Hepatosteatosis refers to the abnormal accumulation of triglycerides in more than 5% of hepatocytes (Lee DH,
2017; Torres,
Williams,
& Harrison,
2012).
Significant etiologies of hepatosteatosis include consumption of alcohol and Non-Alcoholic Fatty Liver Disease (NAFLD) (Miyoshi & Kamada,
2015).
NAFLD is a spectrum ranging from benign Isolated Fatty Liver (IFL) to Non-Alcoholic Steatohepatitis (NASH),
which involves hepatocellular inflammation and the potential to develop fibrosis,
cirrhosis and hepatocellular carcinoma (Torres et al,
2012).
NAFLD and NASH have prevalences of 46% and 12.2% respectively among asymptomatic middle aged Americans (Hamer,
Aguirre,
Casola,
Lavine,
Woenckhaus,
& Sirlin,
2006).
NASH is now the third most common indication for liver transplant in the U.S (Torres et al,
2012).
One-third to one-half of potential donor livers possess some level of steatosis,
which may result in primary graft non-function in liver transplants (Torres et al,
2012).
The clinical importance of early NAFLD diagnosis relates to the increasing prevalence of the condition,
the potential reversibility of the disease,
the possibility to prevent further hepatic damage and its implications in liver transplantation.
Currently a large proportion of hepatic steatosis may be under-reported due to limitations of diagnosing steatosis on single-phase contrast-enhanced abdominal CT scans.
Contrast-enhanced portal venous phase CT,
the most common CT abdomen exam type,
is inaccurate at detecting hepatic steatosis due to the potential for reading errors in attenuation values created by the contrast media inside the liver and spleen as well as by the variability in bolus timing (Lee SS & Park,
2014).
Unenhanced CT Abdomens can be accurate at detecting moderate hepatic steatosis but are less frequently used (Lee SS et al,
2014).
If a more objective and reliable process could be developed to detect and quantify hepatic steatosis when it appears incidentally on Contrast-enhanced venous phase CT,
NAFLD may be diagnosed earlier and more consistently in the general patient population,
allowing for earlier patient lifestyle or therapeutic intervention.
Contemporary advancements in machine learning and deep learning brought about a generation of deep convolutional neural networks (DCNN) that are capable of creating algorithmic models,
apt to perform many image analysis tasks that approach human-level vision abilities,
given sufficient training data and computation.
In addition,
the machine learning paradigm of utilizing transfer learning,
by fine-tuning a pre-trained CNN with generic images,
has shown considerable promise in producing neural networks that offer clinical utility,
with dataset sizes on the order of thousands of images.
These are within the realm of manual label annotation with manual or Natural Language Processing (NLP) supervised learning experiments.
Amongst the many neural network variants now publicly available,
there are 3D CNN’s capable of performing fully automated large solid organ segmentation (e.g.
Liver segmentation) as well as synthesizing multiphasic datasets (e.g.
Plain,
Arterial,
Portal Venous,
and Delay).
The incidental detection of hepatic steatosis is a novel application of 3D/2D DCNN methods that will permit the simultaneous (1) fully automated segmentation of the liver,
and (2) the fully automated classification and grading of steatotic liver disease from contrasted multiphasic CT examinations.
The goal of this study is to apply a series of DCNN’s to facilitate image recognition of hepatic steatosis on multiphasic CT examinations,
on both Plain and Portal Venous phase acquisitions,
utilizing both fully automatic liver organ segmentation,
and image classification to permit the automatic detection and quantification of hepatic steatosis on all routinely performed examinations.