Hierarchical annotation of medical images

Abstract

In this paper, we describe an approach for the automatic medical annotation task of the 2008 CLEF cross-language image retrieval campaign (ImageCLEF). The data comprise 12076 fully annotated images according to the IRMA code. This work is focused on the process of feature extraction from images and hierarchical multi-label classification. To extract features from the images we used a technique called: local distribution of edges. With this techniques each image was described with 80 variables. The goal of the classification task was to classify an image according to the IRMA code. The IRMA code is organized hierarchically. Hence, as classifer we selected an extension of the predictive clustering trees (PCTs) that is able to handle this type of data. Further more, we constructed ensembles (Bagging and Random Forests) that use PCTs as base classifiers.

ImageCLEF 2009 Medical Image Annotation Task: PCTs for Hierarchical Multi-Label Classification

“…For more information, see [10]. According to these experiments and previous research the ensembles of PCTs showed increased performance as compared to a single PCT when applied for hierarchical annotation of medical images [13]. Furthermore, the Bagging and Random Forest methods give similar results and because the Random Forest method is much faster than the Bagging method we submitted only the results for the Random Forest method.…”

Can Computer Vision Problems Benefit from Structured Hierarchical Classification?

“…This is also motivated by the superiority of the hierarchical classification approach in other fields such as text categorization and protein function prediction [7,23,28], where the features are typically high-level and where the possible states of the observed phenomenon are connected via well-understood hierarchical relationships. Enforcing such a hierarchical prior to the classification of visual content has been shown to be advantageous in some works (e.g., [3,6]), but far less often than in other fields [28]. Particularly in our previous work [11], the results of which are also reported in Sect.…”

“…As for classification based on visual features, a hierarchical prior intuitively seems especially appropriate as it reflects the natural way in which humans organize and recognize the objects they see, which is also supported by neurophysiological studies of the visual cortex [2,15,35]. In practice, some results have shown that there is indeed a gain in performance with the hierarchical approach in the visual-based application domain, e.g., for 3D object shape classification [3] and annotation of medical images [6]. A quite active and closely related line of work consists of the supervised construction of class hierarchies from images with multiple tag labels.…”

Automatic Detection of Biomedical Compound Figure using Bag of Words

“…Based on the results shown in Table 1, the Bag of Words (BoW) and Support Vector Machines (SVM) [8] are used as main image representation technique and classifier, respectively. In the past years, the BoW was successfully employed in various medical image retrieval and classification tasks [9][10][11][12][13][14][15][16][17]. Among other classifiers, SVMs have shown a better generalization performance in medical domain compared with other classification techniques.…”