BI-RADS BERT and Using Section Segmentation to Understand Radiology Reports.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Radiology reports are crucial for patient care but exist as unstructured text.
• Extracting structured data from these reports is vital for research and automated care.
• Existing natural language processing (NLP) methods lack section segmentation for radiology reports.

Purpose of the Study:

• To develop and evaluate a BERT-based model for section segmentation of breast radiology reports.
• To assess the impact of section segmentation on the accuracy of clinical information extraction.
• To improve the analysis of breast radiology reports for research and patient care.

Main Methods:

• Pre-trained a BERT model on breast radiology reports.
• Developed a classifier integrating BERT embeddings with global textual features for section segmentation.
• Evaluated the model's accuracy in segregating reports into Breast Imaging Reporting and Data System (BI-RADS) sections.
• Assessed downstream extraction of clinical factors (modality, cancer history, etc.) using segmented reports.

Main Results:

• Achieved 98% accuracy in sentence-level section segmentation of free-text reports.
• Improved overall accuracy for clinical information extraction to 95.9% when using section segmentation.
• Demonstrated a significant 17% improvement in extraction accuracy compared to classic BERT without segmentation (78.9%).

Conclusions:

• BERT models, enhanced with section segmentation, significantly improve the analysis of breast radiology reports.
• Section segmentation is advantageous for accurately identifying key patient factors within reports.
• This approach strengthens the use of NLP for structured data extraction in medical imaging.