Describir: Unsupervised cross-modality domain adaptation via source-domain labels guided contrastive learning for medical image segmentation.