Abstract
Early detection and frequent monitoring are critical for survival of skin cancer patients. Unfortunately, in practice a significant number of cases remain undetected until advanced stages, reducing the chances of survival. An appealing approach for early detection is to employ automated classification of dermoscopic images acquired via low-cost, smartphone-based hardware. By far, the most successful classification approaches on this task are based on deep learning. Unfortunately, most medical image classification tasks are unable to leverage the true potential of deep learning due to limited sizes of training datasets. Investigation of novel data generation techniques is thus an appealing option since it can enable us to augment our training data by a large number of synthetically generated examples. In this work, we investigate the possibility of obtaining realistic looking dermoscopic images via generative adversarial networks (GANs). These images are then employed to augment our existing training set in an effort to enhance the performance of a deep convolutional neural network on the skin lesion classification task. Results are compared with conventional data augmentation strategies and demonstrate that GAN based augmentation delivers significant performance gains.