Dynamic weighted federated contrastive learning for robust medical image classification under heterogeneous data

With the rapid advancement of artificial intelligence technologies, AI-driven medical image classification has attracted considerable attention in recent years. However, increasingly stringent data privacy regulations have significantly restricted data sharing among medical institutions, making it difficult for traditional centralized learning frameworks to effectively exploit distributed medical data resources. As a privacy-preserving distributed machine learning paradigm, federated learning enables multiple institutions to collaboratively train models without directly exchanging raw data. Nevertheless, federated learning still faces several critical challenges in practical applications, including heterogeneous data distributions and model bias caused by non-independent and identically distributed (Non-IID) data. To address these challenges, this paper proposes a dynamic weighted learning method based on federated contrastive learning. Different from approaches that focus solely on local representation optimization or employ fixed aggregation weights, the proposed framework couples supervised contrastive local training with loss-feedback-based multi-factor client weighting, thereby jointly controlling local representation drift and server-side model aggregation. Specifically, contrastive learning is integrated into the federated learning framework by jointly leveraging local and global model representations for supervised contrastive optimization. This strategy effectively mitigates the distribution discrepancy among different clients and gradually aligns local models with the global model as communication rounds progress. Furthermore, during global aggregation, a loss-feedback rule is introduced to accept the candidate global model parameters and aggregation weights only when the candidate loss-feedback value decreases. In addition, a multi-factor weighted aggregation mechanism is introduced to comprehensively evaluate client contributions based on multiple factors, thereby assigning larger aggregation weights to local models that are more beneficial to global model optimization. Experiments were conducted on an osteoporosis knee X-ray dataset and a COVID–Pneumonia–Normal chest X-ray dataset under IID and different Non-IID settings. Under severe Non-IID conditions, the proposed method achieved accuracies of 92.12% and 96.27%, area under the receiver operating characteristic curve (AUC) values of 98.31% and 97.68%, and F1 scores of 90.92% and 93.22% on the two datasets, respectively. Comparisons with representative federated learning methods and the static-weighting ablation setting demonstrate that the proposed method maintains competitive classification performance and stable training behavior under heterogeneous client distributions.