Cue Integration for medical image annotation

Abstract

This paper presents the algorithms and results of our participation to the image annotation task of ImageCLEFmed 2007. We proposed a multi-cue approach where images are represented both by global and local descriptors. These cues are combined following two SVM-based strategies. The first algorithm, called Discriminative Accumulation Scheme (DAS), trains an SVM for each feature, and considers as output of each classifier the distance from the separating hyperplane. The final decision is taken on a linear combination of these distances. The second algorithm, that we call Multi Cue Kernel (MCK), uses a new Mercer kernel which can accept as input different features while keeping them separated. The DAS algorithm obtained a score of 29.9, which ranked fifth among all submissions. The MCK algorithm with the one-vs-all and with the one-vs-one multiclass extensions of SVM scored respectively 26.85 and 27.54. These runs ranked first and second among all submissions. © 2008 Springer-Verlag Berlin Heidelberg.