The purpose of this study is to use machine learning based methods to classify the liver fibrosis staging of chronic liver disease(CLD) using ultrasound images. This study has recruited 187 patients from Ditan Hospital. Liver biopsies were used as the gold standard. Two classification approaches are implemented in our work. The EfficientNet that is based on the conventional convolutional neural network (CNN) is used for classification. The second approach is a radiomics model. We investigated 637 radiomics features and the redundant features were reduced by the least absolute shrinkage and selection operator (LASSO). After reduction, fewer than 20 independent features are used for classifications. The area under the receiver operating characteristic (AUC) of EfficientNet model for cirrhosis (F4), advanced fibrosis (F3+F4), and significant fibrosis (F2+F3+F4) were 0.83, 0.78, 0.84, relatively. The AUC values of radiomics model for cirrhosis, advanced fibrosis, and significant fibrosis were 0.96, 0.81, 0.85, relatively. Machine learning methods can obviously classify liver fibrosis by CLD ultrasound image.

Liver fibrosis, Ultrasound images, Machine learning, Classification