Automatic treatment planning using reinforcement learning for high-dose-rate prostate brachytherapy

Area of Science:

• Medical Physics
• Artificial Intelligence
• Oncology

Background:

• Physician experience is crucial for needle placement in high-dose-rate (HDR) prostate brachytherapy.
• Automating planning can reduce procedure time and ensure consistent quality.

Purpose of the Study:

• To investigate the feasibility of using reinforcement learning (RL) for autonomous needle positioning and dwell time optimization in HDR prostate brachytherapy pre-planning.
• To assess if RL can generate clinically acceptable treatment plans based on patient anatomy.

Main Methods:

• An RL agent was trained to optimize needle placement and dwell times by maximizing a reward function.
• The RL agent iteratively adjusted parameters for multiple needles across several training rounds.
• 100 patient plans were used for training (5, 10, 20 patients) and testing, comparing RL-generated plans to clinical ground truth.

Main Results:

• RL-generated plans used a similar number of needles (14) as clinical plans.
• Significant reductions were observed in Rectum D2cc (3.5%), Urethra D20% (2%), and Prostate V150 (5.5%) compared to clinical plans, when normalized to Prostate V100=95%.
• Performance across RL models trained on different patient numbers was comparable, showing no significant difference in key dosimetry metrics.

Conclusions:

• Reinforcement learning can autonomously generate clinically acceptable HDR prostate brachytherapy plans.
• The RL-based method achieved comparable or superior plan quality to conventional approaches.
• This AI-driven method requires minimal data, demonstrates strong generalizability, and has potential to standardize planning and reduce clinical variability.