Leveraging Serial MRI Radiomics and Machine Learning to Predict Risk of Radiation Necrosis in Patients with Brain Metastases Managed with Stereotactic Radiation and Immunotherapy

Abstract

<h3>Purpose/Objective(s)</h3> Radiation necrosis (RN) is the primary dose-limiting toxicity of stereotactic radiotherapy (SRT) for brain metastases (BMs), especially in patients also receiving immunotherapy (IO). We investigated the utility of serial magnetic resonance imaging (MRI) radiomic features to develop predictive quantitative imaging biomarkers of brain metastases response to SRT and IO. <h3>Materials/Methods</h3> Using an IRB-approved single-institution database, we retrospectively identified adult patients with brain metastases managed with SRT and IO between 2006 and 2021. We collected per-lesion outcomes measures of RN (radiographic and/or symptomatic), progressive disease (PD), or neither (NA). Routine pre- and post-SRT contrast-enhanced T1-weighted MRIs were pushed to MIM Software (Beachwood, OH), where individual BMs were manually contoured. A total of 1061 radiomics features were extracted per lesion using a digital gene expression panel. They were selected from the following categories: shape, intensity, neighborhood intensity difference (NID), grey-level co-occurrence matrix (GLCM), and grey-level run-length matrix (GLRLM). Feature reduction followed using Spearman's correlation coefficient (0.3 cut-off). Features with non-significantly different mean ranks pre- and post-SRT (p<0.01) per Wilcoxon signed-rank test were also excluded. Simple then multiple nominal logistic regressions (SLR and MLR) with Bonferroni correction were used to model the risk of RN, PD, or NA based on pre- and post-SRT, and delta-radiomic ([post-pre]/pre) features. 95% CI of ROC AUC was obtained by 10,000 bootstrapping. Statistical analysis was performed using software. <h3>Results</h3> Our dataset encompassed 301 BMs from 92 patients with NSCLC, melanoma, and renal cell carcinoma, all treated with brain SRT and IO. 74 (24.6%), 75 (24.9%), and 152 developed RN, PD, or NA, respectively. The final radiomic feature set included 39 features that significantly changed post-SRT (p<0.01). MLR resulted in a 3-radiomic feature model that predicted risk of RN (ROC AUC 0.71). The model included post-SRT BM surface area and texture (GLCM 3D Homogeneity) and pre-RT BM roundness. Interestingly, another 5-radiomic feature MLR model predicted RN and PD (ROC AUC 0.71) based on pre-SRT GLCM 3D texture feature, 3 post-SRT shape features (surface area, mean breadth, and roundness), in addition to delta-volume. <h3>Conclusion</h3> Our analysis showed that static and serial MR radiomic features could potentially predict BMs response to SRT and IO. Model external validation, combination with clinical variables, and integration of other MRI sequences are warranted before the large-scale adoption of this novel diagnostic approach.