Learning to Summarize Chinese Radiology Findings With a Pre-Trained Encoder

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence in Healthcare

Background:

• Automatic summarization of radiology reports is crucial for reducing physician workload in computer-aided diagnosis.
• Existing English-based deep learning summarization methods are unsuitable for Chinese reports due to corpus limitations.

Purpose of the Study:

• To develop an effective abstractive summarization approach for Chinese chest radiology reports.
• To address the challenges posed by the lack of suitable Chinese medical corpora for existing summarization techniques.

Main Methods:

• Construction of a pre-training corpus using Chinese medical data and a fine-tuning corpus of Chinese chest radiology reports.
• Introduction of a Pseudo Summary Objective for task-oriented pre-training to enhance encoder initialization.
• Development and fine-tuning of a Chinese medical BERT (CMBERT) model for abstractive summarization.

Main Results:

• The proposed CMBERT-based approach demonstrated significant performance improvements over existing abstractive summarization models.
• The Pseudo Summary Objective effectively improved encoder initialization for the summarization task.
• The method proved effective in overcoming limitations of previous approaches for Chinese radiology report summarization.

Conclusions:

• The developed approach offers a promising solution for automatic Chinese chest radiology report summarization.
• This advancement can significantly alleviate the workload of physicians in computer-aided diagnosis.
• The CMBERT model represents a valuable tool for processing Chinese medical text data.