IMG-13. MRI-Based EGFR Mutation Prediction in NSCLC Brain Metastases Using Vision Transformers

Abstract IMPORTANCE EGFR mutation status guides treatment in non-small cell lung cancer (NSCLC) with brain metastases (BM), but invasive biopsies pose challenges in advanced patients. OBJECTIVE To develop and validate a non-invasive Vision Transformer (ViT) model using MRI to determine EGFR mutation status in NSCLC patients with BM. DESIGN Retrospective dual-center cohort study, 2018-2023, with training/internal validation (157 patients) and external validation (23 patients). SETTING Two tertiary referral centers. PARTICIPANTS 180 NSCLC patients with confirmed BM; training cohort (n=127, mean age 57.65 ± 10.14 years, 80 men/47 women), internal validation (n=30, mean age 58.87 ± 11.62 years, 13 men/17 women), external validation (n=23, mean age 58.39 ± 8.31 years, 10 men/13 women). Eligibility: histologically confirmed NSCLC, BM on MRI, EGFR status known. Consecutive sample, 100% follow-up completion. INTERVENTIONS OR EXPOSURES 3D T1-weighted contrast-enhanced (T1+C) and T2-weighted MRI sequences as model inputs. MAIN OUTCOMES AND MEASURES Primary outcome: EGFR mutation status prediction accuracy, measured by area under the curve (AUC), sensitivity/specificity, and decision curve analysis (DCA) for clinical utility. Pre-specified before data collection. RESULTS Training cohort (n=127): single-sequence ViT models, AUC 0.695 (95% CI, 0.512-0.779) (T1+C) and 0.854 (95% CI, 0.000-0.880) (T2); fused ViT+Transformer model, AUC 0.989 (95% CI, 0.973-1.000), sensitivity 0.978, specificity 0.932. Internal validation (n=30): fusion model AUC 0.953 (95% CI, 0.890-1.000), sensitivity 0.808, specificity 1.000; single-sequence AUCs 0.400 (T1+C) and 0.750 (T2). External validation (n=23): fusion model AUC 0.950 (95% CI, 0.848-1.000), sensitivity 0.833, specificity 1.000. DCA: net benefit at 10%-90% thresholds (P<.001). Conclusions and Relevance: ViT-based fusion model accurately predicts EGFR mutation status in NSCLC BM using multi-sequence MRI, outperforming single-sequence models. Offers non-invasive diagnostic tool for personalized treatment; further validation in larger cohorts warranted before clinical adoption.