Two stage large language model approach enhancing entity classification and relationship mapping in radiology...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Radiology reports present linguistic variability, challenging automated analysis and medical image labeling.
• Accurate structuring of radiologic findings and temporal disease patterns is crucial for clinical decision-making.

Purpose of the Study:

• To develop and evaluate a two-stage natural language processing pipeline integrating Bidirectional Encoder Representations from Transformers (BERT) and large language models (LLMs).
• To improve the analysis of radiology reports for lesion-location mapping in chest CT and diagnosis-episode mapping in brain MRI.

Main Methods:

• A two-stage pipeline was developed: Entity Key Classification using BERT to identify and classify entities, followed by Relationship Mapping using an LLM to infer relationships between extracted entities.
• The pipeline was trained and validated on over 400,000 radiology reports from Seoul Asan Medical Center.

Main Results:

• The pipeline achieved a macro F1-score of 77.39 for chest CT lesion-location mapping.
• The pipeline achieved a macro F1-score of 70.58 for brain MRI diagnosis-episode mapping.
• Integration of BERT and LLM demonstrated effectiveness in enhancing diagnostic accuracy for radiology report analysis.

Conclusions:

• The proposed BERT and LLM integrated pipeline effectively analyzes radiology reports, improving the structuring of findings and temporal pattern capture.
• This approach holds significant potential for advancing medical image labeling and diagnostic accuracy in radiology.