A CT imaging-based deep learning model for predicting EGFR and KRAS mutations in non-small cell lung cancer: toward personalized treatment approaches.

Background: Non-small cell lung cancer (NSCLC) is a leading cause of cancer-related mortality, with mutations in key oncogenes such as epidermal growth factor receptor (EGFR) and Kirsten rat sarcoma viral oncogene homolog (KRAS) affecting treatment response. While genetic testing remains the gold standard for mutation detection, it is invasive, expensive, and limited by sampling bias. Radiomics and deep learning (DL) models based on computed tomography (CT) imaging have emerged as non-invasive alternatives to predict genetic mutations. This study aimed to develop and externally validate a multimodal CT-based DL model integrating clinical variables, for non-invasive prediction of EGFR and KRAS mutations in NSCLC, and to benchmark its performance against other DL models across multiple datasets. Methods: The study used datasets from The Cancer Imaging Archive (TCIA), including NSCLC Radiogenomics, The Cancer Genome Atlas (TCGA) Program-Lung Squamous Cell Carcinoma (TCGA-LUSC) and Lung Adenocarcinoma (TCGA-LUAD). CT scans were preprocessed, and a DL model was trained to predict mutation status based on DL features and clinical variables such as age, sex, and smoking status. Model performance was assessed using area under the curve (AUC), accuracy, sensitivity, specificity, and F1 score, and compared to other machine learning (ML) and DL models. Results: For KRAS mutations, it achieved AUCs of 0.977 [95% confidence interval (CI): 0.943-0.995] in internal validation set and 0.941 (95% CI: 0.846-0.991) in external validation set 1. For EGFR mutations, the model achieved AUCs of 0.976 (95% CI: 0.937-0.993) in internal validation set, 0.960 (95% CI: 0.892-0.990) and 0.943 (95% CI: 0.896-0.989) in external validation sets 1 and 2, respectively. Multimodal models integrating clinical variables further improved accuracy. Compared to other models, Local-Global Mutation Network (LG-MutaNet) consistently outperformed traditional ML and DL algorithms. Conclusions: The LG-MutaNet, combining CT imaging and clinical variables, offers a non-invasive and precise approach for predicting EGFR and KRAS mutations in NSCLC patients, supporting personalized treatment decisions and advancing precision medicine.