S1763
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
Results: The engine achieved sub-second runtimes per Iso Energy Slice (IES), averaging 0.73 ± 0.3 s across three representative energies (short-, mid-, and long-range). Full-field runtimes ranged from 18 s (21 IESs, 1537 PBs) to 81.6 s (35 IESs, 9656 PBs; Figure_1) on a single GPU, with nearly twofold acceleration on a dual-GPU setup (44 s to 24 s). Across seven test patients, the mean gamma pass rate (2 % / 2 mm) was 99.2 ± 0.5 % (Table_1), showing excellent agreement with MC simulation. The model footprint per GPU per energy was < 1.6 MB (FP32), enabling concurrent multi-energy loading and efficient multi-GPU scaling.
therapy plan optimization with energy layer pre- selection driven by organ at risk sparing and delivery time. Phys Med Biol. 2025;70. doi:10.1088/1361- 6560/adad2d[3] Pross D, Wuyckens S, Deffet S, Sterpin E, Lee JA, Souris K. Technical note: Beamlet-free optimization for Monte-Carlo-based treatment planning in proton therapy. Med Phys. 2024;51: 485– 493 Keywords: Monte-Carlo, Beamlet-Free Proffered Paper 3957 A Model-Agnostic AI-based Dose Engine for Real- Time Proton Dose Calculation Ahmad Neishabouri 1 , Amir Abdollahi 1 , Andrea Mairani 2 1 E210, Deutsches Krebsforschungszentrum (DKFZ), Heidelberg, Germany. 2 Heidelberg Ion-Beam Therapy Center, Department of Radiation Oncology, Heidelberg, Germany Purpose/Objective: Proton therapy enables highly precise, individualised dose delivery with superior OAR sparing thanks to the characteristic Bragg peak. This precision increases the demand for real-time dose verification, as even millimetric deviations can compromise target coverage or induce OAR overdose. Modern treatment rooms already acquire continuous imaging to monitor intra/inter-fractional motion for adaptive treatment planning, yet these data streams remain underutilized due to the absence of instantaneous dose engines. AI- based dose calculation can bridge this gap by enabling ultra-fast dose prediction. We present a compact, model-agnostic AI dose engine capable of integrating any pencil-beam (PB) model as its computational core, allowing modular trade-offs between accuracy and speed. Material/Methods: A pre-validated Convolutional ConvLSTM (CC-LSTM) [1] model was incorporated to demonstrate the framework’s capacity for instantaneous, mono- energetic dose calculation under clinically realistic, multi-GPU conditions. The CC-LSTM was commissioned for the clinical beamline using machine- specific data comprising ~35000 PBs across 35 discrete energies (~1000 per energy). Each beam’s dose distribution was simulated with the FLUKA MC code (105 histories per PB) to serve as ground truth for supervised learning and evaluation.Evaluation was performed on seven low-grade glioma head-and-neck patients, none included in model training. Four exhibited beam impinging angles not represented in the training set, testing generalization not only on unseen geometries but also beam pathways. MC- based dose distributions served as reference, and agreement was quantified using 3D gamma analysis.
Conclusion: A model-agnostic, AI-based dose engine was developed and validated through commissioning a machine-specific spatiotemporal model for proton therapy. The system achieved sub-second dose prediction per IES with near-MC accuracy, demonstrating feasibility for real-time deployment. Future work will test the dose engine with different computational cores, including alternative architectures [2], end-to-end MR-based dose calculation models [3], and models for other modalities (Carbon), where AI-based dose calculation can unlock previously unattainable clinical paradigms.
Made with FlippingBook - Share PDF online