S1779
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
task, anatomical and clinical data (CT, RTSTRUCT, prescription) are processed by a diffusion model [2] to generate a compact dose representation, which is decoded into the 3D dose distribution using FoDoM-RT (Figure1). The architecture of the diffusion model was adapted to emphasize critical structures (e.g. PTVs, brainstem) and the relative influence of inputs was optimized for improved clinical relevance. Reconstruction (FoDoM-RT only) and prediction (generative model + FoDoM-RT decoder) performance were assessed using standard DVH metrics and Dose/DVH scores [3] relative to the clinical dose (ground truth). Results were compared across tumor sites to assess generalization capacity. Statistical significance was assessed using the Wilcoxon signed rank test. Results: FoDoM-RT achieved excellent reconstruction, with DVH metrics nearly identical to clinical doses (p>0.05 for all metrics except spinal cord mean dose p ≈ 0.02; Figure2) across all prescriptions and tumor sites. In the dose prediction task, Dose and DVH scores remained consistent across tumor locations, demonstrating the generative model’s ability to generalize. Of twenty-four comparisons (six metrics per site; others excluded from the statistical analysis due to limited sample size), seven showed significant differences, including lower maximum and mean brainstem doses for hypopharyngeal and laryngeal tumors, and reduced spinal cord doses (p<0.05). Predicted doses to target volumes exhibited higher heterogeneity (p<0.05), indicating a limitation of the current implementation. No other differences were significant. Conclusion: FoDoM-RT provides a novel framework for standardized and compact 3D dose representation, enabling consistent dose modeling for treatment planning, quality assurance, and outcome analysis. When applied to a dose prediction task, it achieved robust and controllable predictions across diverse prescriptions and tumor sites. These findings highlight its potential to enhance reproducibility and enable multi-institutional radiotherapy research.
PVDR of 2.87 ± 0.28 and an ADR of 2.16 ± 0.32. This approach not only surpassed VMAT in dose distribution but also ensured efficient treatment duration control. Conclusion: The novel hybrid SFRT strategy combines CyberKnife (delivering high doses to lattice points) with VMAT (providing low-dose coverage to the entire GTV), maximizing the peak-to-valley dose ratio (PVDR) and ablation ratio (ADR) while reducing treatment time. This approach enhances the therapeutic efficacy of SFRT, particularly for large-volume tumors, and will be further validated in future clinical studies. Keywords: Lattice, Cyberknife Digital Poster Highlight 4403 FoDoM-RT: A Foundational Model for Dose Representation in Comprehensive Head-and-Neck Radiotherapy Jeanne Boyer-Chammard 1,2 , Quentin Spinat 1 , Madalina Costea 3 , Pauline Maury 4 , Charlotte Robert 4 , Marie- Claude Biston 5 , Nikos Komodakis 1,6 , Nikos Paragios 7,8 , Eric Deutsch 4 , Vincent Grégoire 5 1 R&D Artificial Intelligence, TheraPanacea, Paris, France. 2 Joint Collaboration : Gustave Roussy INSERM 1030 / Department of Radiation Oncology, Université Paris-Saclay / Centre Léon Bérard, Villejuif / Lyon, France. 3 Cinical Affairs, TheraPanacea, Paris, France. 4 Gustave Roussy, INSERM 1030, Université Paris- Saclay, Villejuif, France. 5 Department of Radiation Oncology, Centre Léon Bérard, Lyon, France. 6 Computer Science Department, University of Crete, Rethymno, Greece. 7 CEO, TheraPanacea, Paris, France. 8 Centrale-Supélec, Université Paris-Saclay, Gif-sur- Yvette, France Purpose/Objective: Variability in radiotherapy planning across institutions challenges treatment standardization, reproducibility, and outcome modeling. We propose FoDoM-RT, a foundational model that learns compact, and fully reconstructible representations of radiotherapy dose distributions. As a demonstration, we apply FoDoM-RT to anatomically-guided dose prediction in Head and Neck (H&N) cancer. Material/Methods: A retrospective dataset of Head-and-Neck cancer patients treated with VMAT (2017–2024) and prescriptions ≥ 50 Gy was used, including one to three PTVs across all tumor sites (512 training, 53 testing cases). FoDoM-RT uses an encoder-decoder architecture: the encoder compresses the dose distributions into a compact, standardized representation, and the decoder reconstructs the corresponding 3D dose. For the dose prediction
Made with FlippingBook - Share PDF online