Classification of diffuse liver diseases based on ultrasound images with multimodal features

This paper describes a method of classification of diffuse liver diseases based on ultrasound images with multimodal features. The CNN network were used to extract image structure features, while Multi-scale Gray-level Co-occurrence Matrix (MGLCM) and Wavelet Multi-sub-bands Co-occurrence Matrix(WMCM) were used to extract image texture features. These two kinds of features were combined into multimodal features, then used as the input of a lightGBM classifier. 2942 liver ultrasound images have been classified using the proposed method. Classification accuracies of normal, fatty liver disease and liver fibrosis are 82.1%, 85.0%, and 80.9%, respectively. It can be seen from contrast experiments that the method proposed in this paper can improve the overall classification accuracy by 5.4%.