Room: Stars at Night Ballroom 1
Purpose: There are in general many possible beam angle configurations in proton therapy and optimizing each candidate is time consuming, making beam angle optimization clinically infeasible. We present a fully automated treatment planning scheme for proton therapy including beam angle selection using Bayesian optimization and constrained hierarchical optimization.
Methods: Our in-house automated treatment planning system for photon therapy, which is based on constrained hierarchical optimization and called ECHO, is adapted for proton therapy. We have integrated Bayesian optimization for beam angle optimization into ECHO. Bayesian optimization is used to efficiently address selection of the optimal beam angles. ECHO is run for some initial beam angle candidates and a clinically relevant treatment score is used to evaluate the resultant plan for each configuration. Bayesian optimization iteratively predicts the score for not-yet-evaluated candidates to find the best candidate to be optimized next with ECHO, and it finds the optimal candidate after only optimizing a few candidates. This technique has been tested on four head-and-neck patients. For each proton plan, two beam angles, out of 36 evenly distributed beams, are selected.
Results: All 558 possible candidates are optimized to find the ground truth optimal configuration (with the lowest treatment score) for comparison purposes. Bayesian optimization finds the optimal configuration after only evaluating on average 9% of the beam configurations (range: 6-14%). Compared with the beam angle configurations chosen by the planner, the optimal configurations reduce parotid mean dose, and cord and mandible maximum doses on average by 4.5%, 1.3%, and 17.3% relative to the prescription dose, respectively, while maintaining the same PTV coverage (D95%).
Conclusion: A fully automated treatment planning workflow including beam angle optimization was implemented, and improvements in OAR sparing were obtained. Bayesian optimization could find the optimal beam angle configuration by evaluating only a few candidates selected intelligently.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by the MSK Cancer Center Support Grant/Core Grant (P30 CA008748).