The Evolution of Radiology Image Annotation in the Era of Large Language Models.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• High-quality, diverse medical imaging datasets with expert annotations are scarce for training AI models.
• Traditional manual annotation is time-consuming and repurposing existing annotations is impractical.
• Previous natural language processing (NLP) methods required custom model training for each use case.

Purpose of the Study:

• To review the evolution of medical image annotation and labeling processes.
• To highlight the potential of large language models (LLMs) in automating label generation.
• To present a scalable solution for efficient AI model training in medical imaging.

Main Methods:

• Review of traditional manual medical image annotation techniques.
• Exploration of semi-automated methods using natural language processing (NLP).
• Application of large language models (LLMs) with prompt engineering to extract labels from clinical radiology reports.

Main Results:

• LLMs enable the generation of accurate, normalized labels at scale directly from clinical reports.
• Combining automatically generated labels with foundation image models facilitates AI model training.
• Semi-automated methods offer a more efficient and scalable approach compared to manual annotation.

Conclusions:

• Large language models represent a significant advancement in overcoming data limitations for medical AI.
• Automated label generation from clinical reports streamlines the creation of robust AI models.
• This approach enhances the efficiency and scalability of developing AI for medical imaging analysis.