ImageCLEF 2009 medical image annotation task: PCTs for hierarchical multi-label classification

Abstract

In this paper, we describe an approach for the automatic medical image annotation task of the 2009 CLEF cross-language image retrieval campaign (ImageCLEF). This work is focused on the process of feature extraction from radiological images and hierarchical multi-label classification. To extract features from the images we used an edge histogram descriptor as global feature and SIFT histogram as local feature. These feature vectors were combined through simple concatenation in one feature vector with 2080 variables. With the combination of global and local features we want to tackle the problem of intra-class variability vs. inter-class similarity and the problem of data unbalance between train and test datasets. For classification we selected an extension of the predictive clustering trees (PCTs) able to handle data types organized in hierarchy. Furthermore, we constructed ensembles (Bagging and Random Forests) that use PCTs as base classifiers to improve the performance.