Detection of MR Imaging Differences of Metastatic Intracranial Lesions in Patients Treated With or Without Immunotherapy Following Stereotactic Radiosurgery Using Radiomics and Machine Learning.

PURPOSE/OBJECTIVE(S) Immune checkpoint inhibitors (ICIs), targeting programmed death-ligand 1 (PD-L1) and programmed death protein 1 (PD-1) receptor, are now routinely used for treatment of metastatic non-small cell lung cancer (NSCLC). Brain MRI changes in response to ICI and stereotactic radiosurgery (SRS) can sometimes be challenging to interpret. Using radiomic analysis, we sought to objectively characterize the differences between NSCLC brain metastases treated with and without immunotherapy, before and after SRS. A machine learning model was applied on post-SRS treatment radiomic features to distinguish between ICI treated versus non-treated patients. MATERIALS/METHODS Patients with NSCLC brain metastases treated with SRS were retrospectively identified. Patients were stratified based on immunotherapy treatment at time of SRS and at 3 month follow up MRI imaging. Using open-source software, 120 different radiomic features were extracted from metastases from T1-weighted post contrast axial MRIs before SRS and 3 months post SRS treatment. Differences between cohorts were compared by univariate and multivariate analysis. Post SRS radiomic features were further analyzed with a machine learning model using extreme gradient boosting (XGBoost) algorithms to predict presence of post-SRS immunotherapy treatment from MR imaging features. RESULTS A total of 40 patients with NSCLC with a total of 89 brain metastases were identified. 24 patients with 57 individual lesions were treated with SRS alone without ICI. 16 patients with 32 individual lesions were treated with SRS and ICI. No significant differences were noted in the two groups prior to SRS treatment. Analysis of follow up imaging in patients on ICI after SRS versus those not on ICI revealed 30 radiomic metrics that were significantly different on univariate analysis (P < 0.05 for all). Following multivariate analysis, three radiomic metrics, flatness, sphericity, and kurtosis were significantly predictive of ICI status (P < 0.05). Machine learning analysis yielded a receiver operator curve AUC of 0.84 ± 0.12. Using this model, we observed that the same three radiomic features (flatness, sphericity, and kurtosis) were the top 3 prominent radiomic differences. CONCLUSION In this observational study, we identified a set of MRI-based radiomic metrics that accurately distinguished NSCLC brain metastases treated with or without ICI after SRS. This was validated using a machine learning model that identified the same 3 radiomic features most predictive of ICI status. Future studies will investigate whether this signature can help in classifying tumor recurrence from treatment effect and predict outcomes in this patient population. AUTHOR DISCLOSURE B.A. Morris: Employee; Epic Systems. H. Premkumar: None. T. Enright: None. P. Yadav: None. A. Burr: None. A. McMillan: None. A.M. Baschnagel: Consultant; HealthMyne, Inc.