KneeTex: an ontology-driven system for information extraction from MRI reports

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Magnetic resonance imaging (MRI) is crucial for diagnosing knee pathologies and planning treatments.
• Clinical narratives in MRI reports contain valuable diagnostic information.
• Previous natural language processing (NLP) studies have not specifically addressed knee MRI reports due to anatomical complexity.

Purpose of the Study:

• To develop KneeTex, an information extraction system tailored for knee MRI reports.
• To leverage ontologies and NLP techniques for accurate information extraction in this domain.

Main Methods:

• KneeTex utilizes an ontology-driven approach, employing automatic term recognition to build a domain-specific ontology (TRAK).
• Sophisticated lexico-semantic rules and minimal syntactic analysis process the sublanguage of knee MRI reports.
• Key steps include named entity recognition, co-reference resolution, text segmentation, and ontology-based semantic typing for template filling.

Main Results:

• The TRAK ontology was expanded to include 1,621 concepts and 2,550 synonyms.
• KneeTex achieved high performance on a test set of 100 MRI reports, with precision of 98.00%, recall of 97.63%, and F-measure of 97.81%.
• Extraction results align with human-level performance.

Conclusions:

• KneeTex is an open-source application for extracting structured clinical findings from knee MRI reports.
• The system outputs findings as JavaScript Object Notation objects, mapped to the TRAK ontology.
• Structured data facilitates efficient searching for epidemiologic studies on knee conditions.