this study proposes a set of novel feature vectors for accurate differentiation of 3 typical types of liver spaceoccupying lesions in ultrasound images. Experiments were performed on 280 cases of liver images, including 112 cases of normal liver images, 90 cases of liver cancer images, 38 cases of liver hemangioma images and 40 cases of liver cyst images. First, we defined two types of region of interest and extracted a series of new features according to general image analysis and clinical diagnosis criteria. Second, the extracted features were roughly screened by U test and correlation analysis. The backwardremoval feature sequences were obtained by quadratic mutual information. Third, the suboptimum feature vectors were determined as input to the three-level back-propagation artificial neural network (BP ANN). Finally, the proposed BP ANN was evaluated on total 280 cases by means of ' leave-one-out ' methods. The precise differentiation rate of liver cancer, liver hemangioma, liver cyst and normal liver are 100%, 94.7%, 95% and 100%, respectively. The results indicate that the new defined features are useful to achieve high accurate differentiation of liver space-occupying lesions.
