Generative adversarial networks to segment skin lesions

The accuracy of skin lesion segmentation has increased in recent years, thanks to advances in machine learning techniques and a large influx of dermoscopy images. However, there is still room for improvement as there exist many considerable challenges mainly due to the large variability in the appearance of lesions (i.e., shape, size, texture, and occlusions). In this work, we present a novel approach for skin lesion segmentation through leveraging generative adversarial networks. Our approach consists of two models: a fully convolutional neural network designed to synthesize an accurate skin lesion segmentation mask (the segmenter), and a convolutional neural network that distinguishes between synthetic and real segmentation masks (the critic). Our experimental results on 1300 images from the DermoFit dataset show that incorporating a critic network to complement a fully convolutional segmenter, like UNet, increases segmentation accuracy.