Computer-aided Quantification Analysis Of Fatty Liver Disease Based On B-scan Ultrasound Images

With the improvement of living conditions, fatty liver diseases have been increasing continuously. The main method to diagnose fatty liver is using B-scan ultrasound, which is often influenced by subjective factors and mainly depends on the experiences of doctors. Therefore, it will be helpful to enhance the accuracy and efficiency of clinical diagnosis by applying texture analysis theory into B-scan liver images to obtain the quantification feature characters and recognize these images.The specific steps of fatty liver computer-aided detection are as follows: selection of region of interest(ROI), feature extraction, classification recognition and disease severity quantification. In the first step, an interactive approach is proposed to select ROI. In the second step, combining with the characteristic of B-scan liver images, four methods of feature extraction have been chosen to differentiate normal and fatty liver images, including Near Field Echo Intensity(NFEI), Gray Level Co-occurrence Matrix(GLCM), Neighborhood Gray-Tone Difference Matrix(NGTDM) and Near Far Field Intensity Ratio(NFFIR). Making use of the statistical difference between normal and fatty liver group and the classification results of different feature combination, we finally set the best feature vector: NFEI, ASM of GLCM, NFFIR. In the part of classification recognition, the RBF kernel SVM classifier is designed by using the general SVM software development kits-LIBSVM. Two kinds of sample selection schemes have been adopted to train the classifier. With the well trained classifier, we classify the images into two groups, normal liver and fatty liver. At last, the similarity between the feature vector of a sample image and the standard feature vector is used to reflect the severity of a sample image. It achieves the goal of further analyzing the severity of fatty liver.The B-scan liver images are provided by the Sixth Hospital of Wuhan City. We choose 93 images from the dataset to carry out experiments. The classification rate for normal liver is 84% and for fatty liver is 97.1%. Good results on classifying mild and moderate fatty liver are also accomplished.