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

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Traditional treatment planning for rotating gamma systems (RGS) is complex, time-consuming, and heavily dependent on planner experience.
• Manual optimization of shot parameters (size, location, weight) involves extensive trial and error to achieve target coverage and dose conformity.
• This process can lead to variability and inefficiency in stereotactic radiosurgery (SRS) planning.

Purpose of the Study:

• To develop and evaluate an automated treatment planning algorithm for RGS using deep reinforcement learning (DRL).
• To assess the feasibility and effectiveness of DRL in generating clinically acceptable RGS treatment plans.
• To compare the quality and efficiency of DRL-generated plans against traditional, planner-driven methods.

Main Methods:

• A deep reinforcement learning (DRL) algorithm, specifically proximal policy optimization (PPO), was developed for RGS treatment planning.
• Shot placement was modeled as a sequential decision-making process, incorporating target geometry and cumulative dose distribution.
• The algorithm was trained and validated on 45 clinical brain tumor cases, using target dose coverage (Cov) and conformal index as reward parameters.

Main Results:

• The DRL algorithm successfully generated clinically acceptable plans for brain tumors across various sizes (0.5-45 cc).
• Achieved average dose coverage (Cov) and selectivity metrics comparable to experienced planners, with plan generation time under 10 minutes.
• Independent validation demonstrated high Cov (96%) and selectivity (0.81), with no statistically significant difference in key planning metrics (Cov, selectivity, CI) compared to manual plans.

Conclusions:

• The DRL-based automated planning algorithm is feasible for RGS treatment planning.
• This approach offers comparable plan quality to manual methods while substantially reducing planning time and user dependency.
• The DRL algorithm shows significant potential to enhance efficiency and consistency in stereotactic radiosurgery (SRS) planning.