Figure 3
From: Deep Cytometry: Deep learning with Real-time Inference in Cell Sorting and Flow Cytometry
Receiver operating characteristics (ROC) and precision-recall (PR) curves for multi-class classification. (a) ROC curves, as evaluators of classifier output quality, are generated based on the predicted probabilities for the test dataset. They are shown for each class along with their macro-averaged and micro-averaged forms. ROC curves show the trade-off between classifier sensitivity and specificity, and an ideal ROC curve reaching the top left corner indicates both high sensitivity and specificity. Gray lines in the figure are the iso-BACC contour lines showing the BACC values from 0.5 to 0.9. (b) PR curve is another evaluator of classifier output quality, especially when dealing with the imbalanced classes. Precision is a measure of correctness, while recall, which is same as sensitivity, measures completeness. The PR curves for individual classes and their averaged forms are preferred to reach upper right corner, where both precision and recall are high. The iso-F1 contour lines show the F1 scores from 0.5 to 0.9. The shaded areas demonstrate the range of variations in each performance curve for five different training runs. The high sensitivity (recall), specificity, and precision regions of the ROC and PR curves are magnified in the insets for superior clarity.
