Classification and Grading of Liver Fibrosis Using Learnt Acoustic Nonlinearity Mapping

Abstract

Liver fibrosis is the liver's response to injury. Long-term liver fibrosis can develop into cirrhosis so early classification and grading of liver fibrosis are important. In this paper, we present a non-invasive machine learning based classification and grading of liver fibrosis using the acoustic nonlinearity parameter B/A, which is a physical parameter that varies between different materials and types of tissues. Instead of using conventional ultrasound B-mode images, which can vary with the different imaging instruments and parameters, we propose four new image functions derived from fundamental and the second harmonic signals of two input power levels, and use machine learning methods to estimate the entire functional dependence from the four sets of signals to fibrosis levels with the intermediate nonlinearity being implicitly learned. We test the method on three grades of rabbit fibrotic liver data using different classifiers: convolutional neural network (CNN), and the combination of CNN and multi-class support vector machine (MSVM) with two different kernel functions. Experimental results of 10-fold cross-validation show that the proposed method reaches a highest classification accuracy of 95.45% and an average overall classification accuracy of 85.19% for grading the stages of liver fibrosis by using the combination of CNN and MSVM with linear kernel function and without averaging features. Accordingly, high classification accuracy and easy implementation makes the prospect of using the proposed method for in-vivo tissue characterization of liver fibrosis possible