S1819
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Purpose/Objective: EPID-based in vivo dosimetry using the forward technique requires the accurate calculation of a predicted (reference) image to be compared (in terms of dose) with the acquired image. In this work, we introduce an original dose prediction method based on artificial neural networks using as input the anatomical model of the patient and DICOM RT Plan. Material/Methods: In our approach, two dose components are computed separately using a Collapsed Cone Convolution algorithm adapted to transit conditions. The attenuator transmitted dose (ATD) is built from the photons directly transmitted from the source, only taking into account the attenuation in the medium. The attenuator scattered dose (ASD) is built from the photons resulting from the interactions between the primary beam and the medium. Dose computations are optimized for a 6MV beam (Elekta SYNERGY with AGILITY multileaf collimator) with a calculation plane located 1600 mm from the source and at 50 mm depth in a water-equivalent medium mimicking the ELEKTA iViewGT.This work uses a U-Net for ATD prediction, while the ASD component is inferred by a cGAN (conditional Generative Adversarial Network). Training data is based on 3978 inputs/ground-truth pairs (3,177 training, 801 testing) extracted from irradiation segments of VMAT treatments of various tumor sites and covering all incidences around patient CT or parallelepipedic/anthropomorphic phantoms. For each configuration, the primary beam fluence, the radiological thickness of the attenuator, and the geometric position of the attenuator between the source and the EPID are used as the inputs.The comparison of the inferred images with the ground- truth data is established by gamma analysis (2%-2mm- global, 10% dose max threshold) performed on both ATD and total dose. Due to its specific characteristics, different gamma criteria (10%-2mm-global) are applied for ASD evaluation. Results: Across the entire test database, the average gamma passing rate (GPR) is 99.97% ( σ = 0.08) for the ATD component and 91.70% ( σ = 20.81) for the ASD component. Given the smaller ASD contribution, an average GPR of 99.97% ( σ = 0.08) is achieved in terms of total dose. Conclusion: This study demonstrates the ability of neural networks to accurately produce predictive total dose calculations in transit dosimetry specific conditions. The combination of a U-Net and a cGAN enables the physical phenomena dissociation in such configurations to deliver an accurate total dose prediction. Keywords: In vivo dosimetry, generative adversarial networks
Conclusion: STF treatments can substantially reduce healthy brain BED2 compared to uniformly fractionated SRS. By adjusting optimization priorities, STF enables flexible trade-offs between minimizing mean brain BED and limiting high-dose brain volumes. References: [1] Torelli, N., Papp, D., & Unkelbach, J. (2023). Spatiotemporal fractionation schemes for stereotactic radiosurgery of multiple brain metastases. Medical physics, 50(8), 5095– 5114. https://doi.org/10.1002/mp.16457. Keywords: fractionation, treatment planning, BED EPID-based in vivo transit dosimetry in external beam radiotherapy: prediction of portal dose images using artificial neural networks Côme Mével Dutertre 1,2 , Alexandre Hakimi 1 , Eric Fadel 1 , François Smekens 3 , Xavier Franceries 2 , François Husson 1 1 Physics R&D, DOSIsoft, Cachan, France. 2 ToNIC, Université Toulouse III Paul Sabatier, Toulouse, France. 3 Pysics R&D, DOSIsoft, Cachan, France Digital Poster 1141
Made with FlippingBook - Share PDF online