Interpretable and Annotation-Efficient Learning for Medical Image Computing