S1713
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
Our results demonstrate that a deep learning-based approach can significantly improve the specificity of SDC controls in MR-linac workflows by accurately correcting for the ERE in CCC-based dose computations. By reducing false positives in PSQA while adding a negligible additional inference time per beam, this method may facilitate broader clinical adoption of automated SDC into the time-sensitive online adaptive workflow. Keywords: MR-linacs, neural networks, PSQA Impact of high linear-energy transfer (LET) particle dose on neuroanatomical changes in pediatric craniopharyngioma Fakhriddin Pirlepesov, Aahan Arif, Vadim P Moskvin, Thomas E Merchant Radiation Oncology, St Jude Children's Research Hospital, Memphis, USA Purpose/Objective: Neuroanatomical changes have been associated with radiotherapy in pediatric brain tumors [1]; however, the effects of proton beam radiotherapy (PBRT) are not yet well understood. This longitudinal study investigates neuroanatomical alterations in children treated with PBRT for craniopharyngioma and examines their relationship to regions exposed to high linear energy transfer (LET) particle doses. Material/Methods: We analyzed 96 T1-weighted MPRAGE MRI scans from 24 patients who completed five years of follow-up in a prospective PBRT trial. Scans were obtained before PBRT and at 1-, 3-, and 5-year intervals. All patients received a single 30-fraction (54 CGE) treatment plan. Field-by-field physical dose and dose-weighted LET (LETd) were calculated using MCSquare Monte Carlo code [2]. Automated segmentation was performed with FreeSurfer (v7.4.1) and manually reviewed [3]. Mini-Oral 1169 Baseline MRI scans were linearly registered to planning CT, with deformable registration for subsequent timepoints. Mean dose and LETd to subcortical structures were calculated, and net volume changes were assessed using Wilcoxon tests and linear regression. Results: The median age of participants was 9.66 years (range: 4.02–20.0), with 54.2% male. Five years after treatment, the right pallidum showed the largest size reduction ( Δ 5yrs = -5.66%), while the left hippocampus exhibited the greatest growth ( Δ 5yrs = 1.87%). Univariate linear regression revealed that younger age at enrollment was significantly associated with reduction of the left amygdala ( β = -0.74, p = 0.038) and growth of the right pallidum ( β = 1.44, p = 0.029).
Digital Poster 1140 Deep Learning-based dose distribution correction of Electron Return Effect in MR-guided radiotherapy Alexandre Hakimi 1 , Eric Fadel 1 , François Smekens 1 , Clément Chevillard 2 , François Husson 1 1 Physics R&D, Dosisoft, Cachan, France. 2 Medical Phycisist, Institut Curie, Paris, France Purpose/Objective: In time-constrained adaptive workflows of MR-linacs, the only feasible form of Patient Specific Quality Assurance (PSQA) is the secondary dose calculation (SDC), which must be both fast and reliable. Although TPS often rely on Monte Carlo (MC) to deliver high- accuracy dose calculations in presence of a magnetic field, the long computation times make MC unsuitable for time-efficient SDC. Collapsed Cone Convolution (CCC) algorithms offer significantly faster computation but their accuracy can be locally compromised in magnetic field environments notably due to the Electron Return Effect (ERE). This work investigates the use of a convolutional neural network (CNN) trained on MC-calculated dose distributions to correct CCC- computed doses for Elekta Unity 1.5T MR-linac, aiming to combine the speed of CCC with the accuracy of MC. Material/Methods: A CNN based on a 3D U-Net architecture was trained to predict ERE correction to dose-to-medium distributions using paired MC (Elekta Monaco GPUMCD v6.2) and CCC per-beam dose calculations as well as patient density matrices. The MC simulations (ground truth) were performed with minimal statistical uncertainty, and CCC doses were computed with ThinkQA SDC solution (Dosisoft) using optimized beam modelling dedicated to Unity MR-linac. This allows the CNN to focus on learning the residual correction specific to the ERE marked at interfaces between tissues and low densities. Model performance was evaluated on an independent set of patients with ERE- sensitive locations, using 3D gamma passing rate (GPR 2%/2 mm, 10% threshold) and DVH comparisons for target volumes and organs of interests. Results: In well-conditioned situations where the ERE is the primary expected source of discrepancy between MC and CCC dose calculations, the use of the CNN improved the GPR from approximately 92% on average with CCC alone to over 98%. DVH analysis showed reductions of D95 discrepancies inside the PTV and decreased dose deviations in organs of interest. Notably, the CNN corrections were mostly localized to regions affected by the ERE, demonstrating the model’s ability to focus on dose perturbations induced by the magnetic field. Conclusion:
Made with FlippingBook - Share PDF online