AI-assisted prediction of radiation failure patterns and survival after stereotactic radiosurgery for brain metastases: A multicenter study.

Abstract

2023 Background: Brain metastases occur in approximately 20%–40% of patients with cancer and are associated with substantial morbidity and mortality. Stereotactic radiosurgery (SRS) is an effective and widely used treatment; however, a significant proportion of patients experience local, distant, or leptomeningeal failure after treatment. Early prediction of radiation failure patterns is critical for guiding individualized surveillance and treatment strategies. Current clinical tools remain inadequate for accurately predicting radiation failure in patients with brain metastases. Methods: In this multicenter retrospective study, we included 1,079 patients with brain metastases from three medical centers. We developed a novel deep learning model, the Global-to-Local Multiple-Instance Learning Mixture-of-Experts (GL-MIL MoE) framework. The model integrates a mask-guided multiscale encoder to process global MRI volumes and a multiple-instance learning (MIL) module to extract features from high-resolution local tumor patches. The primary endpoint was radiation failure pattern, defined as local failure, distant failure, or leptomeningeal failure. The secondary endpoint was overall survival (OS). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), concordance index (C-index), and decision curve analysis (DCA). Model interpretability was assessed using SHAP (Shapley Additive Explanations) and Grad-CAM analyses. Results: The proposed AI model demonstrated strong and consistent performance, achieving AUCs ranging from 0.78 to 0.83 across validation cohorts. C-indices for OS prediction remained robust and significantly outperformed established baseline models (p < 0.05). Multivariable logistic and Cox regression analyses confirmed the model-derived risk score as an independent predictor of both radiation failure patterns and OS (p < 0.05). The model effectively stratified patients into high- and low-risk groups across all cohorts. Decision curve analysis demonstrated meaningful and consistent clinical utility. Conclusions: This AI-based deep learning model enables accurate prediction of radiation failure patterns and survival in patients with brain metastases treated with SRS. Prospective studies are warranted to evaluate its clinical utility in guiding personalized treatment and surveillance strategies in combination with standard clinicopathologic factors.