Deep Reinforcement Learning-Based Automated Treatment Planning for a Rotating Gamma System.

Publication date: May 01, 2026

Introduction Traditional treatment planning for the rotating gamma system (RGS) is a complex and time-consuming process that relies heavily on planner experience. Manual selection of shot size, location, and weight requires extensive trial and error to achieve adequate target coverage (Cov) and dose conformity. Methods We developed an automated treatment planning algorithm for RGS using deep reinforcement learning (DRL). Shot placement was modeled as a sequential decision-making process based on target geometry and cumulative dose distribution. A proximal policy optimization (PPO) algorithm was implemented, using target dose Cov and conformal index as reward-related parameters. Thirty clinical brain tumor cases with target volumes ranging from 0. 5 to 45 cc were used for evaluation. An additional 15 cases that were not included in the training were used for independent validation. Results The DRL-based algorithm generated clinically acceptable treatment plans across a wide range of tumor sizes. Average dose Cov and selectivity were 92% and 0. 79 for tumors 20 cc. In the independent test cohort, average Cov and selectivity were 96% and 0. 81, respectively. Comparisons with clinical treatment plans created by an experienced planner demonstrated no statistically significant differences in Cov, selectivity, and conformity index (CI, Radiation Therapy Oncology Group (RTOG) criteria), although further improvement in conformity remains desirable. Plan generation time for a single target was less than 10 minutes. Conclusion The proposed DRL-based automated planning algorithm demonstrates the feasibility of generating RGS treatment plans. The approach achieves plan quality comparable to planner-driven methods while significantly reducing planning time and reliance on user expertise. These results highlight its strong potential to improve efficiency and consistency in stereotactic radiosurgery (SRS) planning.

Open Access PDF

Semantics

Original Article