DSAI-02 MRI BIOMARKERS FOR BRAIN METASTASES IN LUNG CANCER USING MACHINE LEARNING

Abstract

Abstract Background Brain metastases (BM) occur in approximately 30–40% of lung cancer patients, with substantial morbidity. Management involves surgery, radiation therapy, and systemic treatments, but selecting appropriate therapy is challenging because aggressive interventions benefit some patients while exposing others to toxicity with limited gain. Accurate prognostication is critical for treatment selection. Existing prognostic models rely on clinical variables with limited individualized risk assessment. Imaging biomarkers derived using machine learning (ML) represent a promising non-invasive approach that may better capture tumor heterogeneity. We investigated the prognostic utility of ML-derived MRI biomarkers in metastatic lung cancer patients. Methods Our dataset included 111 patients with lung cancer-associated BM who underwent pre-operative T1-weighted contrast-enhanced and FLAIR MRI. After image preprocessing, 214 radiomic features were extracted. Survival models were trained using clinical-only, imaging-only, and combined clinical-imaging features. Prognostic ability was assessed using concordance index (C-index) across three survival models: Cox proportional hazards (CoxPH), random survival forest (RSF), and DeepSurv. Patients were stratified into high- and low-risk groups and compared by Kaplan-Meier analysis. Pathologic correlates (PDL1 expression, tumor-infiltrating lymphocytes) were compared between imaging-derived risk groups to characterize the biologic basis of imaging features. Results MRI biomarkers were prognostic across all survival models and outperformed clinical variables alone. Combined clinical-imaging features yielded the best performance, with DeepSurv outperforming CoxPH and RSF (C-indices: 0.680 clinical-only, 0.719 imaging-only, 0.728 clinical-imaging). High-risk patients had significantly worse overall survival than low-risk patients (mean OS: 13 vs. 29 months, p < 0.005) and were more likely to be PDL1 negative (78% vs. 47%, p = 0.032) with lower tumor-infiltrating lymphocytes (56% vs. 28%, p = 0.132). Conclusion ML-derived MRI biomarkers have independent prognostic relevance in lung cancer-associated brain metastases. Incorporating these biomarkers into existing clinical risk calculators could improve prognostication and support more personalized treatment strategies.