Application of Deep Learning in Generating Structured Radiology Reports: A Transformer-Based Technique.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Radiology reports are often in free-text format, hindering clinical practice and research.
• Natural Language Processing (NLP) and Deep Learning (DL) offer solutions for automatic information extraction.
• Existing DL models like Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN) have limitations in clinical settings.

Purpose of the Study:

• To propose a transformer-based Named Entity Recognition (NER) architecture for clinical information extraction.
• To develop a method for transforming free-text radiology reports into structured data.
• To evaluate the performance of the proposed model against existing approaches.

Main Methods:

• Collected and annotated 88 abdominopelvic sonography reports.
• Developed a fine-grained NER architecture using transformers.
• Fine-tuned the Text-to-Text Transfer Transformer (T5) and Scisrc (a domain-specific adaptation of T5) models.
• Extracted entities and relations to structure the reports.

Main Results:

• The transformer-based model achieved superior performance compared to ANN and CNN models.
• Achieved ROUGE scores of 0.816 (ROUGE-1), 0.668 (ROUGE-2), 0.528 (ROUGE-L), and BLEU score of 0.743.
• The model successfully transformed free-text reports into an interpretable structured format.

Conclusions:

• Transformer-based models show significant potential for clinical information extraction from radiology reports.
• The proposed approach enhances the utility of unstructured clinical data.
• This method facilitates more efficient data analysis in radiology.