Fig. 4.

Bar plot of the eight most important features for overall model performance as determined by the random forest model by assessment of Gini impurity decrease and recursive feature elimination. Feature importance has been normalized to the most important feature. The features, in order of descending importance are as follows: (1) gray-level co-occurrence matrix difference variance, (2) gray-level zone size matrix zone entropy, (3) gray-level co-occurrence matrix cluster tendency, (4) first-order entropy, (5) gray-level difference method dependence non-uniformity normalized, (6) gray-level zone size matrix large area low gray-level emphasis, (7) gray-level run-length matrix run-length non-uniformity, and (8) neighbourhood gray tone difference matrix busyness. Of note, features 1–5 are associated with image heterogeneity and only one (6) associated with the proportion of large zones with low gray values within the image