Figure Retrieval in Biomedical Literature

Abstract

Automatic classification of figures present in healthcare documents is known to be useful for biomedical document mining. The context of a document is directly reflected in the figures present within them. Embedded text within these figures along with image features have been used for figure retrieval. We demonstrate that image features based on structural properties of figures alone is sufficient for the figure retrieval task. An algorithm for describing structural properties of the embedded images, Fourier Edge Orientation Autocorrelogram, which utilizes spatial distribution of detected edges, is presented. We have shown that Fourier Edge Orientation Autocorrelogram performs better than its predecessor, when most of the edge information is retained. The algorithm is validated on publicly available figures from healthcare literature. Apart from invariance to scale, rotation and non-uniform illumination, the proposed feature descriptor is also shown to be relatively robust to noisy edges. Since there is no standard dataset available for figure classification, comparison of the proposed feature descriptor with four well known binary shape descriptors is demonstrated. The retrieval performance shows an overall improvement over other known methods in figure retrieval task.Keywords-algorithms; feature selection; pre-processing; intelligent information and multimedia systems; intelligent bioinformatics and biomedical systems.

Mining biomedical images towards valuable information retrieval in biomedical and life sciences

“… Mining pathway diagrams ( 44 ) Image and text mining Specific for pathways analysis images. www.biologicalnetworks.org Edge-based image feature descriptor ( 47 ) Image mining Open for all kinds of images health care images. Not publically available online.…”

“…Different machine learning techniques, Fourier descriptors ( 95 ) and neural network-based algorithms ( 96 ) have been proposed and implemented for the feature-based image processing and well applied in various fields of life, science and medicine. During our study we also found and discussed some domain specific and open bioimaging approaches ( 38 , 40 , 41 , 43 , 44 , 47 , 50 ) implementing feature extraction and OCR.…”

“…During this study, we went through different scientific NLP and IR approaches, which have been proposed for text and biomedical image mining e.g. ImageJ ( 30 ), CellProfiler ( 31 ), CellPloc ( 32 ), Vaa3D ( 33 ), Icy ( 34 ), Konstanz Information Miner ( KNIME ) ( 35 ), Fiji ( 36 , 37 ), Framework for the analysis and segmentation of protein-protein interactions (PPI) images ( 38 ), Automatic segmentation of subfigure image panels for multimodal biomedical document retrieval ( 39 ), Ontology based information retrieval from medical Images using Low Level feature extraction method ( 40 , 41 ), Parsing multi-panel collaged figures method for document image understanding ( 42 ), mining images for the detection and analysis of gel diagrams ( 43 ), bioimaging for complex networks and pathways analysis ( 44 ), automatic categorization and spatial distribution analysis of biomedical images ( 45 , 46 ), analysing the embedded structural properties of biomedical figures ( 47 ), Yale Image Finder ( YIF ) ( 48 ), integrating image data into biomedical text ( 49 ) etc. We also found some commercial applications (e.g.…”

“…The authors have validated this approach using different threshold values on 180 published biomedical figures (49 diagram, 40 plot, 42 biomedical images and 49 binary shapes) from 73 biomedical articles. They observed low performance rate at both moderate (30%) and high ( 50 ) threshold for all five methods (FEOAC, EOAC, EOH, HRT and R-SIG) ( 47 ).…”

“…This approach is about analysing embedded structural properties of four different classes (charts, plots, geometrical shapes) of figures and biomedical images from published biomedical literature, using retrieval-based pattern approach ( 47 ). Authors have analysed differences in structural information based on binary representation and utilized the intensity edge statistics of annotated, compressed and enhanced figures.…”