Development and multicenter validation of chest X-ray radiography interpretations based on natural language...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Natural Language Processing

Background:

• Interpreting large chest X-ray (CXR) datasets requires efficient image annotation.
• Artificial intelligence (AI) offers potential for automated CXR interpretation.

Purpose of the Study:

• To extract CXR labels from diagnostic reports using natural language processing (NLP).
• To train and evaluate convolutional neural networks (CNNs) for CXR classification using multi-center data.

Main Methods:

• Utilized 74,082 CXR images and reports for training.
• Employed BERT for entity and relationship extraction from reports, constructing a knowledge graph.
• Developed a 25-label classification system for weakly supervised CNN training.

Main Results:

• Achieved mean AUCs of 0.866, 0.891, and 0.796 in three external test cohorts.
• CNN performance was comparable to radiologists for most signs, with variations across patient groups.
• AI showed superior performance in identifying 6 signs in community clinic patients.

Conclusions:

• An effective CXR interpretation system was constructed and validated using NLP.
• The AI system shows promise for augmenting radiological interpretation and improving diagnostic efficiency.