S1903
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
grid of 3 × 3 × 3 mm3. 5940 segment doses were simulated with 5 × 105/cm2 photons for training (11 patients), 540 segment doses with 5 × 106/cm2 photons for validation (3 patients), and 1293 segment doses with 5 × 106/cm2 photons for testing (5 full plans from 5 testing patients). For BEV coordinate, CNN- ConvLSTM, CNN-Mamba, DoTA [3] and UNet (C3D) [5] models were trained by inputting CT and segment projection BEV cuboids (cuboid shape: 256 × 200 × 200, voxel size: 2 × 2 × 2 mm3). For patient coordinate, a UNet (DeepDose-C3D) model was trained by inputting cropped CT volumes (volume shape: 128 × 192 × 192, voxel size: 3 × 3 × 3 mm3) with four physical inputs [4] (see Figure 1). The global gamma passing rate γ PR (2%/2mm, D>10%Dmax) for segment doses and γ PR (1%/1mm, D>10%Dmax) for plan doses were evaluated (plan doses obtained by summing predicted segments). The dose calculation time including model inference and pre/post-processing among different models on an NVIDIA RTX PRO6000 Max-Q (96 GB) is reported.
Conclusion: BEV coordinate modeling appears more robust than patient coordinate for segment dose prediction. Both methods work well for plan dose prediction. The CNN- ConvLSTM and CNN-Mamba show competitive accuracy and faster inference compared with DoTA, C3D, DeepDose-C3D for segment and plan dose calculation. References: [1] Keall, Paul J., et al. Critical Review: Real-Time Dose- Guided Radiation Therapy. Int J Radiat Oncol Biol Phys. 2025;122(4):787-801.[2] Cheng B, et al. Development and clinical application of a GPU-based Monte Carlo dose verification module and software for 1.5 T MR- LINAC. Med Phys. 2023;50(5):3172-3183.[3] Pastor- Serrano O, et al. Sub-second photon dose prediction via transformer neural networks. Med Phys. 2023;50(5):3159-3171.[4] Kontaxis C, et al. DeepDose: Towards a fast dose calculation engine for radiation therapy using deep learning. Phys Med Biol. 2020;65(7):075013.[5] Liu S, et al. Technical Note: A cascade 3D U-Net for dose prediction in radiotherapy. Med Phys. 2021;48(9):5574-5582. Keywords: Dose calculation, deep learning, VMAT Digital Poster 2828 Standardizing MRI-only radiotherapy commissioning: benchmark dataset and tolerance levels from the MESCAL initiative Davide Cusumano 1 , Matteo Maspero 2 , Luca Vellini 1 , Emilie Alvarez-Michael 3 , Anais Barateau 4 , Igor
Results: Figure 2 (a) presents boxplots of γ PR for predicted segment doses and Figure 2 (b) shows the testing plan dose predicted by different models. The averaged segment and plan dose calculation times for CNN- ConvLSTM, CNN-Mamba, DoTA, C3D, Deep-C3D were 79, 64, 298, 490, 356 ms, and 4.8, 5.2, 18.7, 31.0, 30.3 s, respectively, including overhead BEV resampling computations.
Made with FlippingBook - Share PDF online