Empowering PET imaging reporting with retrieval-augmented large language models and reading reports database: a pilot...

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing in Healthcare
• Medical Imaging Analysis

Background:

• Large Language Models (LLMs) show potential for natural language tasks in clinical settings, including medical imaging reporting.
• Integrating LLMs with clinical data can enhance diagnostic accuracy and reporting efficiency.

Purpose of the Study:

• To evaluate a retrieval-augmented generation (RAG) LLM system for improving medical imaging reporting.
• To assess the efficacy of an LLM integrated with a PET report database for referencing prior cases and supporting decision-making.

Main Methods:

• A custom LLM framework with retrieval capabilities was developed using a decade-long database of PET imaging reports.
• Vector space embedding facilitated similarity-based retrieval for generating context-based answers and identifying similar cases or differential diagnoses.
• Experienced nuclear medicine physicians evaluated the system's performance on relevance and appropriateness scores.

Main Results:

• The RAG LLM system accurately clustered PET reports by diagnosis and study type within an embedded vector space.
• The system demonstrated potential in referencing similar past cases and identifying exemplary ones.
• 84.1% of retrieved cases were deemed relevant similar cases, and the RAG system significantly improved the appropriateness of suggested potential diagnoses compared to a non-RAG LLM.

Conclusions:

• Integrating RAG LLM with a PET report database can support nuclear medicine imaging reading by identifying similar cases and deriving potential diagnoses.
• Advanced AI tools, such as RAG LLM, have the potential to transform medical imaging reporting practices.