Novel Image Features for Categorizing Biomedical Images

Jianqiang Sheng,Songhua Xu,Weicai Deng,Xiaonan Luo

DOI: https://doi.org/10.1109/bibm.2012.6392689

2012-01-01

Abstract:Images embedded in biomedical publications are richly informative. For example, they often concisely summarize key hypotheses, illustrate new methods, and highlight major experimental findings in a research article. Prior studies [1] suggested that images embedded in biomedical publications offer effective clues for retrieving and mining their source documents. To facilitate accessing such valuable imagery resources, image categorization can be helpful. Like many other image processing tasks, extracting discriminative image features is critical for the success of image categorization. For biomedical images, we notice that many of them are embedded with abundant annotation text. Observing this property, we introduce a set of novel image features that exploit the spatial distribution of text information inside an image as essential clues for categorizing biomedical images. Through results of our evaluation experiments, this paper demonstrates the effectiveness of the proposed novel features - compared with conventional image features, our new features can help categorize biomedical images with superior performance using a standard supervised learning based approach.