A framework for biomedical figure segmentation towards image-based document retrieval.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Image Analysis

Background:

• Figures in biomedical publications are crucial for understanding biological experiments and results.
• Multimodal figures, common in bioscience, present challenges for automated analysis due to complex layouts and similar objects.
• Existing image segmentation techniques are insufficient for robustly partitioning these specialized figures.

Purpose of the Study:

• To develop a robust solution for automatically identifying and segmenting unimodal panels from multimodal biomedical figures.
• To improve the accuracy of document classification and retrieval tasks by effectively utilizing figure information.
• To provide an evidence source for derived assertions by accurately linking figure panels to their content.

Main Methods:

• A framework that harvests figure-caption pairs from biomedical articles, leveraging document layout to identify figures and boundaries within PDFs.
• Combines pixel-level image representations with caption information to estimate the number of panels and determine figure layout.
• Utilizes a web-based interface for efficient retrieval of figure panels based on user queries.

Main Results:

• The developed system achieved 96.64% accuracy in segmenting panels from figures in protein-protein interaction (PPI) documents.
• The automatic figure segmentation approach outperformed pure caption-based and image-based methods.
• Experimental results were validated against a gold standard annotated by biologists.

Conclusions:

• The proposed method offers a robust and accurate solution for segmenting multimodal biomedical figures.
• This technique enhances information retrieval and facilitates the integration of figure data into document analysis systems.
• The system provides a valuable tool for researchers to efficiently access and utilize visual information from scientific literature.