Deep learning for assessing liver fibrosis based on acoustic nonlinearity maps: an in vivo study of rabbits

Abstract

Abstract This study aimed to assess liver fibrosis in rabbits by deep learning models based on acoustic nonlinearity maps. Injection of carbon tetrachloride was used to induce liver fibrosis. Acoustic nonlinearity maps, which were built by data of echo signals, were used as input data for deep learning model. Convolutional neural network (CNN), CNN combined with support vector machine (SVM), CNN combined with random forest and CNN combined with logistic regression were used as deep learning model. Nested 10-fold cross-validation was used to search hyperparameters and evaluate performance of models. Histologic examination of liver specimens of the rabbits was performed to evaluate the fibrosis stage. Receiver operator characteristic curve and area under curve (AUC) were used for estimating the probability of the correct prediction of liver fibrosis stages. A total of 600 acoustic nonlinearity maps were used. Model of CNN combined with SVM demonstrated the best diagnostic performance compared with all other methods for diagnosis of significant fibrosis (≥F2, AUC = 0.82), advanced fibrosis (≥F3, AUC = 0.88) and cirrhosis (F4, AUC = 0.90). Model of CNN showed the second highest AUCs. The deep learning model based on acoustic nonlinearity maps demonstrated potential for evaluation of liver fibrosis.