Explainable 1-mm Peritumoral CT Radiomics for EGFR Mutation Prediction in Non-small Cell Lung Cancer: A Vietnamese...

IntroductionEGFR genotyping guides therapy selection in non-small cell lung cancer (NSCLC) but may be delayed by limited tissue and testing logistics. We developed and internally validated an explainable CT radiomics model using peritumoral features to predict EGFR mutation in Vietnamese patients with NSCLC.MethodsWe retrospectively analyzed 200 treatment-naïve NSCLC patients who underwent thin-section CT and EGFR testing. Tumor and 1/3/5 mm peritumoral rings were segmented, yielding 1,316 standardized radiomic features per region. After comparing eight selection methods, mRMR was chosen to build logistic regression, random forest, and XGBoost models using nested 5×3-fold cross-validation. Model discrimination, calibration, and decision-curve performance were assessed. SHAP analysis clarified the contribution of key features, and a clinical-radiomics nomogram combined the optimal ring model with age, sex, smoking, and stage. In a 64-patient NGS subset, radiomic signatures were explored in relation to major signaling pathways.ResultsEGFR mutations were detected in 44% of cases. The 1-mm peritumoral XGBoost model achieved the best performance with an AUC of 0.87, outperforming intratumoral and wider-ring models, p<0.05. SHAP highlighted LoG-filtered texture features capturing heterogeneity at the tumor-stroma interface. The clinical-radiomics nomogram showed strong discrimination (AUC=0.86) and stable calibration. Radiogenomic analysis revealed that peritumoral heterogeneity features were more pronounced in tumors carrying TP53 mutations. In addition, EGFR and KRAS alterations did not co-occur in the NGS subset.ConclusionA compact 1-mm peritumoral radiomics model provides explainable, biologically grounded prediction of EGFR mutation and provides a feasible adjunct for precision oncology in resource-limited settings.