Revisiting Generative Adversarial Networks for Binary Semantic Segmentation on Imbalanced Datasets