Machine Learning for Automated Hepatic Fat Quantification

H. Sagreiya, A. Akhbardeh, I. Durot, D.L. Rubin
University of Pennsylvania,
United States

Keywords: machine learning, fat quantification


Chronic liver disease due to nonalcoholic fatty liver disease (NAFLD) is a global health pandemic with increasing prevalence due to obesity. It affects over 30% of the population in the United States and Europe. Early diagnosis of NAFLD, which can be complicated by inflammation, cirrhosis, and hepatocellular carcinoma, is crucial, as it can be treated with weight loss, insulin-sensitization, and antioxidant agents. NAFLD has historically been diagnosed based on biopsy; an intracellular accumulation of fat droplets exceeding 5% of hepatocytes represents the clinical gold standard. Since limitations of biopsy include risks (bleeding, infection) and sampling error, noninvasive methods for diagnosing fatty liver, or hepatic steatosis, are warranted. Grayscale ultrasound and CT do not have great sensitivity and specificity, especially for mild steatosis, and do not provide reliable fat quantification. Magnetic resonance imaging with proton density fat fraction correlates highly with histology-derived steatosis grade, with high accuracy to quantify fat; while it has been adopted recently as a reference standard for clinical trials, it remains expensive and is not widely available worldwide. In fact, MRI is more than five times more expensive than ultrasound. Ultrasound elastography (USE) techniques noninvasively measure tissue stiffness and are used to manage patients with chronic liver disease. While ultrasound elastography has previously shown high accuracy for grading liver fibrosis, the effect of hepatic steatosis on USE measurements is not clear, and only a few studies have evaluated shear wave properties related to dispersion or viscoelastic models. More studies are warranted to evaluate the potential for quantifying hepatic steatosis by analyzing ultrasound properties such as shear wave velocity. Preliminary data showed that shear wave velocities can predict MR-derived hepatic fat quantification using machine learning in patient cohorts from Stanford and the University of Wisconsin. There was high correlation between the predicted fat quantification from ML and MR-based fat quantification (r=0.98). However, these techniques can be further validated on a much larger dataset. Ultrasound elastography requires specialized ultrasound scanners and trained technologists who manually select regions of interest. Traditional grayscale ultrasound is inexpensive, portable, real-time, and nearly ubiquitous. The hundreds of liver images obtained during a routine abdominal ultrasound, with their associated texture features, serve as a data-rich source to make this prediction. Machine learning has recently gained traction as a method to uncover subtle imaging features that can be leveraged in image recognition and classification, occasionally exceeding human performance. We hypothesize that grayscale ultrasound images can be used as inputs to a convolutional neural network to predict MR-determined fat quantification. There is significant interest from the pharmaceutical industry for inexpensive and reliable ways to serially monitor hepatic steatosis. A primary care physician or hepatologist could monitor hepatic fat in patients with a simple sweep of the ultrasound probe, revolutionizing the management of this disease, making quantitative hepatic fat assessment cheap and widely accessible, especially with the recent development of portable ultrasound probes such as the Butterfly iQ. In future, ultrasound could be used as a quantitative biomarker to monitor hepatic steatosis.