Figure 7.

Comparison of confusion matrices from the 2-class paradigm, 3-class paradigm, and adapted 2-class paradigm. The left panel summarizes the performance of the final DenseNet169 model trained on binary labels and evaluated on the heldout test set (normalized $\mu \pm \sigma$). The middle panel summarizes the performance of an equivalent DenseNet169 model trained on multi-class labels and evaluated on the heldout test set (normalized $\mu \pm \sigma$). The right panel is an adapted confusion matrix representing the same performance of the binary classifier in predicting the outcome of the available three-class labels for comparison.