S1732
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
practice. References: 1.de Jong JI, Habraken SJM, Rojo-Santiago J,
evaluated in the calibration group by simulating 1000 treatments with normally distributed random setup ( σ =1mm) and systematic setup ( σ =0.92mm) and range ( σ =1.5%) errors. Each fraction was assigned the dose distribution of the nearest setup error in the grid and the nearest range error.Probability distributions of clinically relevant dose-volume metrics derived from SPE were compared with those derived from a ground truth based on simulating 35.000 Monte-Carlo calculations (1,000 treatments × 35 fractions) for each patient. Agreement was quantified using the mean percentile error (MPE), defined as the mean absolute difference across percentiles 0.01, 0.02 … 1.The optimal setup error grid (Emax = 3 σ , nsetup error = 33, tavg = 9 minutes) was then used for the validation group patients. Results: The median MPE decreased substantially as the number of setup scenarios increased from 7 to 33, with no further improvement observed between 33 and 123 scenarios (Figure 1). Increasing Emax from 3 σ to 4 σ did not reduce the overall MPE but did improve accuracy for values above the 98th percentile.Applying SPE to the validation set resulted in median errors of 0.02 GyRBE (range:-0.11 to 0.07) for the 10th percentile of the D99.8%,CTV distribution and 0.0 GyRBE (range:-0.14 to 0.23) for the 95th percentile of the D0.03cc,SpinalCord Core distribution (Figure 2).
Lathouwers D, Perkó Z, Breedveld S, Hoogeman MS. Probabilistic evaluation guided IMPT planning with realistic setup and range uncertainties improves the trade-off between OAR sparing and target coverage in neuro-oncological patients. Radiother Oncol. 2025 Sep 27;213:111171. doi: 10.1016/j.radonc.2025.111171. Epub ahead of print. PMID: 41022254. Keywords: Probabilistic evaluation, IMPT, HNC Digital Poster 2406 Deep learning robust intensity modulated proton planning in neurological cancer patients Esther Kneepkens 1 , Esther Decabooter 1 , Rik Emmah 1 , Richard Canters 1 , Cissy Stultiens 1 , Mirthe Pijls 1 , Julie Schoemakers 1 , Inge Compter 1 , Marleen van Iersel - Vet 1 , Danielle Eekers 1 , Rasmus Helander 2 , Christian Hahn 2 , Jacopo Parvizi 2 , Ivan Oleinik 2 , Maja Arvola 2 , Dennie Fransen 2 , Gabriel Carrizo 2 , Peter Thulin 2 , Fredrik Löfman 2 , Mirko Unipan 1 1 Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands. 2 Machine Learning Department, RaySearch Laboratories, Stockholm, Sweden Purpose/Objective: Neurological tumors are often located near critical organs of interest (OOIs), making treatment planning particularly challenging. Proton therapy (PT) offers a more favorable dose distribution that better spares these OOIs, but planning remains complex. To reduce treatment planning time and support rapid decision- making between proton and photon therapy, a deep learning model was developed to predict proton dose distributions directly from OOI contours. Using dose mimicking, these predictions were converted into robust, deliverable treatment plans [1], and their quality was evaluated in a blinded review. Material/Methods: Data from 186 patients with neurological tumors were retrospectively collected, either treated with or evaluated for proton therapy (PT). Manual PT plans were optimized to 50.4–59.4 GyRBE in 28–33 fractions, accounting for 1 mm setup and 3% range uncertainties. After curation, 100 patients remained and were split into training (n=80), validation (20), configuration (20), and test (17) sets. A U-net model predicted dose distributions from beam setup and 22 OOI contours. Predicted doses served as targets in a dose-mimicking step to generate robust, deliverable plans in a research version of RayStation 2024B. Dose volume histogram (DVH) parameters of
Conclusion: Sparse probabilistic evaluation achieves high accuracy while requiring low computation times. This paves the way for embedding probabilistic evaluation in clinical
Made with FlippingBook - Share PDF online