ESTRO 2025 - Abstract Book

S3776

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

test sets at an 8:2 ratio for validation. XGBoost (eXtreme Gradient Boosting) achieved 80% accuracy in predicting ≥grade 1 toxicity. it also demonstrated superior performance in ROC-AUC (Receiver Operating Characteristic Area Under the Curve) metrics among machine learning methods. Feature selection identified an optimal subset of 50 features, yielding 84.0% accuracy and 86.2% ROC-AUC. Model performance for predicting grade 0-3 dermatitis showed ROC-AUC values of 87.5%, 80.0%, 84.7%, and 80.3%, respectively. Key predictive features included average intensity, GLCM (Gray Level Co-occurrence Matrix) maximum probability, shape_elongation, firstorder_maximum (radiomics), firstorder_energy, shape_flatness (dosiomics) based on Shapley value. Conclusion: This study successfully developed and validated machine learning models for predicting radiation-induced dermatitis in patients with breast cancer undergoing hypofractionated radiotherapy. The XGBoost model, utilizing both radiomic and dosimetric features, demonstrated robust predictive performance. These findings provide a promising tool for personalized risk assessment and treatment planning, potentially enabling early intervention strategies to minimize dermatitis-related complications. References: 1. Gabrys HS, Buettner F, Sterzing F, et al. Design and selection of machine learning methods using radiomics and dosiomics for normal tissue complication probability modeling of xerostomia. Front Oncol 2018;8:35. 2. Zhang X, Zheng W, Huang S, et al. Xerostomia prediction in patients with nasopharyngeal carcinoma during radiotherapy using segmental dose distribution in dosiomics and radiomics models. Oral Oncol 2024;158:107000. 3. Keek SA, Beuque M, Primakov S, et al. Predicting adverse radiation effects in brain tumors after stereotactic radiotherapy with deep learning and handcrafted radiomics. Front Oncol 2022;12:920393. Proffered Paper Impact of Assumptions in Causal Inference: Estimating the Effect of Mean Dose to the Medulla Oblongata on Dysphagia After Head and Neck Radiotherapy Eliana M Vasquez Osorio 1,2 , Laia Humbert-Vidan 3 , Azadeh Abravan 1,2 , Carly E A Barbon 4 , Deborah Ganderton 5 , Katherine Hutcheson 4 , Stephen Y Lai 4 , Lip W Lee 2 , Alan McWilliam 1,2 , Amy C Moreno 3 , Gareth Price 1,2 , James Price 2,1 , Michael K Rooney 3 , Marcel van Herk 1,2 , Clifton D Fuller 3 , Matthew Sperrin 6 1 Division of Cancer Sciences, The University of Manchester, Manchester, United Kingdom. 2 Radiotherapy, The Christie NHS Foundation Trust, Manchester, United Kingdom. 3 Division of Radiation Oncology, MD Anderson Cancer Center, Houston, USA. 4 Division of Surgery, MD Anderson Cancer Center, Houston, USA. 5 Speech Therapy, The Christie NHS Foundation Trust, Manchester, United Kingdom. 6 Division of Informatics, Imaging & Data Sciences, The University of Manchester, Manchester, United Kingdom Purpose/Objective: Causal inference provides insights into cause-and-effect relationships by leveraging expert knowledge to correctly address confounding, and is essential for informed clinical decision-making [1] . However, its reliance on expert knowledge, captured as Directed Acyclic Graphs (DAGs), can introduce biases potentially leading to inaccurate conclusions. This challenge is further compounded by the possibility of multiple plausible DAGs representing the same process [1] . In this study, we quantified the impact of differing assumptions —captured in two DAGs— on estimating the causal effect of radiation dose to the medulla oblongata (part of the brainstem) and dysphagia 1-year after head and neck radiotherapy (HN-RT). Keywords: Dermatitis; Machine learning; Breast radiotherapy 2289