<![CDATA[Background: Brain metastases are a major cause of morbidity in cancer patients and are commonly treated with stereotactic radiosurgery. However, accurate prediction of post-treatment outcomes, including local control, recurrence, and radiation-induced toxicity, remains a significant clinical challenge. Deep learning has emerged as a promising approach to improve predictive modeling by integrating high-dimensional imaging and clinical data. Purpose: To systematically evaluate the performance and clinical utility of deep learning models in predicting treatment outcomes in patients with brain metastases following SRS. Method: A systematic review was conducted following PRISMA guidelines. After duplicate removal and screening, eligible full-text articles applying deep learning for outcome prediction were included. Risk of bias was assessed using the PROBAST tool. Results: DL approaches included convolutional neural networks, transformers, ensemble models, and hybrid radiomics-DL frameworks. Reported performance was high, AUC values ranging from 0.71 to 0.99. Multimodal models integrating imaging, clinical, and genomic features achieved superior performance (AUC up to 0.945; accuracy up to 0.98) to unimodal approaches Conclusion: Deep learning models show strong potential for predicting outcomes in brain metastases after SRS, However, methodological limitations and heterogeneity remain significant challenges. Keywords: Brain Metastases; Deep Learning; Stereotactic Radiosurgery; Radiomics; Predictive Modelling.]]>
