MedFed: A Federated Learning Framework for Medical Imaging

The medical imaging data gathered by the hospitals is not utilized to the fullest for research and analysis. The reason is strict Privacy policies for patients’ data. This is also limiting the potential applications of the available data. One of the Solution for the issue is Federated learning. Typically, when data availability is high but computational power is limited Federated learning concept is used, e.g. mobile environments on the contrary, ample of computational resources are available at various medical institutions but the required data is unavailable. This issue can be addressed by a federated learning framework. The proposed solution recommends to use Django and TensorFlow for building the framework which allows large-scale model training within the available hospital infrastructure yet, patient privacy is not required to be compromised. The proposed approach encourages the collaboration among medical institutions by utilizing their computational resources leading to model performance improvisation. This improved the diagnostic accuracy too. An accuracy of 83% is achieved by the global model after training for five epochs across five clients, surpassing the individual accuracy of all participating clients. Here the test data set is hosted on the server. This technique strengthens collaborative medical research as well provides a basis for broader federated learning applications where data privacy is very important. Keywords—Federated Learning, Data Privacy, data availability, medical imaging.