Pixel N-Grams Representation for Medical Image Classification

Abstract

Image classification has wide applications in many fields including medical imaging. A major aspect of classification is to extract features that can correctly represent important variations in an image. Global image features commonly used for classification include Intensity Histograms, Haralick’s features based on Gray-level co-occurrence matrix, Local Binary Patterns and Gabor filters. A novel feature extraction and image representation technique ‘Pixel N-grams’ inspired from ‘Character N-grams’ concept in text categorization is described in this chapter. The classification performance of Pixel N-grams is tested on the various datasets including UIUC texture dataset, binary shapes dataset, miniMIAS dataset of mammography, and real-world high-resolution mammography dataset provided by an Australian radiology practice. The results are compared with other feature extraction techniques such as co-occurrence matrix features, intensity histogram, and bag of visual words. The results demonstrate promising classification accuracy in addition to reduced computational costs, enabling a new way for mammographic classification on low resource computers.