Large language models for structured reporting in radiology: past, present, and future

Area of Science:

• Artificial Intelligence in Radiology
• Natural Language Processing for Medical Reports
• Radiology Information Systems

Background:

• Structured reporting (SR) aims to standardize radiology reports for improved quality and error reduction.
• Widespread adoption of SR has been hindered by various challenges.
• Large language models (LLMs) offer a potential technological solution for automating and facilitating SR.

Purpose of the Study:

• To provide a narrative review of current literature on LLMs for SR in radiology and related fields.
• To explore the capabilities and limitations of LLMs in radiology report processing.

Main Methods:

• Review of existing studies focusing on LLMs (GPT-3.5, GPT-4, Perplexity, Bing Chat, IT5) for SR.
• Analysis of reported results regarding LLM performance in SR tasks.
• Identification of key themes including multilingual capabilities, limitations, and regulatory considerations.

Main Results:

• The current body of literature on LLMs for SR is limited but demonstrates promising outcomes.
• Several studies utilized generative pre-trained transformer (GPT) models, showing positive results.
• Multilingual applications of LLMs for SR were found to be feasible in multiple studies.

Conclusions:

• LLMs possess significant potential to enhance efficiency and accuracy in SR and overall radiology report processing.
• Overcoming limitations such as opaque algorithms and data concerns is crucial for clinical integration.
• The future role of LLMs in radiology hinges on addressing regulatory challenges and further research.