Style transfer for dermatological data augmentation

Abstract

In the field of medical image analysis, there is often a big problem with having too few data points for certain classes for which it is hard to collect samples and associated ground-truths that we need to have in order to do supervised learning. This can pose challenge since nowadays the best image-based computer aided diagnostics solutions use deep convolutional networks which are notorious for requiring large amounts of training data for proper functioning. To combat the problem of too few data points (not just in medicine but in deep learning in general), there have been multiple methods suggested for data augmentation from the simplest ones like random cropping and flipping to the more complicated ones like generating new images from white noise with Generative Adversarial Networks. In this paper, we have proposed a new augmentation method for skin lesion classification using style transfer and have performed some preliminary experiments to test its effectiveness.