New Insights into Automatic Treatment Planning for Cancer Radiotherapy Using Explainable Artificial Intelligence

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Automatic treatment planning in radiotherapy relies on complex algorithms.
• Artificial intelligence (AI) agents, like Actor-Critic with Experience Replay (ACER), are used for optimizing treatment planning parameters (TPPs).
• Understanding the decision-making process of these AI agents is crucial for clinical adoption.

Purpose of the Study:

• To investigate the decision-making process of an AI agent used for automatic treatment planning in prostate cancer radiotherapy.
• To analyze how AI agents utilize dose-volume histogram (DVH) inputs to tune treatment planning parameters (TPPs).
• To assess the efficacy, efficiency, and stability of AI-driven treatment planning.

Main Methods:

• Examined a previously developed ACER network AI agent for automatic TPP tuning in intensity-modulated radiotherapy.
• Applied an explainable AI (EXAI) method to analyze attribution from DVH inputs to TPP-tuning decisions.
• Assessed planning efficacy, efficiency, policy space, and TPP tuning space of multiple AI agent checkpoints.

Main Results:

• ACER agents learned to identify dose violations from DVH inputs and mitigate them through TPP adjustments.
• Organ-wise similarity between DVH attributions and dose-violation reductions ranged from 0.25 to 0.5.
• Agents with stronger attribution-violation similarity showed improved planning efficiency, requiring fewer tuning steps and exhibiting more concentrated TPP-tuning spaces.

Conclusions:

• High-performing ACER agents effectively identify dose violations and employ global tuning strategies for high-quality treatment plans.
• The AI agent's learned TPP-tuning strategies resemble those of experienced human planners.
• Enhanced interpretability of AI decision-making can foster clinician trust and advance automatic treatment planning strategies.