DriverNet: A Clinical and MRI-Based Framework for Noninvasive Pre-Treatment Molecular Triage in NSCLC Brain Metastases

Background/Objectives: Brain metastases (BMs) are a major cause of morbidity and mortality in non-small-cell lung cancer (NSCLC). In this setting, EGFR-mutant and ALK-rearranged tumors represent clinically actionable, CNS-relevant oncogenic subgroups for which matched TKIs are essential to management, yet lesion-level molecular profiling is not always feasible or immediately available. We aimed to develop and externally validate DriverNet, a clinical and MRI-based framework for noninvasive pre-treatment molecular triage of EGFR/ALK status in NSCLC-BM. Methods: In this multicenter study, we analyzed pretreatment clinical, T1CE and T2-FLAIR MRI data to develop unimodal radiomics, 2D/2.5D deep learning (DL), and multimodal fusion models. The final model used ImageNet-pretrained CNNs for feature extraction and a Transformer-based architecture for fusion. The primary cohort was split strictly at the patient level before slice extraction and model development, and two independent external cohorts were used for testing. Clinical-only, imaging-only, and clinical-imaging models were compared using discrimination, calibration, Brier score, and decision-curve analyses. Model interpretability and exploratory prognostic stratification were also assessed. Results: A total of 374 patients from three centers were included. Center 1 comprised 224 patients (Chinese) and was divided into training (n = 179, EGFR/ALK+ 55.3%) and internal validation (n = 45, EGFR/ALK+ 57.8%) sets. External cohorts included 54 patients from Center 2 (Chinese, test 1, EGFR/ALK+ 42.6%) and 96 from Center 3 (Western, test 2, EGFR/ALK+ 12.5%). Among all evaluated models, DriverNet achieved the best overall performance, with AUCs of 0.967, 0.947, 0.962, and 0.952 in the training, internal validation, and two external cohorts, respectively, outperforming the clinical-only and imaging-only models. Model-derived labels were also associated with overall survival in exploratory analyses. Conclusions: DriverNet is a clinical and MRI-based framework for noninvasive pre-treatment molecular triage in NSCLC-BM. It may provide complementary information for future molecular triage studies when lesion-level profiling is unavailable or delayed. Prospective validation in larger and more molecularly balanced cohorts remains necessary before any clinical implementation can be considered.