Automated Grading Method For Liver Fibrosis Based On Machine Learning Ultrasound Image

Liver fibrosis is a common chronic liver disease, which will turn to cirrhosis or liver cancer if patients have no timely diagnosis and treatment. Studies have shown that effective treatment can make the fibrosis reserved, so liver fibrosis staging has important clinical significance.B-mode Ultrasound imaging technology has wide applications in the field of liver disease staging because of the advantages of non-invasive, radiation-free and easy-to-operate. And texture feature of the liver is one of the important assessment indicators. However, in clinical applications, the rating results are highly dependent on the physicians' experience. So we propose a framework for staging and assessment the liver fibrosis automatically based on the machine learning technology, which learns the priori texture feature knowledge firstly to form the classification model for staging the liver fibrosis automatically.Feature extraction of liver parenchyma is based on the dense sample of DAISY descriptor and these features are then encoded by fisher vector. The feature selection and classification of the liver fibrosis are implemented by the sparse support vector machine(SVM) classifier, and leading to the final fibrosis staging results. DAISY descriptor, which extracts feature based on every pixel, has more abundant and smooth feature information because of the descriptor structure of multi-layer circular grids and isotropic Gaussian kernels of different scales. At the same time, DAISY is robust to image scale, rotation, change in illumination and noise. Fisher Vector(FV) fits the local descriptors by several Gaussian Mixture Models(GMM), forming the codebook using the parameters of the fit models. And the feature vectors can be given by the gradients of the log-likelihood. The texture features can be expressed more accurately by the way of model fitting and staging results can be optimal by adding higher order statistics. At last, the framework extracts features and classifies the liver fibrosis based on the sparse SVM.To evaluate the proposed method, 322 liver data from normal, fibrosis and cirrhosis patients are collected. And we get the classification accuracies of 89.10% between normal and abnormal patients and 91.64% between cirrhosis and non-cirrhosis patients based on the 10-fold cross validation.