Results showing within-dataset class performance, within-dataset domain performance, and out-of-sample class performance from training models with the COVIDx dataset.

The task performance (task AUC and task accuracy) shows how well classifiers are able to distinguish between “Normal”, “Pneumonia”, and “COVID-19+” disease labels, while the domain performance (domain AUC) shows how well classifiers are able to distinguish which sub-dataset an image belongs to. We report AUC values as averages of the one-vs-all binary AUCs between all classes, and accuracy (ACC) as the average accuracy over all classes. In all cases class performance (both within-dataset and out-of-sample) is reported from the classifier trained on samples within-dataset, while domain performance is reported from an additional classifier trained to predict domain labels on top of the learned representations, z′, as a measure of how much domain information the representation contains. We observe that using feature disentanglement decreases within-dataset domain performance as expected, and increases out-of-sample class performance—i.e. improves generalization performance.