In-context learning for medical image segmentation