Dual-Space Pyramid Matching for Medical Image Classification

Abstract

AbstractWith the increasing of medical images that are routinely acquired in clinical practice, automatic medical image classification has become an important research topic recently. In this paper, we propose an efficient medical image classification algorithm, which works by mapping local image patches to multi-resolution histograms built both in feature space and image space and then matching sets of features though weighted histogram intersection. The matching produces a kernel function that satisfies Mercer’s condition, and a multi-class SVM classifier is then applied to classify the images. The dual-space pyramid matching scheme explores not only the distribution of local features in feature space but also their spatial layout in the images. Therefore, more accurate implicit correspondence is built between feature sets. We evaluate the proposed algorithm on the dataset for the automatic medical image annotation task of ImageCLEFmed 2005. It outperforms the best result of the campaign as well as the pyramid matchings that only perform in single space.KeywordsFeature VectorMedical ImageFeature SpaceLocal FeatureImage SpaceThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.