Abstract: Deep learning models for medical image segmentation often struggle with task-specific characteristics, limiting their generalization to unseen tasks with new anatomies, labels, or modalities ...