Learning Fully Convolutional Networks on Sparsely Annotated Datasets