TEXTURE FEATURES FOR CLASSIFICATION OF ULTRASONIC LIVER IMAGES

In this paper, the classification of ultrasonic liver images is studied by making use of some powerful texture features, including the spatial gray-level dependence matrices, the Fourier power spectrum, the gray-level difference statistics, and the Laws' texture energy measures. Features of these types are used to classify three sets of ultrasonic liver images-normal liver, hepatoma, and cirrhosis (30 samples each). The Bayes classifier and the Hotelling trace criterion are employed to evaluate the performance of these features. From the viewpoint of speed and accuracy of classification, we have found that these features do not perform well enough, either consuming much time or yielding low classification rate. Hence, a new texture feature set (called multiresolution fractal features) based upon the concepts of multiple resolution imagery and fractional Brownian motion model is proposed to detect diffuse liver diseases fastly and accurately. In our approach, fractal dimensions estimated at various resolutions of the image are gathered to form the feature vector. Texture information contained in the proposed feature vector will be discussed. A real time implementation of our algorithm is performed on a SUN 4/330 workstation and produces about 90% correct classification for the three sets of ultrasonic liver images. This suggests that the multiresolution fractal feature set is an excellent tool in analyzing ultrasonic liver images.