Federated Learning for Enhanced Medical Image Analysis

As artificial intelligence (AI) algorithms continue to advance, researchers have leveraged deep neural networks to address a range of challenges in the medical field. These models require a large-scale dataset and high-quality annotated data for model generalization, which is a major challenge in imaging data due to their limited availability in healthcare institutions. Additionally, it is primarily challenging to work with private patient data and share it with an external entity due to the privacy concerns. These challenges of the traditional centralized learning have led to a more efficient decentralized approach. This approach involves training with a diverse range of data from various domains, which are required to enhance model performance. Hence, many researchers have adopted Federated learning as an emerging paradigm to collaboratively train a machine learning model among multiple healthcare institutions without sharing their local private data. However, medical datasets are sourced from different medical institutions; hence they are often acquired by different protocols, scanner types, data modalities, and from different patient populations. Thus, it is inherently heterogeneous which degrades the global model performance in the federated setting. In this paper, we explore the key motivation for using federated learning in the healthcare field and discuss the challenges posed by the diversity and data heterogeneity of medical data from various institutions. Additionally, we present recent works that help mitigate the non-iid data issue in federated learning. Furthermore, we empirically evaluate the federated learning algorithms alongside centralized learning and one site learning using a benchmark medical dataset. Our analysis demonstrates that the adoption of advanced methods in FL enables us to effectively mitigate the data heterogeneity issue while leveraging data privacy and large-scale datasets within the medical domain.