Integrating Intratumoral Heterogeneity and Peritumoral Features to Quantify Tumor Heterogeneity for Overall Survival Prediction In Non-Small Cell Lung Cancer after Radiotherapy
Abstract
Purpose
This study characterizes the spatial heterogeneity of non-small cell lung cancer (NSCLC) using intratumoral habitat radiomics and explores its predictive capability for overall survival in radiotherapy-treated patients by integrating it with peritumoral radiomic features.
Methods
Retrospective enrollment of 405 histologically confirmed NSCLC patients with pre-radiotherapy computed tomography (CT) scans from four institutions, followed by allocation into training, internal validation, and external validation sets. Based on simple linear iterative clustering (SLIC) and Gaussian mixture model (GMM), the intratumoral space was partitioned into several ecological subregions with distinct radiomic phenotypes. Radiomic features were extracted separately from each subregion, the entire tumor region, as well as the 5-mm and 10-mm peritumoral regions, and eight predictive models were constructed ultimately. The discriminative ability of these models was quantified using the concordance index (C-index) and area under the curve (AUC), while model interpretability was provided via Shapley additive explanations (SHAP) and visualized heatmaps.
Results
Random survival forest (RSF) was the optimal algorithm. The habitat model outperformed other models with a single feature set (C-index: 0.670; 1/3/5-year AUC: 0.713/0.739/0.768). Integrating habitat and 5mm peritumoral features enhanced predictive accuracy (C-index: 0.706; 1/3/5-year AUC: 0.751/0.739/0.768) and showed significant survival differences in Kaplan-Meier curves and log-rank test (p < 0.001). Key features were interpreted using SHAP analysis and heatmap visualization.
Conclusion
The integration of intratumoral microenvironmental heterogeneity features and peritumoral radiomic data presents a new strategy for constructing individualized radiotherapy prognostic models in NSCLC. By virtue of multidimensional radiomic profiling, this model supports the design of differentiated, personalized radiotherapy protocols and thus contributes to precise clinical treatment decision-making.