| Recently, building biological databases from online literature has attracted more attentions, which includes automating the collection, organization and analysis of biological data in the articles. The figures and captions in biomedical articles provide the important information about the crucial data and research results. Biological databases construction contains several steps, which include downloading online biological articles, collecting figure-caption pair from online biological articles, splitting figure-caption pair into panel-annotation pair, extracting image feature and text feature from panel-annotation pair, and identifying the types of panels. Finally, build a knowledge base system that can interpret images in online biological journals, and provide the common research platform and the effective search engine.In this paper, we want to build a magnetic resonance imaging database from online biomedical literature. Three critical processes were studied. Firstly, the recursive panel splitting, Gaussian mixture based magnetic resonance imaging detection, and morphologic operator based region growing methods were proposed to split figures into panels. Secondly, Gaussian mixture based character detection and optical character recognition were proposed to identify the panel serial label in figure. Thirdly, for better comprehensive utilization of image and text features, we proposed three fusion approaches, including merging both features, multiplying both posterior probabilities and co-training algorithm. The experiments show that panels and panel serial labels can be obtained precisely and automatically from the figures in the literature, and experimental results show a significant improvement in the average accuracy of the three fusion classifiers as compared with classifiers only based on image or text features. It provides a foundation for building a knowledge base system that can interpret magnetic resonance imaging images in online articles. |
PDF Full Text Request