An Interpretable Planning Bot for Pancreas Stereotactic Body Radiation Therapy

Area of Science:

• Radiation Oncology
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Pancreas stereotactic body radiation therapy (SBRT) treatment planning is a complex, time-intensive process.
• Optimizing dose distribution requires balancing target coverage with organ-at-risk sparing.
• Current manual planning involves sequential, iterative adjustments by human planners.

Purpose of the Study:

• To develop a reinforcement learning (RL)-based planning bot for pancreas SBRT.
• To systematically address complex tradeoffs in treatment planning.
• To achieve high plan quality consistently and efficiently.

Main Methods:

• Formulated planning interactions as a finite-horizon RL model.
• Defined planning states based on human experience and actions based on common planner steps.
• Developed a reward system guided by physician-assigned constraints; trained the bot on 48 augmented plans.

Main Results:

• Bot-generated plans achieved PTV coverage comparable to clinical plans while meeting all constraints.
• Learned knowledge was interpretable and consistent with human planning expertise.
• Training demonstrated reproducibility, with consistent knowledge maps across sessions.

Conclusions:

• An RL planning bot can generate high-quality pancreas SBRT treatment plans.
• The bot's training is tractable and reproducible, with interpretable acquired knowledge.
• The RL bot has the potential to be integrated into clinical workflows to improve efficiency.