Learning steatosis staging with two-dimensional Convolutional Neural Networks: comparison of accuracy of clinical B-mode with a co-registered spectrogram representation of RF Data

Abstract

We explore the potential of ultrasound raw data as an alternative for non-alcoholic fatty liver disease (NAFLD) staging based on two-dimensional Convolutional Neural Networks (CNNs). Learning is performed on a stacked spectral representation of RF data and compared to co-registered clinical B-mode patches. Our initial results show a superior accuracy with RF data compared to B-mode. Early steatosis stages were classified more accurately than advanced steatosis stages.