Federated Learning for Multi-Disease Retinal Image Classification: A Systematic Review, Taxonomy, and Future Directions

Abstract

Retinal diseases are becoming more frequent, leading to increased demand for large and diverse datasets in applying deep learning techniques in automated diagnoses. The problem with data from different health centers is that, due to various privacy laws, ethical concerns, and limitations on institutional data sharing, it is usually difficult to combine retina images into one centralized system. Federated learning offers a solution by allowing various institutions to train a model in collaboration without merging their patients' data into a single center. The objective of this paper is to review methods for federated learning in the context of classification of retinal diseases based on the analyses of patterns of retina images. A systematic literary search and selection process has been undertaken in order to get the relevant researches published between 2018 and 2025 from the major scientific databases. The selected works are analyzed and classified in accordance with federated learning strategies, distribution of data used, disease coverage, models used, and data protection methods. In addition, a comparative analysis was also conducted in order to figure out major trends, advantages of the discussed approaches, and weaknesses of the existing studies. The current study exposes the weakness of federated retinal learning frameworks, which mostly work towards single-disease classification and base their analyses on the overly simplified claims about data. The issues of communication efficiency, privacy-utility trade-offs, and lack of large-scale clinical validation seriously challenge the application of the existing frameworks. After discussing the results of the review, the authors identify the most important gaps in the available research and develop an approach for deal with the existing issues effectively.