Alpha-trimmed first order statistics for the classification of liver US images

US images are contaminated with speckle noise. To reduce the effect of speckle noise in the extraction of FOS, in this paper it is proposed to remove some of the pixels from both sides of the lower and higher pixel intensities arranged in order of intensity of an image, before the evaluation of FOS features. It is somewhat similar to alpha-trimmed filtering of images and hence in this paper this enhanced FOS method is called as alpha-trimmed FOS. To show the effectiveness of the proposed alpha-trimmed FOS method, features extracted from this method are used for the classification of liver ultrasound images and the classification accuracy obtained is compared with the accuracy obtained using the features extracted from normal FOS method. In the paper performance of proposed method is evaluated using two classifiers (Neural network and SVM). For alpha-trimmed FOS features, best accuracy obtained by neural network is 66.86% which is far better than the classification accuracy of 34.91% obtained with normal FOS features. Classification accuracy obtained by SVM with alpha-trimmed FOS features is 59.76% which is better as compared to accuracy of 53.85% obtained with normal FOS features.