Automatic figure classification in bioscience literature.

Area of Science:

• Biomedical informatics
• Computer vision
• Machine learning

Background:

• Millions of figures are published in biomedical literature annually.
• Effective retrieval of specific figures is crucial for research and knowledge discovery.
• Current search engines lack the ability to accurately classify and retrieve figures based on content.

Purpose of the Study:

• To develop an automated classifier for biomedical figures.
• To categorize figures into five distinct types: Gel-image, Image-of-thing, Graph, Model, and Mix.
• To enhance the performance of biomedical figure search engines.

Main Methods:

• Exploration of rich image features and integration with text features.
• Application of feature selection techniques to identify optimal features.
• Comparison of classification models: rule-based, supervised machine learning (Support Vector Machine - SVM), and a multi-model approach combining both.
• Development of a hierarchical rule-based classifier integrated with an SVM-based classifier.

Main Results:

• Feature selection significantly improved figure classification accuracy.
• Novel image features outperformed previously examined image features.
• The integration of text and image features yielded superior performance compared to using either feature type alone.
• The multi-model classifier, combining rule-based and SVM approaches, achieved the highest performance.

Conclusions:

• Automated classification of biomedical figures is feasible and beneficial for search.
• Combining diverse features (image and text) and classification models enhances accuracy.
• The developed multi-model classifier offers a robust solution for biomedical figure classification, achieving a 76.7% F1-score.