Survival prediction in non-small cell lung cancer using layer-wise radiomics and stacked radscore integration: a multi-institutional study.

[OBJECTIVES] To enhance prognostic modeling in patients with non-small cell lung cancer (NSCLC), we developed and externally validated a novel radiomics framework integrating region-specific feature selection and stacked modeling of radiomic signatures derived from anatomical and dosimetric subregions.

[METHODS] This retrospective, multi-institutional study included 137 patients with advanced-stage NSCLC treated with either three-dimensional conformal radiotherapy (3D-CRT) or volumetric modulated arc therapy (VMAT). From pre-treatment computed tomography (CT) images, 837 radiomic features were extracted per region from segmented volumes including the gross tumor volume (GTV), peritumoral tissue, lung parenchyma, and dose-defined regions. Region-wise feature selection was performed using least absolute shrinkage and selection operator (LASSO)-Cox regression to generate four regional radiomic scores (Radscores). A final Stacked Radiomics model was constructed by combining these scores using survival-guided Cox regression coefficients. Prognostic performance was evaluated using the concordance index (C-index), Kaplan-Meier analysis, and log-rank testing.

[RESULTS] In the training cohort, the Stacked model demonstrated superior prognostic accuracy (C-index = 0.86), compared to the Layer-wise (0.83) and Conventional models (0.79). External validation confirmed the robustness of the Stacked model (C-index = 0.87), outperforming the Layer-wise (0.74) and Conventional models (0.73). Region-specific analysis revealed that GTV and peripheral regions contributed most to survival prediction, while lung parenchyma features had limited generalizability.

[CONCLUSIONS] Our survival-driven, region-aware radiomics framework significantly improves outcome prediction in advanced-stage NSCLC, offering a promising approach for personalized risk stratification and treatment planning.