Federated learning (FL) offers a promising solution for collaborative medical image analysis without requiring the sharing of sensitive data between hospitals and research institutions. However, in practical scenarios, the local data available at each client is often non-i.i.d., with variations arising from factors such as geographic differences, patient demographics, and distinct data collection protocols. These challenges significantly affect the performance of federated learning. In this paper, we introduce FedPGD, a novel method for federated medical image analysis that leverages Projected Gradient Descent (PGD) to synthesize robust adversarial training samples. By generating these challenging samples aligned with both local and global models, FedPGD enhances the model’s generalization and robustness, ensuring improved performance across diverse data distributions. This method enables faster convergence and more effective adaptation to heterogeneous data. We validate our approach on a custom dataset consisting of real and phantom animal X-ray images with catheter and guidewire, demonstrating that while FedPGD achieves slightly better performance than state-of-the-art methods over extended communication rounds, it significantly accelerates model convergence in certain scenarios, making it a more efficient and practical approach for federated medical image analysis.