ability of distinguishing radiation necrosis from tumor recurrence after SRS. Further work is needed to deploy this in a larger multi-institutional cohort and prospectively evaluate its efficacy as a decision-support tool to personalize care in patients with brain metastases. Radiation necrosis can be difficult to non-invasively discern from tumor progression after stereotactic radiosurgery (SRS). In this work, we investigate the utility of radiomics (computerized features) and machine learning to capture per-voxel lesion heterogeneity on routine MRI scans, to differentiate radionecrosis from tumor recurrence in patients with brain metastases treated with SRS. A retrospective analysis was conducted of patients with brain metastases treated with SRS. Eighty-three lesions (n = 56 intact; n = 27 surgical cavity) from 69 patients were identified with median age 68.8 years (range 40.2 – 91.0), of whom 53.6% were male and 33.3% received prior whole-brain radiotherapy (WBRT). Lesion histology included lung (60.2%), renal cell (15.7%), melanoma (10.8%), breast (9.6%), and other (3.6%). Pathologic confirmation was available in 73.5% of lesions. Both intact and resection cavity lesions were included and individually segmented. Image preprocessing and radiomic feature extraction were done using ANTsPy and open-source software. A total of 210 features were extracted from post-contrast T1-weighted (T1w) and T2/FLAIR MRIs. Highly correlated features were removed. Univariate logistic regression was conducted on the remaining T1w and T2/FLAIR features as well as on clinical variables. Multivariate analysis was implemented with various classifiers (Random Forest, Ridge, Lasso, Support Vector Machine [SVM]) on the top-performing features found on univariate logistic regression. Models were assessed using cross-validation to select the best model by area under ROC curve (AUC). Specificity and sensitivity were calculated. On univariate analysis, the top 10 radiomics features consisted of 6 T1w features and 4 T2/FLAIR features (4 GLCM, 3 first order, 1 GLSZM, 1 GLRLM, and 1 shape feature). Age, gender, disease site, prior WBRT, prior fractionated SRS, planning tumor volume, brain-GTV V12 Gy, and immunotherapy before or after SRS were not predictive (AUC less than 62.0%) on univariate analysis compared to radiomic features. Multivariate analysis of top performing radiomic features on both intact and surgical cavities yielded an AUC of 72.0% (standard deviation [SD] ±8.8%). Multivariate analysis of top features on intact lesions alone improved the AUC to 80.5% (SD ±10.8%), with sensitivity of 77.8%, specificity of 72.4%, and positive likelihood ratio of 2.82 in differentiating radionecrosis from recurrence. Radiomics and machine learning tools may improve diagnostic ability of distinguishing radiation necrosis from tumor recurrence after SRS. Further work is needed to deploy this in a larger multi-institutional cohort and prospectively evaluate its efficacy as a decision-support tool to personalize care in patients with brain metastases.