Federated Learning (FL) enables decentralized model training without sharing sensitive data, but statistical heterogeneity introduces biases in BatchNorm (BN) layers, compromising model generalization. This work proposes a strategy for federated learning (FedSyncBN) that synchronizes BatchNorm statistics (mean and variance) during aggregation, combining diferent data sizes and class balances to compute adaptive weights for synchronization. A pre-trained EfficientNet-B0 model adapted for binary classification of 15600 chest X-rays from normal subjects and patients with vascular anomalies was used. A federated environment of 10 clients was simulated, including double heterogeneity: sample sizes distributed via Pareto (α = 2.5; range 240–2400 samples/client) and extreme class proportions generated via Dirichlet sampling (α = 0.52), reflecting real-world clinical disparities. The FedSyncBN was compared against FedAvg and FedProx using local and global metricsS. FedSyncBN strategy improved global generalization, achieving a centralized AUC-ROC of 0.9340 (+1.0% vs. FedAvg) and recall of 0.9400 (+3.5%), with a 20.5% reduction in false positives compared to FedAvg. FedProx showed better local diagnostic balance (F1-Score = 0.9310) but incurred higher false-negative rates (+44.8% vs. FedAvg), limiting its clinical applicability in critical scenarios. FedAvg maintained intermediate balance but poorer generalization in heterogeneous environments. FedSyncBN is proposed as a solution to mitigate BatchNorm statistical biases in FL with heterogeneous data, enhancing generalization and diagnostic balance in collaborative radiology models, and facilitating practical deployment in multicenter clinical settings.