S2119
Physics - Inter-fraction motion management and daily adaptive radiotherapy
ESTRO 2026
Conclusion: During daily online adaptive MR-guided SBRT for pancreatic cancer, inter-observer variability can significantly influence OAR doses. Although dose constraints were occasionally exceeded in individual fractions, these did not consistently occur in the same patients or the same regions, and the accumulated doses across fractions remained within the initial constraints. Therefore, the current treatment approach appears safe; nonetheless, reducing inter- observer variability may further enhance safety and enable cautious dose escalation. Keywords: Pancreatic cancer, Observer Variability, MRgSBRT Proffered Paper 1057 Introduction and Dosimetric Evaluation of a Fully Automated Workflow for Target Volume Propagation in MRI-Guided Radiotherapy Using Foundation Models Denis Dudas 1,2 , Tom Julius Blöcker 1 , Chengtao Wei 1 , Stefanie Corradini 1,3 , Claus Belka 1,2 , Christopher Kurz 1 , Guillaume Landry 1,2 1 Department of Radiation Oncology, LMU Klinikum, Munich, Germany. 2 Bavarian Cancer Research Center, BZKF, Munich, Germany. 3 Department of Radiation Oncology, Universitätsklinikum Erlangen, Erlangen, Germany Purpose/Objective: One of the time-consuming steps in online adaptive radiotherapy on MR-Linacs is the daily adaptation of target volume. Our study investigates the potential of promptable foundation models, namely nnInteractive [1], for automatic contour adaptations in lung cancer patients. We explored the use of clinically approved contours from treatment planning data as prompts to guide target volume segmentation on daily MR images. We evaluated the accuracy of such a workflow and assessed its risks regarding dosimetric quality of adapted daily treatment plans. Material/Methods: Clinically approved planning contours were propagated to daily MR images using affine or deformable image registration using the TransMorph model (TM) presented in the study by Wei et al. [2]. The propagated contours were utilized to create prompts for auto-segmentation using nnInteractive. Resulting refined contours were compared with clinically approved (clinical) contours.No training data is required with nnInteractive. Three lung cancer datasets (acquired using 0.35 T MR-linac systems) were employed: a validation dataset (16 patients, 143 fractions), an internal testing dataset (21 patients, 96 fractions), and
Made with FlippingBook - Share PDF online