ESTRO 2025 - Abstract Book

S3321

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

Conclusion: Targeting separate breathing phases resulted in statistically and potentially clinically significant variation. The optimal phase depends on the (model) endpoints, so (patient-specific) weighting can be possibly used to select the optimal phase for a patient group or individual patient. In future research possible variations in targetability need to be addressed.

Keywords: FLASH-PT, Lung Cancer, Motion Management

3644

Digital Poster Enhancing online adaptive radiotherapy with uncertainty-based segmentation error detection Marissa van Lente 1,2 , Josien Pluim 2 , Robin Strand 3,4 , Samuel Fransson 3 , David Tilly 5,6 1 Medical Imaging, Radboud University Medical Center, Nijmegen, Netherlands. 2 Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands. 3 Surgical Sciences, Uppsala University, Uppsala, Sweden. 4 Information Technology, Uppsala University, Uppsala, Sweden. 5 Medical Physics, Uppsala University, Uppsala, Sweden. 6 Immunology, Genetics and Pathology, Uppsala University, Uppsala, Sweden Purpose/Objective: Anatomical segmentation is one of the biggest sources of uncertainty in all radiotherapy but also in the online adaptive radiotherapy workflow [1]. The aim of this study is to investigate the relation between the estimated uncertainty in deep learning (DL) based segmentation and the accuracy of the predicted segmentations. Material/Methods: The Monte Carlo dropout method was applied to estimate the uncertainty of a DL segmentation model of magnetic resonance (MR)-guided radiotherapy prostate cancer images, trained to segment the clinical target volume (CTV), bladder, and rectum. The training/validation set consisted of 26 patients and the test set of 10 patients, with multiple scans per patient. Concrete dropout was added to a U-Net model to obtain multiple samples from a stochastic distribution for each input image. Predictive entropy (PE) was used to capture predictive (model and data)