Autonomous Radiotherapy Planning Via Agentic Orchestration Using a Multimodal TPS-Integrated Compound AI Platform
Abstract
Purpose
Radiation therapy(RT) planning remains a manual, trial-and-error process that consumes significant clinical time and yields inconsistent results. We present a compound AI platform integrated with a commercial treatment-planning-system(TPS), combining autonomous planning orchestration with directive-conditioned 3D dose prediction. We hypothesize this site-agnostic system will match or surpass expert plan quality while reducing planning time from days to hours.
Methods
Our platform interfaces with a commercial TPS API, enabling autonomous operation ready for clinical deployment. It features a multimodal multi-agent architecture: (i)a plan review agent analyzing CT, dose, and dose-volume-histograms(DVHs) to identify emphasis points, (ii)a user-liaison agent integrating optional physician feedback, (iii)a critic agent evaluating proposals against DVH trajectories, and (iv)a planning agent executing plan modifications. Agents leverage GPT-5.2(HIPAA-compliant) with in-context learning, memory, prompt chaining, and tools to modify plans in an iterative loop. Four tool categories enable autonomy: Retrieval-Augmented-Generation(RAG) for querying plan data and institutional protocols, configuration tools for beam parameters and optimization objectives, execution tools for dose optimization/calculation, and plan analysis tools. Agents can also invoke dose prediction to guide initial objectives and accelerate trade-off exploration. The platform was evaluated on 15 cases (5 each: brain, lung, prostate) for 5 iterations each, and AI-generated plans were benchmarked against clinical plans.
Results
AI-generated plans achieved 87.1±6.9% of institutional organ-sparing criteria versus 82.8±8.6% for clinical plans. The automated feedback loop consistently refined suboptimal objectives across iterations, up from 82.0±6.3% at first iteration. Each iteration took 8.3±0.5 minutes, with AI contributing only 1.8±0.1 minutes.
Conclusion
This extensible platform demonstrates feasibility of fully autonomous RT treatment planning in a clinical environment. Dose prediction as an agent-invoked tool enables rapid trade-off exploration and informed objective initialization, improving convergence efficiency. This site-agnostic framework has potential to standardize plan quality, reduce planning burden, and accelerate adaptive workflows while supporting expansion to additional predictive models and tasks.