Ultrasonic liver steatosis quantification by a learning-based acoustic model from a novel shear wave sequence.

An efficient and accurate approach to quantify the steatosis extent of liver is important for clinical practice. For the purpose, we propose a specific designed ultrasound shear wave sequence to estimate ultrasonic and shear wave physical parameters. The utilization of the estimated quantitative parameters is then studied. Shear wave attenuation, shear wave absorption, elasticity, dispersion slope and echo attenuation were simultaneously estimated and quantified from the proposed novel shear wave sequence. Then, a regression tree model was utilized to learn the connection between the space represented by all the physical parameters and the liver fat proportion. MR mDIXON quantification was used as the ground truth for liver fat quantification. Our study included a total of 60 patients. Correlation coefficient (CC) with the ground truth were applied to mainly evaluate different methods for which the corresponding values were − 0.25, − 0.26, 0.028, 0.045, 0.46 and 0.83 for shear wave attenuation, shear wave absorption, elasticity, dispersion slope, echo attenuation and the learning-based model, respectively. The original parameters were extremely outperformed by the learning-based model for which the root mean square error for liver steatosis quantification is only 4.5% that is also state-of-the-art for ultrasound application in the related field. Although individual ultrasonic and shear wave parameters were not perfectly adequate for liver steatosis quantification, a promising result can be achieved by the proposed learning-based acoustic model based on them.