You are here
Biomedical image representation and classification using an entropy-weighted probabilistic concept feature space.
This paper presents a novel approach to biomedical image representation for classification by mapping image regions to local concepts and represent images in a weighted entropy based probabilistic feature space. In a heterogeneous collection of medical images, it is possible to identify specific local patches that are perceptually and/or semantically distinguishable. The variation of these patches is effectively modeled as local concepts based on their low-level features as inputs to a multi-class SVM classifier. The probability of occurrence of each concept in an image is measured by spreading and normalizing each region’s class confidence score based on the probabilistic output of the classifier. Furthermore, importance of concepts is measured as Shannon entropy based on pixel values of image patches and used to refine the feature vector to overcome the limitation of the “TF-IDF”- based weighting. In addition, to take the localization information of concepts into consideration, each image each segmented into five overlapping regions and local concept feature vectors are generated from those regions to finally obtain a combined semi-global feature vector. A systematic evaluation of image classification on two biomedical image data sets demonstrates improvement of more than 10% for the proposed feature representation approach compared to the commonly used low level and visual word-based approaches.