ESTRO 2025 - Abstract Book

S3725

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Purpose/Objective: This study seeks to develop an innovative time-to-event prediction model, leveraging dosiomics features from the dynamic relationship between CT and Biological Effective Dose (BED), to predict treatment failure in Non-Small Cell Lung Cancer (NSCLC) patients following SBRT treatment. Material/Methods: We selected 179 lung cancer patients with primary NSCLC who were treated with SBRT between 2007 to 2022. We created five interaction matrices that represent the interplay between CT and BED from each patient and extracted features, using radiomics mathematics, from CT, BED, and the interaction matrices from GTV, PTV, peritumoral region outside PTV, and 50% of the prescription isodose volume outside the PTV (ISO50). We randomly selected 151 cases and 28 cases as training and test datasets, and we trained four prediction models utilizing an in-house time to-event workflow. Our workflow utilized a nested cross-validation approach to optimize a Cox Proportional Hazards model with elastic net regularization. Using a grid search approach within our nested cross-validation framework, our workflow simultaneously selects the most important features and builds the model. Each of the four models was built using the same number of selected features, derived from CT, BED, a combination of CT and BED, and a composite of CT and BED with their interaction matrices from the same training set, respectively.