Learn the new, keep the old: Extending pretrained models with new anatomy and images