S1721
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
doses. The novel dose mimicking converts DL- predicted doses into deliverable positions and intensities of pencil-beam spots. It uses quadratic overdose penalties on all voxels. For targets, also quadratic underdose penalties are used. In addition, constraints for voxelwise (VW) minimum and maximum doses are set (ensuring clinical hard constraints are met in all scenarios), as well as
seen with proton therapy compared to photon plans with similar target extension.This finding is important for the small subgroup of patients with a prophylactic high or ultra-high target, in whom the risk of severe side effects may be reduced. Therefore, IMPT should be considered based on individual plan comparisons for this patient group. References: [1] Nilsson MP, Undseth C, Albertsson P, Eidem M, Havelund BM, Johannsson J, et al. Nordic anal cancer (NOAC) group consensus guidelines for risk-adapted delineation of the elective clinical target volume in anal cancer. Acta Oncol. 2023 Aug;62(8):897-906.[2] Pan YB, Maeda Y, Wilson A, Glynne-Jones R, Vaizey CJ. Late gastrointestinal toxicity after radiotherapy for anal cancer: a systematic literature review. Acta Oncol. 2018 Nov;57(11):1427-1437.[3] Sodergren SC, Vassiliou V, Dennis K, Tomaszewski KA, Gilbert A, Glynne-Jones R, et al; EORTC Quality of Life Group. Systematic review of the quality of life issues associated with anal cancer and its treatment with radiochemotherapy. Support Care Cancer. 2015 Dec;23(12):3613-23. Keywords: Anal Cancer, IMPT, Side Effects Fast, fully-automated and constrained IMPT dose mimicking to convert deep learning predicted doses into high quality deliverable plans Hazem Nomer 1 , Franziska Knuth 1 , Joep van Genderingen 1 , Linda Rossi 1 , Uwe Oelfke 2 , Jos Elbers 1 , Ben Heijmen 1 , Sebastiaan Breedveld 1 1 Department of Radiotherapy, Erasmus MC Cancer Institute, University Medical Center, Rotterdam, Netherlands. 2 Joint Department of Physics, The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, London, United Kingdom Purpose/Objective: Deep learning (DL) offers instantaneous dose prediction, but not the machine parameters for dose delivery. We developed a fast, fully-automated and constrained dose mimicking approach to accurately convert DL-predicted IMPT doses into deliverable Digital Poster 1719 plans that closely resemble ground truth (GT) autoplans. Like GT-plans, mimicked plans have guaranteed plan robustness and strictly adhere to clinical hard constraints. Material/Methods: A dataset of 1250 head-and-neck patients with various tumour locations was split into training (1000), validation (100) and test (150) sets. For all patients, GT- plans were generated using constrained multi-criterial autoplanning for our 70/54.25 Gy protocol, using 21 robustness scenarios. An HD-Unet for DL dose predictions was trained using the test/validation GT
constraints ensuring deliverability. The training/validation patients were used for
development and parameter tuning of the mimicking routine. For final evaluations, mimicked plans for the 150 test patients were compared with the DL- predictions they were based on, and with their GT autoplans. Robustness of the plans was evaluated through VWminimum/maximum. NTCPs for important side-effects were calculated[1]. Wilcoxon signed-rank tests were used to test the statistical significance of observed differences. Results: Clinical and delivery constraints were met for all GT, predicted and mimicked plans. Median deviations of mimicked doses from GT and predicted doses were close to zero (Figure 1a) with small IQR. Figure 1b shows deviations for NTCPs. All eight median NTCP differences were ≤ 0.5%-point, while the maximum IQR was 0.94%-point, with most IQRs even much smaller. For the mimicked plans compared to the GT-plans, NTCPs improved for xerostomia for 63 patients (both grades), and 95 and 104 patients for dysphagia, grade ≥ 2 and ≥ 3 respectively. Figure 2 shows that the dose distributions stay highly similar. On average the mimicking optimizations take 6.8 minutes. Conclusion: The proposed fast and fully-automated dose mimicking accurately converted DL-predicted IMPT doses of head-and-neck patients into deliverable plans that fully respected all imposed clinical and delivery constraints. Mimicked doses also closely resembled ground truth autoplans. No manual fine tuning of mimicked plans was needed.
Made with FlippingBook - Share PDF online