Improving Radiology Report Conciseness and Structure via Local Large Language Models.

Area of Science:

• Medical Imaging Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation Improvement

Background:

• Radiology reports are often lengthy and unstructured, hindering rapid identification of critical findings.
• This can lead to challenges for referring physicians and an increased risk of missed information.
• Efficient information extraction from clinical reports is crucial for effective patient care.

Purpose of the Study:

• To enhance radiology reports by improving conciseness and structure.
• To organize findings by relevant organs for better readability.
• To evaluate the efficacy of locally deployed large language models (LLMs) for this task.

Main Methods:

• A retrospective study utilizing 814 radiology reports from seven board-certified body radiologists.
• Implementation of private, locally deployed large language models (LLMs) within institutional firewalls.
• Testing of five prompting strategies using the LangChain framework, with evaluation of models including Mixtral and Llama.

Main Results:

• The Mixtral LLM demonstrated superior adherence to formatting requirements compared to other models.
• An optimal strategy involved report condensation followed by structured formatting, reducing verbosity.
• The Mixtral LLM reduced redundant word counts by over 53% across all reports and radiologists.

Conclusions:

• Locally deployed, open-source LLMs show significant potential for streamlining radiology reporting.
• Concise and well-structured reports generated by LLMs enhance information retrieval for referring physicians.
• This technology can improve clinical workflows and patient care by ensuring critical findings are easily accessible.