Figure 2.

UMAP projections of features learned from models trained with and without feature disentanglement on unmasked imagery. Each point represents a CXR from the COVIDX dataset. The top row colors points by their domain label – which subdataset of the COVIDx dataset they are in – while the bottom row colors points by their disease label. We observe that without feature disentanglement, the learned representations easily separate datasets – despite not being trained for this task – however, with feature disentanglement, the learned representations do not clearly separate datasets.