Performance of the Gradient-Boosted Decision Tree (GBDT) model trained with Input C.

(A) displays the normalized confusion matrix. The greatest difficulty this model faced was in distinguishing early arterial phase (E-AP) from late arterial phase (L-AP) and late arterial phase (L-AP) from portal venous phase (PVP). (B) displays the precision-recall curves (PRC) using a One vs. Rest (OvR) approach. The OvR approach measures the ability of the model to distinguish that phase from all the other phases. For each pair of graphs, the top row displays histograms of the probability calculated by the model that a given scan is the target label. If that probability is greater than 0.5, then the model will classify it as the target label. Scans with the target label with probabilities greater than 0.5 and scans not of the target label (“Rest”) with probabilities less than 0.5 are correctly classified. The bottom row of graphs in each pair displays the PRCs, which graph the model’s precision (positive predictive value) against recall (sensitivity or true positive rate). A no-skill classifier is displayed as a horizontal line of the number of scans of the target label divided by the total number of scans. As shown by the PRC, the model has the most difficulty classifying E-AP scans.