CXRAgent: Director-Orchestrated Multi-Stage Reasoning for Chest X-Ray Interpretation

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Radiology

Background:

• Chest X-ray (CXR) interpretation is crucial for clinical diagnosis.
• Existing AI models for CXR analysis struggle with adaptability and complex reasoning.
• Current LLM-based agents lack reliability assessment and diverse diagnostic pipelines.

Purpose of the Study:

• To introduce CXRAgent, a novel director-orchestrated, multi-stage agent for improved CXR interpretation.
• To enhance adaptability, credibility, and reliability in AI-driven CXR analysis.
• To enable robust multi-step reasoning and collaborative decision-making for diagnostic tasks.

Main Methods:

• CXRAgent employs a central director to orchestrate tool invocation, diagnostic planning, and collaborative decision-making.
• The Evidence-driven Validator (EDV) normalizes and verifies tool outputs, grounding diagnoses with visual evidence.
• An expert team is assembled and coordinated for adaptive, collaborative reasoning based on task requirements and intermediate findings.

Main Results:

• CXRAgent achieved 67.0% accuracy on CheXbench and 75.6% on Medical-CXR-VQA.
• The agent demonstrated strong performance with reliable visual grounding.
• A RaTEScore of 0.569 was attained on MIMIC-CXR for report generation.

Conclusions:

• CXRAgent offers a reliable and adaptable solution for complex CXR interpretation tasks.
• The multi-stage, director-orchestrated approach enhances diagnostic credibility through evidence grounding and collaborative reasoning.
• The proposed agent framework advances the capabilities of AI in medical imaging analysis.