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

Purpose.In high-dose-rate (HDR) prostate brachytherapy procedures, needle placement solely relies on physician experience. We investigated the feasibility of using reinforcement learning (RL) to provide needle positions and dwell times based on patient anatomy during pre-planning stage. This approach would reduce procedure time and ensure consistent plan quality.Materials And Methods.We trained an RL agent to adjust the position of one selected needle and all the dwell times on it to maximize a pre-defined reward function after observing the environment. The RL agent moves on to the next needle to adjust its location and dwell times until all needles are adjusted. Multiple rounds of process are played by the agent until the maximum number of rounds is reached. Plan data from 100 prostate HDR boost patients treated in our clinic were included in this study. Three RL models were trained on 5, 10 and 20 patients respectively, and the rest of 5 for evaluation, and 75 for testing. The dosimetry metrics and the number of used needles of RL plan were compared to those of the clinical results (ground truth).Results.On average, plans from all three models trained with different number of patients use 14 needles as clinical plans do, but reduces with statistical significance from clinical plans Rectum D2cc by 3.5%, Urethra D20% by 2%, and Prostate V150 by 5.5% at least when both plans are normalized to Prostate V100 = 95%. Meanwhile, all three RL plans trained on different number of patients have similar performance, with no statistically significant difference on Rectum D2cc and Urethra D20%.Conclusion.We present the first study demonstrating the feasibility of using RL to autonomously generate clinically acceptable HDR prostate brachytherapy plans. This RL-based method achieved equal or improved plan quality compared to conventional clinical approaches. With minimal data requirements and strong generalizability, this approach has substantial potential to standardize brachytherapy planning and reduce clinical variability.