371 - Orchestrated Multi-Agent Large Language Model Framework for Automated Prostate Radiotherapy Planning

09:55am - 10:05am ET

Room 107

Presenter(s)

Andy Qin, PhD - Johns Hopkins Medicine, Baltimore, MD

K. Zhang¹, X. Jia², and A. Qin³; ¹Johns Hopkins Medicine, Baltimore, MD, United States, ²Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins Medicine, Baltimore, MD, ³Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD

Purpose/Objective(s): To quantitatively evaluate an orchestrated multi-agent large language model (LLM) framework for automated prostate radiotherapy planning. We hypothesized that a role-specialized multi-agent system deriving prescription and prioritization directly from structured diagnostic metadata would reproduce prescription-compliant target coverage and autonomously generate a deliverable treatment plan while revealing measurable dosimetric trade-offs in dose-escalated cases.

Materials/Methods:

A structured multi-agent architecture simulated multidisciplinary workflow using constrained prompting with structured JSON outputs to enforce role boundaries. Four LLM agents were implemented: (1) a Radiation Oncologist agent that ingests structured ProstateDiagnosis metadata and automatically generates prescription and OAR prioritization; (2) a Dosimetrist agent operating in dual modes: prospective strategy generation and DVH-based compliance assessment: while generating executable inverse-planning scripts with iterative optimization; (3) a Medical Physicist agent that evaluates dosimetric endpoints, verifies technical feasibility, enforces QA thresholds, and assesses IGRT requirements; and (4) a Coordinator that orchestrates execution, validates data integrity, and enforces two mandatory consensus checkpoints (RO–DOS and RO–MP). Unsafe scenarios (e.g., incomplete staging or unavailable prior RT data) triggered automated workflow termination.

The framework was validated across 11 diverse prostate cases spanning all risk groups from an open-source database. Prostate volume ranged from 27–141 cc (mean 53.5 ± 30.3 cc), with Gleason scores 6–8 representing low-, intermediate-, and high-risk disease. Prescription strata included 7800 cGy (n=7), 7000 cGy (n=1), and 6000 cGy (n=3). Quantitative endpoints were extracted from finalized plans. Correlation between prostate volume, prescription level, and rectal V75 was evaluated.

Results: PTV_High met prescription in 100% of cases. In 7800 cGy plans, mean PTV_High D98 was 7686 ± 124 cGy (98.5% ± 1.6% of prescription). Rectal V75Gy averaged 7.0 ± 2.3%, with 64% exceeding the 5% constraint. Rectal V75 correlated weakly with prostate volume (r=0.38) but more strongly with prescription level. Bladder violations occurred in <15% of cases, and small bowel high-dose thresholds were exceeded in 55% of escalated plans.

Conclusion:

An orchestrated multi-agent LLM framework translated structured diagnostic metadata into prescription, prioritization, and executable planning scripts while enforcing multidisciplinary review and safety logic. Dose escalation was associated with increased rectal V75, highlighting quantifiable optimization trade-offs in automated prostate radiotherapy planning.