How to improve the quality of deep learning masks?

How to improve the quality of instance segmentation annotations in projects that require deep learning?

You can do iteration or improve the image quality.