Ultrasound liver steatosis diagnosis using deep convolutional neural networks

Abstract

One of the most common liver diseases is nonalcoholic hepatic steatosis. Until now, the standard method used for direct fatty liver quantification in hepatic tissue samples is liver biopsy. However, this method is invasive and involves certain risks for the patient. The goal of this paper is to find a non-invasive, cost-effective and wide available method for hepatic steatosis diagnosis that can replace the standard invasive procedure. The solution proposed is to use ultrasound images and deep convolutional neural networks. We implemented two models of deep convolutional neural networks and used 550 ultrasound images from 55 obese patients (only 17 with healthy liver) to train and test them. Our best model obtained an average accuracy of 87.49%.