Comparison of deep learning algorithms for quantitative diagnosis of fatty liver using multidimensional-moments heatmaps

Abstract

Non-alcoholic fatty liver disease (NAFLD) carries a high risk of progressing to non-alcoholic steatohepatitis (NASH), cirrhosis, and hepatocellular carcinoma. Therefore, the quantitative diagnosis of NAFLD, i.e., determining hepatic fat deposition, is desired during ultrasound inspection. It is known that the deposition of fat droplets within hepatocytes increases the attenuation in ultrasound images. Because the texture of the attenuated ultrasound image becomes a homogeneous speckle, the probability density function of the echo envelopes can be approximated by a Rayleigh distribution. The diagnosis methods of NAFLD using echo-envelope statistics of ultrasound images have been reported. In this study, we focus on the distribution and variance of echo-envelope statistics and propose a new diagnosis method to analysis the characteristics of echo-envelope statistics using a deep learning algorithm (DLA). In the proposed method, first- and third-order moments as echo-envelope statistics are calculated at each pixel in the ultrasound image. Then, the heatmap is formed from the distributions of both moments in each region of interest within the liver. The heatmaps are classified into the stages of fatty liver using DLA such as a convolutional neural network or a vision transformer. In this presentation, a comparative study on the accuracies by various DLA are reported.