Abstract A feared complication within oncology is the development of brain metastases (BM), as this is associated with grim prognosis and substantial morbidity. Despite the remarkable response of immune checkpoint inhibitors (ICI) for unresectable, chemotherapy-refractory tumors of varied histologies, many patients do not respond to ICI. Furthermore, median overall survival for BM patients remains poor as many responses to ICI are not durable. Reasons for this wide variability in outcomes, as well as predictive biomarkers for ICI response, remain unknown. Recently, a growing body of work has demonstrated the importance of the tumor immune microenvironment (iTME) in dictating ICI response. For example, the degree of tumor-infiltrating lymphocytes (TILs) and expression of inflammatory markers, such as interferon-gamma (IFNγ) and perforin, have been correlated with patient outcomes and durability of ICI response. However, in order to integrate these biomarkers into clinical precision medicine, non-invasive assays are needed. Brain biopsies to characterize the BM phenotype are rarely feasible due to the clear morbidity associated with surgery. Answering the need for a non-invasive method of stratifying patients by likelihood of response to ICI, we leveraged a cohort of 44 patients with brain metastases of diverse primary histologies treated with pembrolizumab. To determine if risk factors are discernable on MRI, we trained a deep learning model to predict overall survival (OS) after ICI. The model achieved 0.755 AUC-ROC in predicting OS greater than one year (HR=1.46, 95% CI: 1.11-1.92, p=0.01). In comparison with commonly used measures of progression, change in risk score from the pretreatment MRI to the first follow-up MRI was significantly associated with increasing severity of Response Assessment in Neuro-Oncology criterion for Brain Metastases (RANO-BM) (p=0.05). By developing non-invasive immunotyping for brain metastases, we anticipate our findings have the potential to result in more precise tailoring of existing therapies. To parse the degree to which model predictions were driven by tumor phenotype vs tumor immune microenvironment (iTME) we further applied the model to a unique dataset of surgically-excised melanoma metastases. The model generalized to this cohort with HR=1.37, 95% CI: 0.79-2.37, p=0.26. Feature representations learned by the convolutional neural network (CNN) were subsequently correlated with immune-relevant biomarkers, such as PD1, PDL1, and MHCII, obtained through cyclic immunofluorescence (CyCIF). Finally, features of the pretrained survival network were applied to the secondary task of ICI response prediction in the melanoma metastases cohort (accuracy=0.875). Transfer learning of the ICI survival model reveals the degree to which learned imaging patterns are mutually salient to both survival and response assessment, paving the way to identify molecular processes synonymous to both tasks. Citation Format: Mishka Gidwani, Albert Kim, Elizabeth R. Gerstner, Jayashree Kalpathy-Cramer. Convolutional neural network assessment of immune activation state of brain metastases [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-026.