Fig. 2

Flow chart shows the steps of data analysis. We manually highlighted anatomic liver segments on portal venous phase CT scans of patients with chronic liver diseases. A three-dimensional texture analysis generated 1117 features out of each segment. The highly correlated features were filtered out from the dataset before normalization to the interquartile range. An unsupervised k-means and hierarchical clustering were performed with all segments. The univariate classification rate of the features for low-grade vs. high-grade fibrosis was tested in a receiver operating curve analysis. The cutoff at 9.5 kPa of shear-wave elastography was used as a reference. A machine learning pipeline was used to build models that could predict high-grade vs. low-grade fibrosis based on selected texture features. In the first analysis, the segments were randomly split between equal size train and test sets. In the second analysis, the segments scanned with a 64-slice scanner constituted the train set, and segments scanned with a 16-slice scanner were assigned to the test set. The models were optimized and validated on the corresponding training and the test sets, respectively