Abstract
PURPOSE: Artificial intelligence has emerged as a powerful technique for data analysis and predictive modeling. However, traditional centralized learning methods, which require aggregating large and diverse datasets at a central location, present considerable privacy and security risks, particularly in sensitive areas such as healthcare. Federated learning (FL) offers a promising alternative by enabling collaborative model training without the need to share raw data. We aim to systematically examine the current state of the art in the application of FL within the healthcare domain, with a focus on computational pathology. APPROACH: We conducted a systematic review of the published literature on FL in healthcare, with a specific focus on imaging-based applications relevant to computational pathology. Our analysis includes studies utilizing a range of medical imaging modalities such as whole-slide histopathology images, magnetic resonance imaging, computed tomography, and positron emission tomography. The selected studies were categorized based on a taxonomy of FL architectures, with a focus on understanding each study's motivations, implementation strategies, and targeted problems. We also evaluated recurring technical challenges such as system and data heterogeneity, privacy preservation mechanisms, and communication efficiency, as well as the integration of complementary technologies such as blockchain, homomorphic encryption, and multi-modal learning. RESULTS: The literature demonstrates a growing adoption of FL across healthcare applications, with increasing interest in computational pathology. Studies report promising outcomes for tasks such as patient outcome prediction, disease classification, and tissue segmentation using decentralized datasets. Notably, federated approaches often match or outperform centralized models in terms of accuracy while maintaining data privacy across institutions. In the case of computational pathology, federated training has proven feasible and effective. However, challenges persist across studies, including data modality heterogeneity, communication overhead, and slow model convergence. Several papers propose novel FL frameworks to address these issues, although standardization across implementations remains limited. CONCLUSIONS: FL holds significant promise for enabling secure, privacy-preserving collaboration in healthcare, particularly within computational pathology. The reviewed studies highlight the feasibility of applying FL across diverse data types without the need to centralize sensitive information. Nevertheless, key challenges such as system interoperability, data heterogeneity, and model interpretability continue to hinder real-world adoption. Future research should focus on developing scalable, standardized FL infrastructures, improving model robustness across heterogeneous sources, and addressing ethical concerns around fairness and accountability to support safe and effective clinical integration.