Automatically pairing measured findings across narrative abdomen CT reports

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology

Background:

• Radiological measurements are crucial for oncology response assessment, guided by standards like WHO and RECIST.
• These measurements are often embedded in unstructured, free-text radiology reports.
• Automating the extraction and pairing of these measurements is essential for efficient data analysis.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) pipeline for extracting and pairing radiological measurements from narrative reports.
• To improve the standardization and objectivization of response assessment in oncology care.
• To facilitate data mining, advanced search, and workflow support for healthcare professionals.

Main Methods:

• A natural language processing pipeline was designed to extract measurements from radiology reports.
• The pipeline automatically pairs extracted measurements with those from prior reports for the same clinical finding.
• A Random Forest classifier was trained on 15 features using a manually created ground truth from 50 abdominal CT reports.

Main Results:

• The end-to-end evaluation demonstrated high performance: precision of 0.910, recall of 0.878, and an F-measure of 0.894.
• The Area Under the Curve (AUC) reached 0.988, indicating strong classification accuracy.
• Utilizing UMLS concepts did not enhance the pipeline's performance.

Conclusions:

• The developed NLP pipeline effectively extracts and pairs radiological measurements from free-text reports.
• This technology offers significant potential for improving data management in oncology care.
• Applications include enhanced data mining, search capabilities, and workflow optimization for clinicians.