Fig. 1.

Diagram of the CycleGAN training loop and the novel process for generating unsupervised COVID-19 lesion segmentations. The model is composed of four networks: two generators and two discriminators. Beginning with an original image from the COVID domain, the COVID-to-Healthy generator converts it into the healthy domain. The image is then converted back into the COVID domain by the healthy-to-COVID generator, and it is compared to the original COVID image to calculate cycle consistency loss. Along the way, the original healthy and generated healthy images are given as input to the healthy discriminator, which attempts to correctly classify the image as original or generated and subsequently calculate adversarial loss. This process is then repeated with a real image from the healthy domain. After training, the COVID-to-Healthy generator is used to convert unseen COVID images to generated healthy equivalents. These are then subtracted from the original lesioned tissue to create a difference map, which can then be thresholded to produce a lesion segmentation.