A tandem reinforcement learning framework for localized prostate cancer treatment planning and machine parameter...

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Volumetric Modulated Arc Therapy (VMAT) machine parameter optimization (MPO) is computationally intensive.
• Existing machine learning methods often supplement, not replace, traditional optimizers and rely on extensive training data.
• Deep reinforcement learning (RL) offers a novel approach for VMAT MPO by learning optimal strategies through trial-and-error.

Purpose of the Study:

• To develop and validate a deep RL algorithm for automated VMAT MPO.
• To generate clinically comparable prostate cancer treatment plans independent of commercial treatment planning system (TPS) optimizers.
• To ensure generated plans meet machine constraints.

Main Methods:

• A dataset of 100 prostate cancer patients was used for training, validation, and testing (70-10-20 split).
• A Proximal Policy Optimization (PPO) algorithm trained two convolutional neural networks to optimize multi-leaf collimator (MLC) positions and monitor units (MUs).
• The RL networks used dose, structure masks, and machine parameters as inputs, optimizing a dose-volume histogram (DVH)-based reward function.

Main Results:

• The RL algorithm generated VMAT plans in an average of 6.3 ± 4.7 seconds.
• RL-generated plans showed improved bladder and rectum sparing compared to reference plans.
• Statistically significant improvements were observed in PTV D2% and reduced mean rectal dose (Dmean), with all clinical objectives met.

Conclusions:

• A deep RL framework was successfully developed and validated for VMAT MPO.
• The algorithm rapidly produces high-quality VMAT prostate cancer plans comparable to manual optimization.
• This demonstrates RL's potential to fully automate VMAT planning, reducing time while maintaining quality.