Early detection and accurate diagnosis for liver disease are very important. Due to the defects inherent in the ultrasound images and the complexity appearance of diseases, automatic classification for liver diseases in ultrasound images is a challenging task. In this paper, we introduce a novel method to classify focal liver lesions in ultrasound images. At first, we use an automatic image segmentation algorithm to delineate the lesion region. Then, according to the characteristics of liver lesions, we design a new image feature which is discriminative to liver lesions. Finally, six image features are processed by an improved sparse representation classifier to identify the diseases. We expand the sparse representation dictionary to optimize the classifier. Experimental results have shown that the proposed method could improve the classification accuracy in comparison with other state-of-the-art classifiers. It should be capable of assisting the physicians for liver disease diagnosis in the clinical practice. © Springer International Publishing AG 2017.