Use of a novel set of features based on texture anisotropy for identification of liver steatosis from ultrasound images: a simple method

Abstract

Detection of fatty liver disease (steatosis) from the ultra sound (US) images using pattern recognition techniques is attempted by many authors. Different pre-processing methods, feature extraction, feature selection and classification models are reported in the literature. The present work uses a hitherto unexplored property of liver texture. A careful visual observation reveals that the liver texture is anisotropic. A novel set of features exploiting this anisotropy is explicitly proposed. These anisotropy features are derived from gray level difference histogram (GLDH), pair correlation function (PCF), probabilistic local directionality statistics and randomness of texture (GLCM-χ8). For comparison with other features, the most popular gray level co-occurrence matrix (GLCM) derived features are also extracted. Accordingly, three alternative data sets are prepared to classify the images with five different classifiers –Bayesian, multilayer perceptron (MLP), probabilistic neural network (PNN), learning vector quantisation (LVQ) and support vector machine (SVM). A comparative evaluation in terms of specificity, sensitivity, discrimination score and accuracy has been made while classifying US images of human livers. On the basis of results this paper enumerates as to how the anisotropy feature provides better entity for classification purpose in the present context. It is also shown that the highest accuracy of 99% is obtained using anisotropy features with PNN. Anisotropy features leads to 100% sensitivity with PNN and SVM. The present classification system with anisotropy features outperforms the existing models available in the literature keeping in mind the simplicity of the algorithm.