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 is complex and often opaque.
• Artificial intelligence (AI) agents offer potential for optimizing treatment planning parameters (TPPs).
• Understanding AI decision-making is crucial for clinical trust and adoption.

Purpose of the Study:

• To elucidate the decision-making process of an AI agent for automatic treatment planning.
• To analyze how AI agents learn to tune treatment planning parameters (TPPs) for prostate cancer radiotherapy.
• To correlate AI's understanding of dose-volume histogram (DVH) inputs with TPP adjustments.

Main Methods:

• Examined actor-critic with experience replay (ACER) AI agents at various training stages.
• Applied explainable AI methods to analyze input-output relationships (DVH to TPP tuning).
• Assessed planning efficacy, efficiency, policy space, and TPP tuning space.

Main Results:

• ACER agents progressively learned to identify dose violations from DVH inputs and mitigate them.
• Agents with stronger attribution-violation similarity showed higher planning efficiency and stability.
• High-performing agents required fewer tuning steps and concentrated TPP adjustments.

Conclusions:

• High-performing AI agents effectively identify dose violations and employ global tuning strategies.
• AI agents demonstrate learning strategies comparable to experienced human planners.
• Improved interpretability of AI decision-making can enhance clinician trust and advance automatic treatment planning.