S1863
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
variant of the MedNeXT architecture [3], which is an encoder-decoder equipped with ConvNeXT blocks. The predicted and clinical dose distribution were compared with 3%/3mm gamma passing rates (GPR). The predicted doses were imported into the clinical treatment planning system (Eclipse v16, Varian Medical Systems), where dose objectives were automatically extracted for planning. The optimised doses were compared to the clinical doses with dose- volume metrics, the conformity index CI100, and the modulation complexity score (MCS). Plan complexity increases with decreasing MCS. Plan quality was evaluated with the number of achieved clinical goals. The time to generate a plan was recorded.
(Figure 1). Some organs at risk, such as the oral mucosa and salivary glands in Case 2, show in the DVHs that up to 10% and 20% of the volumes receive about twice their mean doses. In case 1, out of the beam, the most representative organs received less than 1.3 mSv, corresponding to the right eye lense.The isodose distribution (Figure 2) indicates a strong contribution from scattered electrons and bremsstrahlung radiation. Conclusion: The peripheral dose values found are sufficiently significant to warrant consideration, highlighting the need for careful dosimetric evaluation in electron FLASH treatments. Further studies could aid in designing applicators and shielding to improve patient protection for the accelerator model under study. Keywords: Flash, UHDR, Electrons, peripheral dose Deep-learning dose prediction-driven automated radiation therapy planning for same-day single- fraction treatment of solitary metastasis in the lung Mathieu Gaudreault 1,2 , Nicholas Hardcastle 1,2 , Katrina Woodford 3,2 , Jason Li 4,2 , Kenton Thompson 3,2 , Susan Harden 3,2 , Sandro Porceddu 3,2 , Vanessa Panettieri 1,2 1 Physical Sciences, Peter MacCallum Cancer Centre, Melbourne, Australia. 2 Sir Peter MacCallum Department of Oncology, the University of Melbourne, Melbourne, Australia. 3 Radiation Oncology, Peter MacCallum Cancer Centre, Melbourne, Australia. 4 Bioinformatics Core Facility, Peter MacCallum Cancer Centre, Melbourne, Australia Proffered Paper 1963
Purpose/Objective: Same-day single-fraction stereotactic ablative
radiation therapy (SABR) to solitary metastasis in the lung requires an efficient planning workflow [1,2]. We introduce and compare two deep-learning (DL) dose prediction workflows generating a clinically acceptable plan from the computed tomography (CT) image and tumour segmentation. Material/Methods: Treatment data for consecutive patients with lung metastasis or primary tumour at a single institution between 11/2018 and 11/2024 were included. The prescription dose was either 28 Gy in one fraction (metastasis only) or 48 Gy in four fractions (metastasis and primary). Both workflows used as inputs a density map derived from the CT image and a synthetic dose distribution produced from the tumour segmentation. One workflow also used DL-predicted organs at risk (OAR) as inputs (with-OAR), whilst the other workflow did not (without-OAR), but instead used generic structures for planning (Fig. 1). The network was a
Results: The dataset was split into 200 / 56 for training / testing. The testing dataset included the single fractionation only. By considering all voxels, GPR mean ± standard deviation were 99.1 ± 0.9% / 99.0 ± 1.0% in the with- / without-OAR workflow. All clinical goals were achieved for all patients in plans from both workflows, which resulted in 25% patients with improved dosimetry (Fig. 2). As compared with clinical settings, dose conformity was improved (CI100 = -3.9 % / -5.3% of the clinical value in the with- / without- OAR workflow, p-value < 0.001), at the cost of higher complexity (MCS = -19% of the clinical value in both workflows, p-value < 0.001). On average, a plan was generated in 9.8 min / 8.9 min by the with- / without-
Made with FlippingBook - Share PDF online