ESTRO 2025 - Abstract Book

S3249

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

ESTRO 2025

2213

Proffered Paper Foundation AI models for MRgRT real-time target localization Tom Julius Bloecker 1 , Elia Lombardo 1 , Sebastian N. Marschner 1 , Claus Belka 1,2,3 , Stefanie Corradini 1 , Miguel A. Palacios 4 , Marco Riboldi 5 , Christopher Kurz 1 , Guillaume Landry 1 1 Department of Radiation Oncology, LMU University Hospital, Munich, Germany. 2 German Cancer Consortium (DKTK), Partnership between DKFZ and LMU University Hospital Munich, Munich, Germany. 3 Bavarian Cancer Research Center (BZKF), Partner site Munich, Munich, Germany. 4 Dept. of Radiation Oncology, Amsterdam UMC, Vrije Universiteit Medical Centre, Amsterdam, Netherlands. 5 Department of Medical Physics, Ludwig-Maximilians Universität (LMU), Munich, Germany Purpose/Objective: This study evaluated real-time target localization in magnetic resonance imaging (MRI) guided radiation therapy based on foundation artificial intelligence models, thus bypassing the need for time and resource consuming application-specific model training. Material/Methods: The segment-anything 2 model (SAM2) for video segmentation, a foundation model not trained on medical images, was used in this work. It was applied to 0.35 T MRI-linac 2D sagittal cine MRI data from two institutions and countries, containing scans from 33 patients with 8060 labeled frames, with annotations from 2 to 5 observers per frame, totaling 29179 ground truth segmentations. Specifically, two variants of SAM2 were compared: SAM2-large (226M parameters) and SAM2-small (46M parameters). The SAM2-based approach was evaluated and compared against inter-observer (IO) variability, static registrations from breath-hold (ST-BH) and non-breath-hold (ST-NBH) frames, and a transformer-based image registration model, TransMorph (31M parameters), with patient-specific (PS) fine-tuning (TM-PS) and without (TM) (Lombardo et al.). The segmentations produced were compared with the ground truth using the Dice similarity coefficient (DSC), the Euclidean center-of-mass distance (ECD), and Hausdorff distances, with only the first two reported in this abstract due to space. Results: Results (see Figure 1) showed that SAM2 in its large (median DSC 0.93±0.03 and ECD 1.6±1.1 mm) and small (DSC 0.91±0.04 and ECD 2.2±1.5 mm) variants produced target segmentations comparable or superior to TransMorph w/o PS fine-tuning (DSC 0.91±0.07 and ECD 2.6±1.4 mm) and within uncertainty to TransMorph w/ PS fine-tuning (DSC 0.94±0.03 and ECD 1.4±0.8 mm). The foundation models and TransMorph w/ PS fine-tuning both outperformed inter-observer variability (DSC 0.90±0.06 and ECD 1.7±0.7 mm, using STAPLE). The differences between all models but the 2 SAM2 variants were found to be statistically significant in a Friedman and post-hoc Nemenyi test.