End-to-End PET/CT Interpretation and Quantification with an LLM-Orchestrated AI Agent: A Real-World Pilot Study.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Deep learning enhances PET analysis but lacks end-to-end automation for clinical reporting.
• Real-world PET/CT interpretation faces challenges in heterogeneous data settings.

Purpose of the Study:

• To develop and evaluate an autonomous large language model (LLM)-orchestrated agent for end-to-end PET/CT interpretation.
• To assess the feasibility of automating the workflow from raw DICOM to quantitative clinical reports.

Main Methods:

• An LLM agent was designed to autonomously select DICOM series, perform registration, SUV conversion, segmentation, detection, and report generation.
• The system utilized a reasoning-based LLM and a vision-enabled LLM for interpretation.
• Retrospective evaluation involved 170 lung cancer patients undergoing baseline [18F]FDG PET/CT.

Main Results:

• The agent successfully automated the entire workflow for all 170 cases without human intervention.
• Primary tumor detection achieved 100% sensitivity.
• Nodal and distant metastasis detection showed variable sensitivity and specificity, with identified systematic false-positive and false-negative findings.

Conclusions:

• LLM-orchestrated agents can automate PET/CT workflows from raw data to draft reports in real-world settings.
• While primary tumor detection is reliable, nodal and metastatic assessment requires further refinement and expert oversight.