Feature analysis and automatic classification of B-mode ultrasound images of fatty liver

Abstract

Fatty liver refers to the pathological changes caused by excessive fat accumulation in hepatocytes with various causes. The medical analysis of B-mode ultrasound images(BMUI) is an important method for the identification of the degree of fatty liver. In this paper, a transfer learning method based on Densely Connected Convolutional Networks (DenseNet) and Light Gradient Boosting Machine (LightGBM) was proposed to automatically recognize the four classifications of BMUI of fatty liver. In comparison with the traditional gray and texture feature analysis, this method has the advantages of avoiding the manual extraction the region of interest of BMUI required by the traditional machine learning method for the analysis of feature performance. The complete BMUI were used to train the pathological feature classification model of fatty liver, and this method not only provides the comparison of ultrasonic echo intensity in the near field and far field of BMUI, but also retains the internal texture features and gray information of fatty liver. On the experimental data set, the classification model combined with DenseNet201 network and LightGBM proposed in this paper improved the recognition accuracy by 17.5% on average compared with the support vector machine and migration models based on ResNet101 network. In comparison with the two models above, the size was reduced by 99% on average, indicating remarkably improvement.