ESTRO 2025 - Abstract Book

S3796

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

2621

Proffered Paper Predictive modelling of contrast-enhancing brain lesion risk after proton or photon radiotherapy Martina Palkowitsch 1,2 , Larissa S Kilian 1 , Fabian Hennings 1,2 , Armin Lühr 3,4 , Justus Thiem 4 , Annekatrin Seidlitz 1,4,5 , Esther G C Troost 1,2,4 , Mechthild Krause 1,2,4 , Steffen Löck 1,4,6 1 OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany. 2 Helmholtz-Zentrum Dresden - Rossendorf, Institute of Radiooncology – OncoRay, Dresden, Germany. 3 TU Dortmund University, Department of Physics, Dortmund, Germany. 4 Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany. 5 National Center for Tumor Diseases Dresden (NCT/UCC), Germany:, German Cancer Research Center (DKFZ), Heidelberg, Germany; Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany; Helmholtz-Zentrum Dresden-Rossendorf (HZDR), Dresden, Germany. 6 German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany Purpose/Objective: Recent studies on contrast-enhancing brain lesions (CEBL) in glioma patients post-proton therapy underscore the relevance of variable relative biological effectiveness (RBE) of protons and elevated radiosensitivity in the periventricular region (PVR) [1,2]. Predictive models were developed to estimate: (a) voxel-wise CEBL risk based on absorbed dose (D), dose-averaged linear energy transfer (LET d ), and PVR information, and (b) patient-wise CEBL risk using dose-volume parameters in the PVR considering variable RBE. This study aims to: confirm LET-dependent proton RBE and increased PVR radiosensitivity by (1) independently validating the models, (2) refining the existing models, and (3) developing CEBL risk models for photon therapy. Material/Methods: Data from 156 brain tumour patients treated with proton (90) or photon therapy (66) were analysed. CEBLs were delineated on follow-up MRI scans (contrast-enhanced T1). Logistic regression models for (a) voxel- and (b) patient level CEBL prediction were independently validated on the proton cohort, refined on the proton cohort, and developed on the photon and combined cohort. Predictor variables included D, proton LET d , clinical and anatomical factors. Model performance was evaluated using the mean area under the receiver operating characteristic curve (AUC mean ), with bootstrapping applied for validation and 3x333-fold cross-validation for refinement and development. Results: 97 CEBLs were identified in 45 of 90 proton patients (50%), and 41 CEBLs in 21 of 66 photon patients (32%). (1) Validation of the (a) voxel-wise and (b) patient-wise NTCP models demonstrated good discriminative performance (AUC mean : (a) 0.83, (b) 0.81 / 0.80; Table). (2) Model refinement improved AUC mean values to (a) 0.94 and (b) 0.82. The LET-dependent proton RBE and the increased PVR radiosensitivity were confirmed: the best voxel-wise performance was achieved using dose, LET d and PVR information (Figure), while at patient level, D 2ml in the PVR considering variable RBE, and prescribed dose were strongest predictors. (3) For photons, predictive models for CEBLs were developed, achieving AUC mean values of (a) 0.94 and (b) 0.81. Increased radiosensitivity in the PVR was also observed in the photon cohort: dose and PVR location were key voxel-level predictors (Figure). In combined patient-level analysis, D 3ml in the PVR (variable proton RBE) was among the top predictors (AUC mean : 0.82), further supporting increased PVR sensitivity and RBE variability.