Applications of Generative Adversarial Networks to Dermatologic Imaging

While standard dermatological images are relatively easy to take, the availability and public release of such data sets for machine learning is notoriously limited due to medical data legal constraints, availability of field experts for annotation, numerous and sometimes rare diseases, large variance of skin pigmentation or the presence of identifying factors such as fingerprints or tattoos. With these generic issues in mind, we explore the application of Generative Adversarial Networks (GANs) to three different types of images showing full hands, skin lesions, and varying degrees of eczema. A first model generates realistic images of all three types with a focus on the technical application of data augmentation. A perceptual study conducted with laypeople confirms that generated skin images cannot be distinguished from real data. Next, we propose models to add eczema lesions to healthy skin, respectively to remove eczema from patient skin using segmentation masks in a supervised learning setting. Such models allow to leverage existing unrelated skin pictures and enable non-technical applications, e.g. in aesthetic dermatology. Finally, we combine both models for eczema addition and removal in an entirely unsupervised process based on CycleGAN without relying on ground truth annotations anymore. The source code of our experiments is available on https://github.com/furgerf/GAN-for-dermatologic-imaging.