Gamma Knife treatment planning using knowledge-based reinforcement learning.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Inverse planning in Gamma Knife radiosurgery uses manually tuned weights to achieve clinical objectives.
• Manual tuning is case-specific, increasing clinical workload and potentially affecting plan quality consistency.

Purpose of the Study:

• To train a deep reinforcement learning (DRL) agent to automate inverse planning.
• The DRL agent uses a reward function incorporating clinical metrics from historical plans to guide optimization.

Main Methods:

• A neural network-based DRL agent was developed to adjust inverse planning weights.
• The agent inputs current plan metrics, dose distribution, and target/organ-at-risk masks.
• The approach was validated on single-target metastases and acoustic neuroma datasets.

Main Results:

• The DRL agent achieved significantly higher plan scores on the metastases test set (p=0.0136).
• The agent also showed improved plan scores on the acoustic neuroma test set (p=0.4493).
• Agent-generated plans demonstrated greater similarity to clinical plans across key quality metrics.

Conclusions:

• The DRL agent successfully learned to generate plans consistent with historical clinical decisions.
• Future research will explore incorporating additional inputs to further enhance agent performance by explaining planning variability.