Feature statistic analysis of ultrasound images of liver cancer

Abstract

ABSTRACT In this paper, a specific feature analysis of liver ultrasound images including normal liver, liver cancer especially hepatocellular carcinoma (HCC) and other hepatopathy is discussed. According to the classification of hepatocellular carcinoma (HCC), primary carci noma is divided into four types. 15 features from single gray-level statistic, gray-level co-occurrence matrix (GLCM), and gray-level run-length matrix (GLRLM) are extracted. Experiments for the discrimination of each type of HCC, normal liver, fatty liver, angioma and hepatic abscess have been conducted. Corresponding features to potentially discriminate them are found. Keyword: ultrasound images, liver cancer, gray-level co-occurrence matrix, single gray-level statistic, gray-level run-length matrix, feature analysis 1. INTRODUCTION Liver cancer is located in the second position of malignant tumors. B-mode scanner is the first selection method of popular diagnostics. Ultrasound images has gained widespread acceptance as an effective diagnostic tool. Affected by the quality of ultrasound images of liver cancer, benign expression of malignant tumors and observer’s visual fatigue, careless mistakes, and diagnostic level, not all liver cancers can be detected accurately. Thus, it’s necessary to provide the medical operators a computer aided diagnosis system whic h is helpful to reduce the possibility of wrong diagnosis or missed diagnosis. Due to hepatocellular carcinoma (HCC) is at 90% of liver cancer, we mainly discriminate HCC, fatty liver, angioma, hepatic abscess and normal liver images. According to the ultrasound image repres entation of hepatocellular carcinoma (HCC) and echo type in the tumors, there are five types of primary carcinoma of liver, which are correspo nding to low echo type, equal echo type, high echo type, mixed echo type and diffuse type respectively. The shape of low echo carcinoma region is similarly a circle, and most of this type have a clearly edge, but echo distribution is not homogeneous, and echo is lower than normal region. Low echo halo is always around equal echo carcinoma region, and ech o distribution is not homogeneous. High echo carcinoma region is always large, and has clear edge, and echo in it is higher than in normal region. There are many type echoes inmixed echo carcinoma region, such as no echoes, high echo regions, and mixed low and high echo regions. Hepatocellular carcinoma’s focal region of this type is alwa ys large, and also has homogeneous echo distribution. Thelast type has numerous tumors diffused in the whole liver, an d can easily diagnose by operat ors themselves, so we do not discuss this type in this pape r. There are several benign tumors easily conf used with HCC, such as fatty liver, angioma, hepatic abscess, which we need to address. In this paper, we aim to find the most sensitive and stable groups of features by summing up those features of liver cancer showed on ultrasound images, mathematically analyze and compare them. However, features of ultrasound liver cancer images may have a wide range including the distribution of echo, text ure feature, or boundary information. It is difficult to analyze mathema tically. Spatial gray-level dependence matrices, Fourier power spectrum, gray-level