ESTRO 2025 - Abstract Book

S2459

Physics - Autosegmentation

ESTRO 2025

2 Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiooncology – OncoRay, Dresden, Germany. 3 Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany. 4 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. 5 Department of Particle Therapy, University Hospital Essen, West German Proton Therapy Centre Essen (WPE), West German Cancer Center (WTZ), Germany, German Cancer Consortium (DKTK), Partner Site Essen, Germany. 6 German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany Purpose/Objective: For the segmentation of organs-at-risk (OAR) in the brain, magnetic resonance imaging (MRI) is used, as it provides superior soft tissue contrast compared to computed tomography (CT). However, MRI can be highly diverse in the scanning parameters influencing the image contrast. This may complicate the development of broadly applicable automatic segmentation models when only a limited amount of training data is available. Here, we developed automatic segmentation models based on T1-weighted MRI, compared their performance to CT-based approaches and analysed their robustness. Material/Methods: We trained 3D U-Nets using the nnU-Net framework [1] to automatically segment the cerebellum, the temporal and frontal lobes and the hippocampi using (i) CT, (ii) MRI (T1 or contrast-enhanced T1 (T1ce)) and (iii) combined CT+MRI as input. We included data from patients with brain tumours treated with protons at two different centres [N 1 =109 (N 1,training =76, N 1,validation =33), N 2,validation =53]. The planning CT (pCT), MRI (T1 or T1ce;Centre1:different scanning sequences,Centre2:identical scanning sequences), and the planned dose distribution were available. The OARs were manually delineated based on the pCT and MRI, serving as reference ground truth (GT). We evaluated quantitative metrics including the Dice similarity coefficient (DSC) and dose-volume-histogram (DVH) parameters. The results of the different approaches were compared using Wilcoxon signed-rank tests. Results: Our models showed good performance across all approaches with median DSC values > 0.85 for all OARs except for the hippocampi, see Figure 1. The best results were obtained for the cerebellum with the CT approach in the internal validation (median(DSC 1,CT )=0.94) and the MRI approach in the external validation (median(DSC 2,MRI )=0.93). For the moderately performing hippocampi, the MRI approach led to improved results in the internal cohort (not significant) and the external cohort (p CT-MRI <0.001). In general, segmentation performance was reduced in external validation, and median DSC values were slightly higher for the MRI models. Differences in the DVH parameters, as exemplarily illustrated in Figure 2 for D2% for the cerebellum and the left hippocampus, agree with the quantitative findings.