ViVo: visual vocabulary construction for mining biomedical images

Given a large collection of medical images of several conditions and treatments, how can we succinctly describe the characteristics of each setting? For example, given a large collection of retinal images from several different experimental conditions (normal, detached, reattached, etc.), how can data mining help biologists focus on important regions in the images or on the differences between different experimental conditions? If the images were text documents, we could find the main terms and concepts for each condition by existing IR methods (e.g., tf/idf and LSI). We propose something analogous, but for the much more challenging case of an image collection: We propose to automatically develop a visual vocabulary by breaking images into n × n tiles and deriving key tiles ("ViVos") for each image and condition. We experiment with numerous domain-independent ways of extracting features from tiles (color histograms, textures, etc.), and several ways of choosing characteristic tiles (PCA, ICA). We perform experiments on two disparate biomedical datasets. The quantitative measure of success is classification accuracy: Our "ViVos" achieve high classification accuracy (up to 83 %for a nine-class problem on feline retinal images). More importantly, qualitatively, our "ViVos" do an excellent job as "visual vocabulary terms": they have biological meaning, as corroborated by domain experts; they help spot characteristic regions of images, exactly like text vocabulary terms do for documents; and they highlight the differences between pairs of images.