Clinical VMAT machine parameter optimization for localized prostate cancer using deep reinforcement learning.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Volumetric Modulated Arc Therapy (VMAT) machine parameter optimization (MPO) is computationally intensive and sensitive to dose objectives.
• Reinforcement learning (RL) offers a potential solution through machine learning via trial-and-error.

Purpose of the Study:

• To develop and evaluate an RL approach for VMAT MPO on a clinical linear accelerator (linac).
• To rapidly and automatically generate deliverable VMAT plans for localized prostate cancer.
• To compare the dosimetry of RL-generated plans with existing clinical plans.

Main Methods:

• Extended a previous RL approach using a policy network for VMAT MPO of a 3D beam model.
• Trained RL to minimize a dose-based cost function using data from 136 prostate cancer patients.
• Applied the trained RL VMAT to an independent cohort of 15 patients and compared dosimetry to clinical plans.
• Integrated RL with a clinical treatment planning system (TPS) for automated plan refinement.

Main Results:

• RL training produced 40,000 plans during exploration.
• Mean execution time for deliverable VMAT plans was 3.3 ± 0.5 seconds, with TPS refinement taking an additional 77.4 ± 5.8 seconds.
• RL+TPS plans achieved similar target coverage and overall maximum dose, with a significantly lower mean rectum dose (17.4 ± 7.4 Gy) compared to clinical plans (21.0 ± 6.0 Gy).

Conclusions:

• Developed and applied an RL approach for VMAT MPO on a clinical linac model.
• The RL VMAT approach, when combined with a clinical TPS, can rapidly generate high-quality plans for localized prostate cancer.
• This method shows potential for discovering advanced linac control policies through automated trial-and-error.