A Survey on Recent Federated Learning Techniques in Healthcare: Challenges and Future Trends

Deep learning algorithms typically require access to large volumes of raw data. However, in the medical field, such data cannot be shared due to privacy constraints. Consequently, medical AI-based systems use Federated Learning (FL), which enables the processing and training of models on sensitive data without requiring it to leave local devices. In this paper, we outline the very recent developments of FL in the medical field, particularly in medical oncology, imaging, and drug discovery. We conclude that with the advantages of processing data in a federated system, FL does encounter the federation opacity issue. This latter causes training data to be opaque to the stakeholders, thus creating a lack of transparency that can lead to data poisoning, lack of accountability to institutions, and novel types of vulnerabilities. In addition, in this paper, we highlight the ethical issues that need to be resolved, in particular data heterogeneity, and the privacy-utility trade-off. Without an ethical approach that focuses on improving access to healthcare for all patients, we will not see the benefits of the technological advancements that are in the field of FL.