Classification of CT Figures in Biomedical Articles Based on Body Segments

Figures in biomedical articles provide important information that can be utilized to enrich user experience in biomedical article retrieval. One method to improve retrieval performance is to categorize figures into various modalities. We have previously used a hierarchical classification strategy that significantly improves retrieval performance. In this paper, we extend the hierarchy and add body segment classification, i.e., classifying the figures in CT (computed tomography) modality into different body segments, such as head, abdomen, pelvis, or thorax. To address the large variety of article images, we extracted a wide set of feature types (feature vector length of 2321) and applied a multi-class SVM classifier. Feature selection was applied to reduce the feature vector to length 50. Evaluation of the proposed method on a dataset consisting of 2465 figures from a subset of open access biomedical articles from the National Library of Medicine´s (NLM) PubMed Central® repository achieves classification accuracy of over 90%. This demonstrates its effectiveness and potential to become a vital component in biomedical document retrieval systems such as OpenI, a multimodal biomedical literature search system developed at NLM.