ESTRO 2025 - Abstract Book

S3250

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

ESTRO 2025

Model runtime was 200 ms for SAM2-large and 109 ms for SAM2-small and was sufficient for real-time application at 4-8 Hz. Both TransMorph variants shared an inference time of 32 ms. TransMorph was trained for 4 days. TransMorph w/ PS fine-tuning furthermore requires the contouring of 8 frames from a patient cine-scan and 10 minutes for further training. The foundation models only required a single contoured frame as initialisation as shown in Figure 2.

Conclusion: This study demonstrates that foundation models can achieve high-quality real-time target localization in MRgRT, matching state-of-the-art methods without PS fine-tuning. Developments like medical imaging specific foundation models such as MedSAM should be explored.

Keywords: MRI-linac,respiratory motion,deep learning

References: Lombardo, Elia, et al. "Patient-specific deep learning tracking framework for real-time 2D target localization in MRI guided radiotherapy." International Journal of Radiation Oncology* Biology* Physics (2024).

2262

Digital Poster A patient-specific 2D contour prediction model for sub-millimetre tracking in real-time pancreas SBRT using deep learning Jeremy T Booth 1,2 , Abdella Ahmed 1 , Levi Madden 1 , Maegan Stewart 1 , Gabrielle Metz 1 , Meegan Shepherd 1 , Carlito Coronel 1 , Paul Keall 3 , Andrew Kneebone 1 , George Hruby 1 1 Northern Sydney Cancer Centre, Royal North Shore Hospital, Sydney, Australia. 2 Institute of Medical Physics, School of Physics, University of Sydney, Sydney, Australia. 3 Image-X Institute, University of Sydney, Sydney, Australia