Improving Skin Lesion Segmentation via Stacked Adversarial Learning

Segmentation of skin lesions is an essential step in computer aided diagnosis (CAD) for the automated melanoma diagnosis. Recently, segmentation methods based on fully convolutional networks (FCNs) have achieved great success for general images. This success is primarily related to FCNs leveraging large labelled datasets to learn features that correspond to the shallow appearance and the deep semantics of the images. Such large labelled datasets, however, are usually not available for medical images. So researchers have used specific cost functions and post-processing algorithms to refine the coarse boundaries of the results to improve the FCN performance in skin lesion segmentation. These methods are heavily reliant on tuning many parameters and post-processing techniques. In this paper, we adopt the generative adversarial networks (GANs) given their inherent ability to produce consistent and realistic image features by using deep neural networks and adversarial learning concepts. We build upon the GAN with a novel stacked adversarial learning architecture such that skin lesion features can be learned, iteratively, in a class-specific manner. The outputs from our method are then added to the existing FCN training data, thus increasing the overall feature diversity. We evaluated our method on the ISIC 2017 skin lesion segmentation challenge dataset; we show that it is more accurate and robust when compared to the existing skin state-of-the-art methods.