Introducing soft topology constraints in deep learning-based segmentation using projected pooling loss.