Background: Brain metastases (BMs) incidence is high in patients with metastatic non-small cell lung cancer (NSCLC). Immune Checkpoints Inhibitors (ICIs) are now standard of care for metastatic NSCLC in the first-line or beyond settings. However, less than half of patients will have a tumor response, and the determinants of BMs response to ICIs remain unknown. This study aims to evaluate the value of radiomics to predict control/progression of BMs at a lesion level, and outcomes at a patient level in the BM+ NSCLC population treated by ICIs. Methods: We conducted a retrospective multicenter (5 European centers) study including consecutive patients with NSCLC BMs and available baseline MRI before ICIs (evaluation cohort) or chemotherapy (control cohort) with a 2:1 ratio, between 2011 and 2021. After modified RANO (Response Assessment in Neuro-Oncology) assessment of individual BMs, we developed a radiomic model to predict BMs control/progression to ICIs at the first brain follow-up imaging. The ICIs cohort was split into two datasets used for (i) model training with 5-fold cross-validation and (ii) model testing. The Chemotherapy cohort was used to ensure the specificity of our model to ICIs-treated patients. Results: Ninety-four and 49 patients were included in the ICIs and Chemotherapy cohorts respectively, of which 56 (59.6%, N = 227 BMs) and 39 (79.6%, N = 192 BMs) patients had available brain follow-up imaging. The ICIs cohort was specifically enriched in radiomic features which were significantly associated to BMs progression. Our final model, based on extreme gradient boosting (xgboost) on BMs > 10mm from the training set (N = 39 BMs), could predict individual BMs progression to ICIs with an area under curve (AUC) of 0.77 (p-value = 0.029, 95% CI [0.56-0.98]) in the test set (N = 20 BMs). We further generated a radiomic score to stratify BM+ NSCLC ICIs-treated patients between High-Risk or Low-Risk groups according to the predicted individual BMs progression. High-risk patients were associated with worse overall survival (OS) (median OS of 6.3 months, 95% CI [3.02-10.49]) compared to Low-risk patients (median OS of 11.87 months, 95% CI [7.02-21.90], p = 0.042). The prognostic value of the radiomic score on OS was validated in a multivariate analysis in the ICIs cohort (Table). Conclusions: To our knowledge, this is the first study to explore the value of radiomics in the prediction of BMs response to immunotherapy. Prospective evaluation will confirm the generalizability of our model in clinical practice.