Describir: Calibrating the Dice Loss to Handle Neural Network Overconfidence for Biomedical Image Segmentation.