S2499
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Purpose/Objective: To evaluate the predictive potential of integrating computed tomography (CT)-derived radiomic features with clinical variables to assess radiological response after Stereotactic Ablative Radiotherapy (SABR) in early-stage non-small cell lung cancer (NSCLC) patients. Material/Methods: A retrospective cohort of 137 early-stage NSCLC patients treated with SABR between 2013–2024 was analyzed. Radiological response at 3 months post- SABR according to RECIST criteria. A total of 16 clinical variables were collected and 1,688 radiomic features were extracted from planning CT images using PyRadiomics after tumor segmentation. Several clinical and clinical-radiomic predictive models were developed considering 6 different feature selection methods (LASSO, ANOVA, Random Forest, MRMR, and Optuna-based optimization) and 6 supervised classifiers (Logistic Regression, KNN, SVM, Random Forest, XGBoost, and Multi-Layer Perceptron Classifier), combined with 4 oversampling strategies (ROS, ADASYN, SMOTE, SMOTE-Tomek) and 2 hyperparameter tuning approaches (Random Search CV and Optuna). Performance was assessed using ROC-AUC, precision-recall curves, and confusion matrices on a stratified validation set (70/30 split). Results: Among 137 patients (median age: 72 years, 75% male), 65% showed a complete or partial radiological response. The best-performing ML model combined clinical and radiomic features selected using the MRMR method, with classification performed by SVM, achieving a ROC-AUC of 0.70 and a precision-recall AUC of 0.78 in the validation set. No oversampling and hyperparameter optimization further improved model performance. Texture-based radiomic features and treatment parameters were key contributors to radiological response prediction. Conclusion: The integration of radiomic and clinical data enables moderately accurate prediction of SABR radiological response in early-stage NSCLC patients. These findings support the potential of radiomics-based ML models to advance precision radiotherapy and highlight methodological considerations when analyzing small, high-dimensional, and imbalanced clinical datasets. Keywords: Lung cancer, radiomics, SBRT
Conclusion: Visualization from Vision Transformer revealed high- attention areas associated with treatment failure predominantly in the peritumoral region. References: 1. Front Oncol. 2024 Dec 12;14:1438861. doi: 10.3389/fonc.2024.14388612. https://github.com/junyuchen245/ViT-V- Net_for_3D_Image_Registration_Pytorch Keywords: Lung SBRT, BED, Vision Transformer Predicting Radiological Response to Stereotactic Ablative Radiotherapy in Early-Stage NSCLC Using Clinical and Radiomics-Based Machine Learning Models Marta Canela-Capdevila 1 , Raquel García-Pablo 1,2 , Albert Moragas-Fernández 3 , Mauricio Murcia-Mejía 2 , Rocío Benavides-Villarreal 2 , Víctor Henández- Masgrau 4,3 , Jordi Camps-Andreu 5 , Meritxell Arenas- Prat 2,3 1 Radiation Oncology, Institut d'Investigació Sanitària Pere Virgili, Reus, Spain. 2 Radiation Oncology, University Hospital Sant Joan de Reus, Reus, Spain. 3 Ciències Mèdiques Bàsiques, Universitat Rovira i Virgili, Reus, Spain. 4 Medical Physics, University Hospital Sant Joan de Reus, Reus, Spain. 5 Unitat de Recerca Biomèdica, Institut d'Investigació Sanitària Pere Virgili, Reus, Spain Digital Poster 5153
Made with FlippingBook - Share PDF online