Integration of intratumoral/peritumoral radiomics and deep learning for predicting overall survival in non-small cell lung cancer patients: a multicenter study

Background: Prognostic assessment of non-small cell lung cancer (NSCLC) relies on TNM staging, yet tumor heterogeneity limits its accuracy. This study aimed to develop a model for improving the prediction of overall survival (OS) in NSCLC patients receiving radiotherapy, which integrated intratumoral/peritumoral radiomics features and 3D deep learning (DL) features. Methods: A total of 303 NSCLC patients from three centers were retrospectively enrolled. Radiomics features were extracted from intratumoral and 3/6/9 mm peritumoral regions on CT scans. A network named 3D-SE-ResNet was proposed to extract DL features. Lasso-Cox and principal component analysis (PCA) were used to integrate multidimensional features to establish a combined model. Performance was evaluated via the concordance index (C-index) and area under the curve (AUC). Survival differences were visualized through Kaplan–Meier curves. Results: The 6 mm expansion peritumoral radiomics features analysis showed the best performance (C-index: 0.63). The DL features outperformed the radiomics features (C-index: 0.74 vs 0.63). The combined model achieved the highest accuracy (C-index: 0.77/0.73 across datasets). K–M analysis confirmed significant survival differences (log-rank P < 0.001). Conclusion: The combined model integrates intratumoral/peritumoral radiomics features and 3D DL features and effectively predicts the OS of NSCLC patients, offering a novel tool for personalized radiotherapy strategies.