DeTEXT: A Database for Evaluating Text Extraction from Biomedical Literature Figures

Area of Science:

• Biomedical Informatics
• Computer Vision
• Natural Language Processing

Background:

• Biomedical literature contains millions of figures with vital experimental evidence.
• Extracting text from these figures is crucial for information mining.
• Automated systems require high-quality ground truth data for development.

Purpose of the Study:

• To introduce DeTEXT, the first publicly available, human-annotated, large-scale dataset for evaluating text extraction from biomedical figures.
• To provide a robust benchmark for developing and assessing automated figure-text extraction systems.

Main Methods:

• Selection of 500 biomedical figures from 288 open-access full-text articles.
• Development of annotation guidelines and tools for human annotators.
• Annotation of 9308 text regions within the selected figures.
• Analysis of inter-annotator agreement and annotation reliability.

Main Results:

• Creation of DeTEXT, a comprehensive dataset comprising 288 articles, 500 figures, and 9308 annotated text regions.
• Demonstration of high-quality, human-annotated data suitable for training and evaluating text extraction models.
• Summary of dataset statistics and provision of evaluation protocols.

Conclusions:

• DeTEXT serves as a valuable resource for advancing automated text extraction from biomedical figures.
• Identified challenges and future research directions in figure-text detection and recognition were discussed.
• The DeTEXT database is publicly accessible for research use.