S2271
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
Results: A 3 mm PTV margin ensured ≥ 95% CTV coverage in 100% (50/50) of fractions. A 2 mm margin achieved 95% coverage in 80% (40/50) of fractions, and a 1 mm margin in 46% (23/50). Vertical and longitudinal shifts were applied in 92% (46/50) and 90% (45/50) of fractions, respectively, with median shifts of 2 mm (SD 1.8 mm vertical, 1.7 mm longitudinal). Pearson correlation coefficients between CTV coverage (3 mm margin) and translational shifts were -0.24 (vertical) and -0.12 (longitudinal), indicating weak correlations. Conclusion: A 3 mm PTV margin remains necessary to ensure robust CTV coverage with a 3 mm gating threshold in MR-guided prostate SBRT. Reducing the margin to 2 or 1 mm results in insufficient coverage (<95%), indicating the need for tighter motion control. References: 1. Vanspeybroeck B, Bezuidenhout J, Soete G, et al. Proseven trial: MR-guided prostate stereotactic body radiotherapy in seven days - First results. Radiother Oncol. Published online October 17, 2025. doi:10.1016/j.radonc.2025.111208 Keywords: Intrafraction Motion, MRgRT, Prostate SBRT Digital Poster Highlight 4478 AI-based real-time liver localization on fixed-angle stereoscopic X-ray images Domagoj Radonic 1 , Tom Julius Blöcker 1 , Lili Huang 1 , Philipp Freislederer 2 , Elia Lombardo 1 , Chengtao Wei 1 , Vanessa Filipa da Silva Mendes 1 , Lucas Pieper 3 , Florian Putz 3 , Christoph Bert 3 , Claus Belka 1,4 , Stefanie Corradini 1,5 , Marco Riboldi 6 , Guillaume Landry 1 , Christopher Kurz 1 1 Department of Radiation Oncology, LMU University Hospital, LMU Munich, Munich, Germany. 2 Brainlab AG, Brainlab AG, Munich, Germany. 3 Department of Radiation Oncology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany. 4 Bavarian Cancer Research Center (BZKF), Bavarian Cancer Research Center (BZKF), Munich, Germany. 5 Now at Department of Radiation Oncology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany. 6 Department of Medical Physics, LMU Munich, Munich, Germany Purpose/Objective: X-ray fluoroscopy suffers from limited soft tissue contrast in the abdominal area, which makes accurate targeting difficult. This contribution examines methods for markerless localization of the liver as surrogate structure on stereoscopic X-ray images. Using prior anatomical knowledge and deep learning (DL)-based
auto-segmentation models, the liver is localized in real-time to facilitate markerless tracking in X-ray guided radiation therapy. Material/Methods: To simulate a dataset of abdominal X-ray images, CT scans were projected to digitally reconstructed radiographs (DRRs) according to the geometry of the ExacTrac Dynamic (ETD) (Brainlab AG, Munich, Germany). Projected motion phases of a 4DCT yielded a temporal sequence of DRRs (tDRR). The corresponding ground-truth liver contours were obtained by projecting the 3D contours—which were automatically generated by TotalSegmentator [1] from each 4DCT phase—onto their respective tDRRs. Three DL models, each implemented as a single, angle- agnostic network, were explored utilizing projected planning CTs (pDRR) and associated liver contours for stereoscopic tDRR segmentation: (1) a pre-trained SAM2 [2] received the pDRR contour as a prompt for tDRR contour localization; (2) a three-channel nnU- Net [3] received pDRR and pDRR contour as additional input channels to tDRR; (3) a DL-based deformation model (TransMorph [4]) deformed the pDRR contours to match tDRR anatomy. A total of 392 planning CTs (pCT) with associated 4DCTs were used for nnU-Net and TransMorph training. All the models were tested on an independent multi-center test set of 37 pCT- 4DCTs including non-same-day pairs to test robustness to anatomical variations. The quality of the predicted tDRR contours was assessed using Surface Dice (SD) ( τ = 3 mm), Hausdorff distance (HD), as well as the Euclidian Centroid Distance (ECD). Finally, a sequence of 7 ETD X-rays containing the liver, acquired periodically in free breathing for positional verification of a patient during hypofractionated spine SBRT, were used for validation on real X-ray data. Here, liver contours were delineated manually, while the pDRR contour was generated from the patient’s pCT. Results: TransMorph and nnU-Net outperformed SAM2 on the test dataset in terms of the metrics examined (Table 1). When validated on ETD X-rays of the liver (Figure 1), TransMorph achieved performance comparable to that on the DRR-only dataset. Inference times were under 40 ms, thus suitable for real-time applications.
Made with FlippingBook - Share PDF online