Author: William T. Hrinivich, Junghoon Lee, Lina Mekki 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Department of Biomedical Engineering, Johns Hopkins University, Johns Hopkins University 🌍
Purpose: Volumetric modulated arc therapy (VMAT) planning is a computationally expensive process. In this work, we propose a reinforcement learning (RL) framework to automatically optimize dose rate and multi-leaf collimator (MLC) positions for dual arc prostate VMAT planning.
Methods: The proposed method consists of a multi-task convolution-based policy network including a joint encoder paired with two task-specific decoder paths trained to predict actions related to dose rate and MLC positions over the range of two arcs for prostate VMAT. For each gantry angle, the model takes as input a patient’s CT scan, planning target volume (PTV) and organs at risk contours (OARs), and machine parameters at all previous gantry angles, to predict the machine parameters at the next gantry angle. To capture the long-range dependencies between gantry angles over two arcs, a long-short term memory layer was added to the policy network architecture. VMAT plans for 15 test cases were generated using the proposed method and compared to plans directly optimized using a clinical treatment planning system (TPS).
Results: Preliminary results showed that the proposed framework resulted in a PTV coverage of 87.0±7.29 % compared to 97.2±1.4 % for TPS plans, and comparable dose to OARs, in 5.3±0.74 seconds. Additionally, fine-tuning the RL-generated plans in TPS improved the PTV coverage to clinical standard while reducing the maximum dose to the body, rectum, and bladder. Specifically, the average maximum doses over 15 test cases were 62.5±0.36 Gy, 62.1±0.57 Gy, and 62.3±0.38 Gy, respectively, for the RL fine-tuned plans compared to 64.6±1.31 Gy, 63.3±0.96 Gy, and 63.2±0.90 Gy for the TPS plans.
Conclusion: We demonstrated that optimization of two-arc VMAT machine parameters could be achieved using RL. Preliminary findings indicate that the proposed method provides near clinically acceptable plans, which can be refined to meet clinical goals with minimal fine-tuning in TPS.