S2429
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Purpose/Objective: A key barrier to clinical radiomics adoption is the lack of rigorous external validation with pipelines. Here, we externally validate a CT-based radiomics workflow for NSCLC survival prediction and assess both survival regression and 2-year survival classification to evaluate segmentation effects and cross-dataset generalizability. Material/Methods: The internal cohort included 365 NSCLC patients who underwent PET/CT imaging at Cardiff University Hospital (2011–2017). CT radiomic features were extracted following IBSI guidelines using SPAARC- Radiomics. External validation was performed on 36 NSCLC patients from the TCIA NSCLC-Radiogenomics dataset. A LASSO-penalized Cox model was trained to predict overall survival, with radiomic features z-score normalized and the L1 penalty optimized via 5-fold cross-validation using the concordance index (C- index). Patients in the validation cohort were stratified into high- and low-risk groups based on the median risk score, and survival differences were evaluated using Kaplan–Meier and log-rank tests. Model performance was quantified using Harrell’s C- index.For comparison with classification-based modelling, 2-year survival status was predicted using Random Forest, Support Vector Machine, and XGBoost, with AUC used for discrimination. Two tumour segmentation strategies—PET-guided thresholding and manual delineation—were evaluated to assess segmentation-related variability. Results: In the internal cohort (n = 365), PET-guided segmentation (30% SUVmax) achieved the best survival modelling performance (training C-index = 0.617, validation = 0.594) and significant risk stratification (log-rank p = 0.010). Manual CT segmentation performed weaker (training C-index = 0.609, validation = 0.560; p = 0.1759). For 2-year survival classification, PET-based contours produced higher and more consistent validation AUCs (0.627– 0.645) than manual CT segmentation (AUC 0.587– 0.619).In the external TCIA cohort (n = 36), the radiomics survival model generalized without retraining, yielding a C-index of 0.655, consistent with prior literature (C-index 0.643). Kaplan–Meier stratification showed similar trends (p = 0.108). In contrast, external 2-year classification performance dropped substantially (AUC 0.356–0.367), indicating limited robustness of binary prediction compared with continuous survival modelling.
Conclusion: These results show that segmentation and endpoint design strongly influence radiomics generalizability. PET-guided segmentation outperformed manual CT contours, indicating that biologically informed delineation better captures survival-relevant tumour heterogeneity. Cox survival modelling generalized well across datasets (external C-index 0.655, similar to prior reports of 0.643), demonstrating reproducibility. In contrast, 2-year survival classification failed to generalize (external AUC 0.356–0.367), revealing high sensitivity to domain shift. Thus, while radiomics pipelines can achieve stable external survival prediction, classification endpoints are more vulnerable to dataset heterogeneity and require cautious clinical implementation. References: [1] Whybra, P., Spezi, E. Sensitivity of standardised radiomics algorithms to mask generation across different software platforms. Sci Rep 13, 14419 (2023).[2] Bakr, S., Gevaert, O., Echegaray, S., Ayers, K., Zhou, M., Shafiq, M., Zheng, H., Zhang, W., Leung, A., Kadoch, M., Shrager, J., Quon, A., Rubin, D., Plevritis, S., & Napel, S. (2017). Data for NSCLC Radiogenomics (Version 4) [Data set]. The Cancer Imaging Archive. [3] Chen W, Qiao X, Yin S, Zhang X, Xu X. Integrating Radiomics with Genomics for Non-Small Cell Lung Cancer Survival Analysis. J Oncol. 2022 Aug 27; 2022:5131170. doi: 10.1155/2022/5131170. Keywords: Radiomics, Survival, Validation PROTECT-DTI: A Multicenter Federated Imaging and Clinical Integration Program to Preserve White Matter in Stereotactic Brain Radiotherapy Paul RETIF 1,2 , Markus HECHT 3 , Guillaume VOGIN 4 , Nicolas DEMOGEOT 5 , François LALLEMAND 6 , Nicolas MEERT 7 , Xavier MICHEL 8 1 Medical Physics Unit, CHR Metz-Thionville, Metz, France. 2 CRAN, Université de Lorraine, Nancy, France. 3 Department of Radiotherapy and Radiooncology, Saarland University Medical Center Homburg/Saar, Homburg/Saar, Germany. 4 Department of Radiotherapy, Centre François Baclesse, Esch-sur- Alzette, Luxembourg. 5 Department of Radiotherapy, Institut de Cancérologie de Lorraine, Vandoeuvre-lès- Nancy, France. 6 Department of Radiation Oncology, CHU of Liège, Liège, Belgium. 7 Department of radiotherapy, Civil Hospital Marie Curie University of Charleroi, Charleroi, Belgium. 8 Radiotherapy- Digital Poster 1609
Made with FlippingBook - Share PDF online