An enhanced CRFs-based system for information extraction from radiology reports

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Information extraction from free-text radiology reports is crucial for clinical applications.
• Supervised learning methods are employed to annotate text segments with radiological concepts.
• Challenges include handling text segments that do not align with sentence boundaries.

Purpose of the Study:

• To present two novel approaches for information extraction (IE) in radiology reports.
• To evaluate the effectiveness of a cascaded, two-stage IE method.
• To assess a confidence-weighted ensemble method combining standard and novel IE techniques.

Main Methods:

• Development of a cascaded, two-stage method using linear-chain conditional random fields (LC-CRFs).
• Implementation of a confidence-weighted ensemble method integrating standard LC-CRFs and the two-stage approach.
• Introduction and utilization of 'positional features' to aid automatic text annotation based on concept location.

Main Results:

• The proposed ensemble method demonstrated superior performance compared to a traditional single-stage CRFs system.
• Experiments were conducted on a dataset of mammography reports.
• The ensemble method showed significant improvements in two distinct, application-relevant scenarios.

Conclusions:

• The developed ensemble method offers an effective solution for information extraction from radiology reports.
• The incorporation of positional features enhances the accuracy of automatic text annotation.
• These advancements hold promise for improving the utility of free-text radiology data.