A multimodal MRI-radiomics deep learning model for survival risk stratification after gamma knife radiosurgery in patients with brain metastases: A multicenter retrospective study.

Abstract

Brain metastasis (BM) is a high-mortality complication occurring in 20-40% of cancer patients. While the Gamma Knife (GK) is a primary treatment, individualized prognostic prediction remains limited. This study develops and validates a multimodal deep-learning framework for overall survival risk stratification after GK. This multicenter retrospective study includes 875 patients across three centers. A mask-guided multi-scale encoder is applied to extract MRI features. The proposed model integrated full MRI, grid-based MRI patches, radiomics, and consistently available clinical variables to generate a patient-level log-risk score for overall survival. Performance is assessed via time-dependent AUC, C-index, and Decision Curve Analysis (DCA). The model achieves 1-year AUCs of 0.870 (Training), 0.755 (Internal Val), 0.740 (External Val 1), and 0.788 (External Val 2). C-indices remain moderate across validation cohorts (0.655, 0.653, and 0.649). Multivariable Cox regression showed that the model-derived risk score was independently associated with overall survival across all cohorts. Using a training-derived exploratory threshold of 0.17, the model stratified patients into high- and low-risk groups with significant differences in overall survival across all cohorts. DCA suggests the potential net benefit at 12 months. The proposed multimodal model showed consistent but moderate discrimination for overall survival stratification in BM patients. By integrating multimodal data, the framework may provide incremental prognostic information for post-GK risk stratification. Further recalibration, incorporation of comprehensive clinical variables, and prospective validation are warranted before clinical implementation.