Fine-tuning ERNIE for chest abnormal imaging signs extraction

Area of Science:

• Medical informatics
• Natural Language Processing (NLP)
• Radiology

Background:

• Chest imaging reports are crucial for clinical research and medical tasks.
• Automated extraction of abnormal signs from these reports is vital but underexplored in Chinese.
• Existing methods face challenges with data insufficiency.

Purpose of the Study:

• To develop an effective method for extracting abnormal imaging signs from Chinese chest imaging reports.
• To address the challenge of limited data in this domain.
• To improve the accuracy and efficiency of information extraction for clinical applications.

Main Methods:

• Formulated abnormal imaging sign extraction as a sequence tagging and matching problem.
• Proposed EASON (fine-tuning ERNIE with CRF for Abnormal Signs ExtractiON), a model using pretrained ERNIE as a backbone to overcome data insufficiency.
• Developed a tag2relation algorithm to assign attributes (body part, degree) to extracted signs.

Main Results:

• EASON demonstrated significant and consistent improvement over baseline methods.
• The proposed method effectively addresses the problem of data insufficiency in Chinese chest imaging report analysis.
• The tag2relation algorithm successfully assigned relevant attributes to abnormal signs.

Conclusions:

• The developed EASON model provides a robust solution for extracting abnormal signs from Chinese chest imaging reports.
• This advancement facilitates clinical research and downstream medical tasks by improving information extraction.
• The study highlights the potential of transfer learning and novel algorithms in medical NLP.