Learning joint segmentation of tissues and brain lesions from task-specific hetero-modal domain-shifted datasets.