[Protocol Generation for Contrast-enhanced CT Using Retrieval-augmented Generation with LLM].
Authors
Affiliations (1)
Affiliations (1)
- Department of Radiology, NHO Fukuokahigashi Medical Center.
Abstract
To evaluate how output format affects inference accuracy in contrast-enhanced CT protocol generation using a large language model (LLM) combined with retrieval-augmented generation (RAG) based on information obtained from a radiology information system. A total of 335 contrast-enhanced CT examinations with radiologist-determined protocols were retrospectively analyzed. Among them, 292 cases were used to build the RAG knowledge base, and 43 cases were used for validation. Two types of prompts were prepared: a structured format that outputs each imaging phase (unenhanced, arterial, portal venous, and equilibrium) in binary form (0/1), and a free-text format. The inference results generated by each format were evaluated. Complete agreement across all imaging phases was achieved in 16 of 43 cases (37.2%). Phase-specific accuracies were 76.2% for the unenhanced phase, 64.3% for the arterial phase, 78.6% for the portal venous phase, and 81.0% for the equilibrium phase. In the free-text evaluation, 65.1% of the generated contrast-enhanced CT protocols were rated as either "no issues" or "acceptable." An LLM integrated with RAG may serve as a potential decision-support tool for determining contrast-enhanced CT protocols.