Reinforcement Learning for Radiotherapy Dose Fractioning Automation

Area of Science:

• Oncology
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• External beam radiotherapy is a cornerstone of cancer treatment, aiming to maximize tumor destruction while minimizing damage to surrounding healthy tissues.
• Current radiotherapy dose fractionation schedules are often standardized and may not be optimally tailored to individual tumor responses and tissue tolerances.
• Automating the optimization of complex treatment parameters like dose fractionation is a significant challenge in radiation oncology.

Purpose of the Study:

• To investigate the application of deep reinforcement learning (DRL) for automating and optimizing external beam radiotherapy dose fractionation schedules.
• To develop and evaluate DRL agents capable of learning optimal treatment strategies from simulated tumor and healthy cell interactions.
• To compare the efficacy of DRL-derived fractionation schedules against traditional constant dose per fraction approaches.

Main Methods:

• Utilized deep reinforcement learning algorithms, specifically deep Q network (DQN) and deep deterministic policy gradient (DDPG).
• Developed a 2D tumor growth simulation model to represent radiation effects on both cancerous and healthy tissues.
• Trained DRL agents within this simulated environment to learn and adapt dose fractionation strategies dynamically.

Main Results:

• The DRL agents learned to favor initiating radiotherapy with a higher dose per fraction.
• The optimal strategy identified involved a gradual reduction in dose per fraction throughout the treatment course.
• This adaptive fractionation approach demonstrated superiority over conventional constant dose per fraction schedules in simulations.

Conclusions:

• Deep reinforcement learning holds significant promise for automating and optimizing radiotherapy treatment planning.
• Adaptive dose fractionation, starting high and decreasing, represents a potentially more effective strategy than constant fractionation.
• Further research and clinical validation are warranted to translate these simulated findings into improved patient outcomes.