From Diagnosis to Delivery: An Orchestrated Role-Specialized Multi-Agent LLM Framework for Automated Radiotherapy Planning
Abstract
Purpose
To explore the application of LLM agents for simulating clinical-team collaboration in prostate cancer radiotherapy treatment planning, integrating multidisciplinary expertise to automate end-to-end radiotherapy plan generation.
Methods
A multi-agent system automates prostate cancer radiotherapy planning through four agents with specialized skills:(1)the Radiation Oncologist (RO) handles clinical decisions and dose prescriptions;(2)the Dosimetrist (DOS) manages planning strategies and DVH compliance while producing RayStation scripts for plan generation,(3)the Medical Physicist (MP) evaluates feasibility and QA protocols, and(4) a coordinator orchestrates workflow and data verification. Each agent operates within non-overlapping skill sets defined by dedicated documents. Professional coordination is enforced through two mandatory checkpoints: an iterative RO–DOS review and a joint RO–MP clinical–technical assessment. The system was validated across 15 diverse cases spanning all risk groups from an open-source database, demonstrating robust handling of complex clinical scenarios.
Results
Patients with smaller prostate volumes achieved excellent organ-at-risk sparing, with rectal V7596%. For anatomically challenging cases with posterior extracapsular extension, systematic posterior target compromise to D98 of 93–95% was applied to preserve rectal constraints, demonstrating appropriate clinical prioritization. When geometric limitations resulted in rectal V75>5% or V70>20%, rectal spacer placement was identified as the primary mitigation strategy, with predicted V70 dose reductions of 5–7%. The automated workflow safety mechanism terminated processing in 13% of cases when critical staging data were unavailable or when post-radiation recurrence required specialized salvage therapy beyond the scope of standard planning agents.
Conclusion
This study demonstrates the feasibility of a multi-agent LLM system for simulating multidisciplinary collaboration in prostate cancer radiotherapy planning. The system integrates expert knowledge and automatically generates clinically acceptable treatment plans with stable dosimetric performance across risk groups. Built-in safety mechanisms enable early termination of unsafe workflows, highlighting effective risk stratification.