Tree-Based Beam Angle Optimisation for Proton Therapy: An Automated Planning Tool
Abstract
Purpose
To develop and test an automated pipeline to pre-determine optimal beam angles for proton beam therapy before planning begins. This provides treatment planners with clinically viable beam configurations without the time investment typically required for manual exploration, alongside additional metrics such as LET and RBE to support informed decision-making.
Methods
Dose profiles were computed for a set of starting beam angles spanning the design space using an open-source GPU Monte Carlo software [Schiavi et al, 2017]. Starting angles were selected using a novel fast selection algorithm—without dose calculation—to allocate computation efficiently. For each angle, beam weights were optimised using gradient-descent to produce fields that robustly covered the target volume. The final dose considered the worst-case dose delivered to organs due to set up errors. A tree-based search algorithm with Bayesian optimisation was employed to iteratively select subsequent beam angles, prioritising regions of the design space with highest expected improvement. Each evaluated angle maintained robust target coverage while meeting organ-at-risk constraints to produce a set of Pareto optimal beam angles. Multi-beam configurations were optimised using a multi-objective algorithm that minimises dose to critical structures and LET at overlapping Bragg peaks, producing optimal beam sets for a user-defined number of beams.
Results
The pipeline evaluates offline without user intervention, enabling systematic exploration of the design space beyond what manual selection permits. For brain cases, optimal angles can be found after a few hours. Generated plans achieve D95% target coverage whilst allowing on-the-fly angle exploration and LET-informed optimisation before computing dose with a Treatment Planning System. The approach has been tested on brain cases with expansion to additional treatment sites planned for future work.
Conclusion
GPU-accelerated dose calculation combined with model-guided search enables generation of optimal proton beam angles within clinically practical timeframes, potentially improving planning efficiency and consistency.