A Multi-Agent AI Framework for Intelligent, Knowledge-Based Quality Assurance In Online Adaptive Radiotherapy
Abstract
Purpose
The increasing adoption of MRI-guided online adaptive radiotherapy (oART) has improved treatment personalization but introduced substantial technical complexity and time pressure. Existing QA approaches remain largely manual and retrospective, with limited ability to ensure continuity between physician intent defined in the preplan and its realization during online adaptation. This study presents an intelligent multi-agent framework designed to accelerate MR-Linac plan QA while enabling context-aware adaption, providing a foundation for both preplan validation and extension to real-time adaptive workflows.
Methods
A LangGraph-based multi-agent workflow was engineered via five specialized agents: (1) Key Fact Extractor to extract data from Monaco internal files, (2) Structure Quality Analyzer (contours naming, density overrides, layering, accuracy, etc), (3) Parameter Validation Agent (auditing key plan parameters), (4)ART Agent (track fractional changes) and (5) Superior Checker (final verification). The system utilizes a global and site-specific reference rules as working memory with task-specific instructions embedded in each agent's prompt. For validation, a preliminary study of three mock prostate plans containing simulated errors were used. Quantitative performance was measured alongside qualitative scores from three experienced medical physicists, who rated the reports on clarity, readability, and overall satisfaction (1–5 scale).
Results
Out of total 47 simulated errors, the system identified 46, achieving 97.9% sensitivity and 93.9% precision (one missed minor deviation; three false positives). It generated structured, reproducible QA reports with a stable processing time of 45±5 seconds. Physicist evaluations confirmed high performance across report clarity (4.3±0.6), readability (4.2±0.3), and overall satisfaction (4.1±0.6).
Conclusion
This work demonstrates the feasibility of a multi-agent large language model framework for automated MR-Linac plan QA, achieving high accuracy, speed, and report consistency. While evaluated using preplans in this study, the framework is inherently extensible to online adaptive checks and human-AI co-piloting, positioning it as a scalable solution for next-generation, context-aware ART quality assurance.