Bone marrow cell classification plays a crucial role in diagnosing hematological disorders such as leukemia, anemia, and multiple myeloma. However, the variability in imaging protocols, staining methods, and the distribution of data across multiple institutions present significant challenges for developing accurate and generalized models. Privacy concerns and data access limitations often hinder traditional centralized approaches, while federated learning (FL) offers a decentralized solution. However, conventional FL methods face limitations, including issues with data heterogeneity and insufficient feature diversity, which impact the model’s generalization and classification performance. In this paper, we propose Feature Exchange Federated Learning (FEFL), an innovative framework designed to enhance the learning capabilities of federated models. By incorporating Hierarchical Feature Exchange Layers (H-FEL), FEFL allows for the exchange of feature representations across three layers: shallow, intermediate, and deep. Rather than sharing raw model parameters or data, the proposed framework exchanges meaningful feature embeddings, improving the model's ability to generalize across heterogeneous datasets. This hierarchical feature exchange enables richer feature learning, mitigates overfitting, and accelerates convergence in federated training.