Metric Privacy in Federated Learning for Medical Imaging: Improving Convergence and Preventing Client Inference Attacks