Deep Learning to Classify Radiology Free-Text Reports.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Informatics

Background:

• Pulmonary embolism (PE) is a critical diagnosis.
• Extracting PE findings from unstructured radiology reports is challenging.
• Automated analysis of thoracic computed tomography (CT) reports can improve efficiency.

Purpose of the Study:

• To compare a deep learning convolutional neural network (CNN) model with a traditional natural language processing (NLP) model.
• To evaluate their performance in extracting PE findings from thoracic CT reports.
• To assess performance across two institutions.

Main Methods:

• Utilized contrast-enhanced chest CT examinations from 1998-2016.
• Human radiologists annotated PE presence, chronicity, and location.
• Compared a CNN model with unsupervised learning to the PeFinder NLP application.
• Determined sensitivity, specificity, accuracy, and F1 scores for internal and external validation.

Main Results:

• The CNN model achieved 99% accuracy and an AUC of 0.97.
• For internal validation, the CNN model had a significantly higher F1 score (0.938 vs. 0.867) for PE classification.
• No significant differences in sensitivity, specificity, or accuracy were found internally.
• No statistical performance difference was observed between models for external validation.

Conclusions:

• Deep learning CNN models can classify free-text radiology reports effectively.
• CNN performance is equivalent to or surpasses traditional NLP models for PE extraction.
• Automated analysis holds promise for improving radiology report interpretation.