Early Stopping Criteria for Training Generative Adversarial Networks in Biomedical Imaging