Large and diverse datasets are needed for the training of robust deep learning models in medical image classification. The authors of this paper explore federated learning to develop such collaborative models in the real world. Federated learning enables model training across decentralized devices holding local data without exchanging the data. Seven clinical institutions around the world collaborated on breast density classification using mammography images, despite large variations in mammography systems, class distribution, and dataset size. In the absence of centralizing data, we have successfully trained AI models within FL for image analysis tasks. We have found that, compared to the baseline trained on data only from one site, results across all testing data from other participating sites show a relative improvement in model generalizability of 45.8%. This large gain exemplifies the potential for FL to engender such methods for surmounting traditional limitations of centralized data and heterogeneity within medical imaging data toward the goal of realizing enhanced and collaborative diagnostics with AI at its core. Therefore, through the availability of such heterogeneous datasets and expertise, a more robust generalization across various medical imaging AI models can be achieved with high fidelity toward better diagnostic outcomes and, by extension, better patient care.