S2265
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
tuning gains highlight its potential for real-time tumour tracking, as the amount of high quality labelled data is scarce in this domain.
University Eindhoven, Eindhoven, Netherlands. 3 Department of radiation oncology, Catharina Hospital Eindhoven, Eindhoven, Netherlands. 4 Department of Electrical Engineering, Technical University Eindhoven, Eindhoven, Netherlands Purpose/Objective: Vendor-provided tumour tracking solutions typically rely on image registration. In research, deep learning- based segmentation models for real-time tracking on 2D cine MRI are increasingly explored. Early models were lightweight and task-specific, often built on U-Net architectures. More recently, we see a shift toward large, general-purpose foundation models such as the Segment Anything model (SAM), trained on extensive non-medical datasets. This trend is reflected in the TrackRAD2025 submissions. This study compares the MedSAM-2 [1] foundation model with the conventional nnUNet [2] architecture and evaluates the impact of fine-tuning on a 2D cine MR-linac dataset. Material/Methods: A 2D nnUNet was trained from scratch on 0.35 T cine sequences from 24 patients from the TrackRAD2025 dataset [3], using 19 patients for training and 5 patients for validation. The test set included an additional 14 patients each annotated by 5 observers. The simultaneous truth and performance level estimation (STAPLE) algorithm was used to generate a single ground truth. Training ran for 300 epochs selecting the best checkpoint for inference.MedSAM-2 was fine-tuned on the same 19 patients for 15 epochs, with the final checkpoint used for inference, as validation showed consistent performance across checkpoints. Additionally, baseline MedSAM-2 was deployed as zero-shot, without finetuning, on the test set. Both models received a fixed input (first image and its contour) and a variable image from the cine time series. nnUNet used no temporal context while MedSAM-2 incorporated past frames. Performance was evaluated using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), average surface distance (ASD), and center of mass (COM) Despite careful nnUnet optimization, MedSAM-2 showed better performance. Notably, zero-shot MedSAM-2 performed comparably to the fine-tuned version, suggesting fine-tuning may be unnecessary. Figure 1 shows model performance. MedSAM-2 results fell within interobserver variability. Average inference times were 38 ms for MedSAM-2 and 144 ms for nnUNet. Conclusion: distance, compared to the manual labels. Interobserver variability was also calculated. Results: Zero-shot MedSAM-2 achieved improved performance and faster inference times. The foundation model's strong generalization capabilities and minimal fine-
References: [1] Ma J et al., arXiv preprint 2504.03600, 2025. https://doi.org/10.48550/arXiv.2504.03600[2] Isensee F et al., Nature Methods 18, 2021. 10.1038/s41592-020-01008-z[3] Wang Y et al., Medical Physics 52, 2025. https://doi.org/10.1002/mp.17964 Keywords: Foundation models, Autosegmentation, Deep learning
Proffered Paper 4169 Real-Time Dose-Guided Motion Mitigation
Strategies for Prostate SBRT: A Simulation Study Karolina Alexandra Klucznik 1 , Thomas Ravkilde 2 , Laura Happersett 3 , Brian Leong 3 , Pengpeng Zhang 3 , Grace Tang 3 , Per Poulsen 1,2 1 Danish Center for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark. 2 Department of Oncology, Aarhus University Hospital, Aarhus, Denmark. 3 Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA
Made with FlippingBook - Share PDF online