Autonomous radiotherapy planning via agentic orchestration using a multimodal TPS-integrated compound AI platform

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiation Oncology

Background:

• Radiation therapy (RT) treatment planning is complex, requiring multi-day optimization.
• Subjective planning strategies lead to variability in plan quality.
• Current computational methods automate only parts of the workflow.

Purpose of the Study:

• To develop a compound artificial intelligence (AI) platform for fully autonomous RT treatment planning.
• To integrate multi-agent large language model (LLM) orchestration with 3D dose prediction.
• To create an end-to-end pipeline from physician directive to deliverable plan.

Main Methods:

• A seven-agent AI system navigated optimization using clinical reasoning and TPS modifications.
• Directive-conditioned 3D dose prediction autonomously derived initial optimization objectives.
• Retrieval-augmented generation (RAG) incorporated institutional knowledge.
• Evaluated 60 retrospective cases (brain, lung, prostate) for IMRT and VMAT plans.

Main Results:

• AI plans met 89.8% of dosimetric criteria vs. 85.2% for clinical plans (p < 0.001).
• IMRT plans improved in 25/30 cases (94.1% vs 84.3%, p < 0.001); VMAT showed no significant difference.
• Each autonomous plan iteration completed in 20.2 minutes.

Conclusions:

• Compound AI demonstrates feasibility for end-to-end, fully autonomous RT treatment planning.
• Integrated dose prediction resolved dependency on templates and manual input.
• This approach overcomes limitations of previous LLM-based planning systems.