ESTRO 2025 - Abstract Book

S2918

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2025

accelerate the planning process; however, their utility is often limited as decision support systems due to the closed nature of commercial treatment planning systems (TPS). The aim of this study was to investigate the feasibility and quality of a pipeline for integrating DL dose prediction outputs into clinical plan adaptation during MRgRT. Material/Methods: Structure sets and dose maps from 262 fractions of 25 prostate cancer patients, treated with 60 Gy in 20 fractions, were acquired using an automatic treatment planning approach based on a particle swarm optimization for a high consistency in plan quality [1]. The resulting dataset was used to train a DL based dose prediction model for MRgRT. The model, based on the SwinUnetr architecture, employed a combined loss function of root mean square error and dose volume histogram difference to create 3D dose distributions from structure maps. This model was integrated into a pipeline that used the dose prediction results to calculate dose constraints, which were subsequently applied to a plan template to generate applicable plans within the clinical TPS Monaco without human intervention (Figure 1). The resulting plans were evaluated regarding compliance to institutional planning criteria (Table 1). Due to the limited availability of MRgRT patients, a separate dataset of 20 fractions from 20 patients, which were treated at a conventional linear accelerator with offline MR-Images available, was used for evaluation.

Results: In 17 out of 20 test cases, at least 13 of 14 clinical criteria were achieved (Table 1). Target coverage was the most common reason for failure, with three and six plans missing the criteria for the respective target volumes slightly. With the exception of two cases showing notable deviations in bladder V56 Gy or rectum circumference, all other