ULTRASONIC LIVER IMAGE CLASSIFICATION

Abstract

Ultrasound (US) medical images are widely used for imaging soft tissue such as lungs, liver, rectum etc. for the diagnosis of abnormalities because of its non invasive technique & non deleterious effect. Image textures from different tissues have differing appearances, which can assist in clinical image interpretation. However image interpretation by humans is limited due to the search pattern of humans, the presence of noise in the image & the presentation of complex disease states requiring vast amount of clinical information. So a number of computer assisted diagnosis (CAD) techniques have been developed to provide a computer output which assists the radiologist's image reading. In the present study, the proposed CAD system includes two main steps (1) image feature extraction from the abnormal area of the image. (2) Classification of images between different types of abnormalities using extracted features. The five classes in the present study are Normal, Hepatocelular carcinoma (HCC), Metastases (MET), Hemangioma (HEM) & Cyst. Different texture based feature extraction techniques have been successfully implemented by various researchers for diagnosis of US liver images. Though, Fractal dimension (FD), Neighborhood Gray Tone difference matrix (NGTDM), Gray Level Difference Method (GLDM) based features have been used in C.T images of focal liver disease, these features have not been used for US focal liver disease analysis. So an attempt has been made in this study to find out the effectiveness of these methods for diagnosis of US focal liver images. The existing FD & NGTDM methods are found to be not so effective in case of US focal liver images, so modified FD & NGTDM methods have been proposed in this study which improves the diagnosis accuracy. Multiresolution FD features based on wavelet packet decomposition have been proposed here. Out of total 170 numbers of extracted features, 99 numbers of features found to be optimal. The multiclass Support Vector Machine (SVM) with radial basis kernel function have been used for classification. The classification accuracy of 62.72% has been found with the selected optimal features, but the proposed features combine with other features may increase the accuracy in classifying focal abnormalities.