S1789
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
2 Faculty of Physics and Astronomy, Heidelberg University, Heidelberg, Germany. 3 Scientific Computing Center, Karlsruhe Institute of Technology, Heidelberg, Germany Purpose/Objective: The application of AI-based methods gains increasing attention within the radiotherapy planning workflow, e.g. for dose calculation or outcome prediction. To date, however, there are only a limited number of open source software packages designed for embedding and exploring such methods. We advertise pyRadPlan (Wahl et al. 2025), a Python-based treatment planning software (TPS) toolkit, which is specifically tailored towards the use of AI-driven features. Material/Methods: pyRadPlan is an open source radiotherapy treatment planning toolkit based on the established preclinical TPS matRad (Wieser et al. 2017). It provides a matRad- interoperable planning framework for different treatment modalities (photons, ions, VHEE), with the flexibility to chain together dose calculation and optimization algorithms. Figure 1 displays the overall structure and a generic example for the pyRadPlan interface.Across all workflow steps, pyRadPlan supports AI-based tools by leveraging CPU/GPU array backends (numpy, cupy, pytorch) through its agnostic code using a strict implementation of the Python array API. For validation and serialization, we rely on pydantic, which includes AI-empowered functionalities via their Python agent framework, e.g. prompting beam angles from a large language model. pyRadPlan’s dose calculation module has two branches, one for pencil beam kernel methods, and one for Monte Carlo engine interfaces. This module also facilitates the integration of AI-based dose calculation engines. Beyond dose calculation, pyRadPlan features conventional treatment plan optimization methods with the opportunity to step up towards outcome-guided planning using machine learning prediction models.
Results: We showcase the treatment planning capabilities of pyRadPlan by generating conventional and machine learning outcome model-based proton treatment plans for a head-and-neck patient. In both plans, we apply Hong’s pencil beam algorithm for dose calculation, while for dose optimization, we impose conventional objective functions on the left/right parotid, skin and PTVs. Moreover, we rely on the bounded L-BFGS implementation from SciPy for problem solving. The machine learning outcome model-based plan additionally incorporates a logistic regression model fitted on a cohort of 153 patients with 4 features (mean/maximum dose, dose skewness/kurtosis) for each parotid, predicting risk for long-term grade 2+ xerostomia. We depict the optimization results in Figure 2, indicating the clinical acceptability of the plan and the feasibility of risk mitigation.
Made with FlippingBook - Share PDF online