Context-Driven Concept Annotation in Radiology Reports: Anatomical Phrase Labeling.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Radiologists frequently review prior patient history for comparative analysis during readings.
• Structuring radiology reports for efficient data access is crucial for Radiology Information Systems (RIS).
• Current methods for anatomical site labeling in reports struggle with linguistic variations among radiologists.

Purpose of the Study:

• To develop a novel framework for context-driven anatomical labeling in radiology reports.
• To address the limitations of conventional ontology-based keyword matching in anatomical phrase identification.
• To enhance the precision and recall of anatomical labeling in clinical text.

Main Methods:

• Proposed a novel framework utilizing two parallel Recurrent Neural Networks (RNNs).
• One RNN infers sentence context, while the other performs word (token)-level labeling.
• Trained the framework on a large dataset of radiology reports from a clinical site.

Main Results:

• The proposed framework demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements in correctly labeling ambiguous anatomical cases.
• Validated on radiology reports from two independent clinical sites.

Conclusions:

• The context-driven anatomical labeling framework offers a more robust solution for radiology report analysis.
• This approach effectively handles linguistic variability in anatomical descriptions.
• Potential to enhance the accuracy and efficiency of Radiology Information Systems.