2104 - Survival Prediction after Stereotactic Radiosurgery in Patients with Brain Metastases Using a Multimodal 3D MRI-Radiomics Deep Learning Model
Presenter(s)
Y. Chen1, Q. Yuan2, C. K. Cramer3, C. A. Helis3, G. He1, A. R. Choi3, P. J. Young3, Y. Wang4, F. Xing5, J. Ruiz6, J. Tan1, Q. Lyu7, C. T. Whitlow8, M. T. Munley3, J. Willey3, M. D. Chan1, and Y. Jiang1; 1Wake Forest University School of Medicine, Winston Salem, NC, 2Southern Medical University, Guangzhou, Guangdong, China, 3Department of Radiation Oncology, Wake Forest University School of Medicine, Winston-Salem, NC, 4Department of Molecular and Cellular Bioscience, Wake Forest University School of Medicine, Winston-Salem, NC, 5Department of Cancer Biology, Wake Forest University School of Medicine, Winston-Salem, NC, 6Department of Cancer Medicine, Wake Forest University School of Medicine, Winston-Salem, NC, 7Yale School of Medicine, New Haven, CT, 8Department of Radiology and Biomedical Imaging, Yale School of Medicine, New Haven, CT
Purpose/Objective(s): Brain metastases (BM) occur in 20–40% of cancer patients and are associated with high mortality. Stereotactic radiosurgery is a standard treatment; however, outcomes vary substantially even among patients with good performance status, and existing clinical prognostic systems lack sufficient precision for individualized survival prediction. This study aimed to develop and validate a multimodal deep learning framework integrating 3D MRI, radiomics, and clinical data to improve overall survival (OS) prediction in patients with BM.
Materials/Methods: This multicenter retrospective study included 1,079 patients from three medical centers. We developed a Global-to-Local Multiple Instance Learning Mixture of Experts (GL-MIL MoE) model that combines a mask-guided multi-scale encoder to process global MRI volumes with Multiple Instance Learning (MIL) to extract features from high-resolution local image patches. Model performance was evaluated using time-dependent area under the curve (AUC), concordance index (C-index), and decision curve analysis (DCA). Model interpretability was assessed using SHAP (Shapley Additive Explanations) and Grad-CAM analyses to identify key prognostic features.
Results: The GL-MIL MoE model demonstrated strong discriminative performance, achieving 1-year AUCs of 0.870 in the training cohort, 0.755 in the internal validation cohort, and 0.740 and 0.788 in two independent external validation cohorts. The model maintained robust C-indices across all validation cohorts, significantly outperforming established baseline prognostic models (p < 0.05). Multivariable Cox regression confirmed the model-derived risk score as an independent predictor of survival (p < 0.05). A validated log-risk threshold consistently stratified patients into high- and low-risk groups across all cohorts. Decision curve analysis demonstrated a positive net clinical benefit at 12 months. SHAP analysis identified intratumoral heterogeneity and primary tumor type as the dominant contributors to model predictions.
Conclusion: This multimodal global-to-local deep learning framework provides a precise, validated, and interpretable approach for individualized survival prediction in patients with brain metastases treated with stereotactic radiosurgery. Integration of 3D MRI, radiomics, and clinical data offers superior prognostic performance and clinical utility, with potential to support personalized treatment planning and post-treatment surveillance.