Abstract
OBJECTIVES: Epidermal growth factor receptor (EGFR) mutation status is an essential biomarker guiding targeted therapy selection in lung adenocarcinoma. This study aimed to develop and validate a non-invasive predictive model that integrates radiomics and deep learning using CT images for accurate assessment of EGFR mutation status. METHODS: A total of 220 patients with lung adenocarcinoma were retrospectively enrolled and randomly divided into training and testing cohorts at a 7:3 ratio. Radiomics features were extracted from CT images using PyRadiomics, and deep learning features were obtained from five pretrained architectures: ResNet34, ResNet152, DenseNet121, ShuffleNet, and Vision Transformer (ViT). Feature selection used the intraclass correlation coefficient, Spearman correlation, and LASSO regression. The deep learning architectures were compared within the training set using cross-validation, and the best-performing architecture, ViT, was retained for downstream modeling. Based on the selected features, we constructed a radiomics model (Rad model), a ViT-based deep learning model (ViT model), and two fusion models (early fusion and late fusion) integrating radiomics and ViT features. Model performance was evaluated using receiver operating characteristic (ROC) curves, area under the curve (AUC), accuracy, sensitivity, specificity, precision, F1-score, and decision curve analysis (DCA). RESULTS: The fusion models outperformed both radiomics and deep learning models in predicting EGFR mutation status. In the testing set, the early fusion model achieved the highest predictive performance (AUC = 0.910), exceeding the late fusion model (AUC = 0.892), the ViT model (AUC = 0.870), and the Rad model (AUC = 0.792). It also demonstrated superior accuracy (0.848), sensitivity (0.872), and specificity (0.815). Decision curve analysis further confirmed its clinical utility. CONCLUSION: Our study demonstrated that integrating radiomics and deep learning contributed to EGFR mutation prediction, providing a non-invasive approach to support personalized treatment decisions in lung adenocarcinoma.