AI Accelerated Human-in-the-loop Structuring of Radiology Reports

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Rule-based Natural Language Processing (NLP) requires extensive domain knowledge.
• Standard NLP methods often miss critical terminology in radiology reports.
• AI can accelerate lexicon development for structuring reports.

Purpose of the Study:

• To develop a domain-specific lexicon ontology for structuring radiology reports.
• To improve the accuracy of NLP in extracting information from Chest X-ray (CXR) reports.

Main Methods:

• Utilized AI techniques for concept expansion and phrasal grouping.
• Developed a domain-specific lexicon ontology using Chest X-ray (CXR) reports.
• Applied a simple NLP pipeline with the enhanced vocabulary.

Main Results:

• Achieved high recall (93.83%) and precision (94.87%) in extracting 83 abnormalities.
• Extracted 47 abnormality/normality descriptions for key anatomies.
• Identified additional label mentions and captured expert insights on descriptions and image quality.

Conclusions:

• A robust vocabulary significantly enhances NLP pipeline performance in radiology.
• The developed CXR ontology enables anatomically structured labeling of reports.
• AI-driven lexicon development is crucial for accurate medical report analysis.