Evaluation of Federated Learning Techniques on Edge Devices Using Synthetic Medical Imaging Datasets

Artificial intelligence (AI) holds great promise in healthcare as it can significantly advances disease diagnosis using diverse medical datasets. However, learning generalizable machine learning models requires data from multiple data providers and should not compromise patient privacy to abide by federal laws. Federated learning (FL) is an attractive approach of learning such models from different data providers without exchanging the actual data owned by the providers. In certain medical scenarios, there is limited labeled data to train models via FL. Therefore, in this paper, we explore the effectiveness of synthetic medical imaging datasets for training deep learning models using FL. We specifically focus on FL using edge devices that have limited computing and storage capacity. Our federated learning environment was implemented using the open source Flower framework and a network of edge devices. We selected a publicly available histopathology image dataset related to lung and colon cancer. We evaluated the performance of three well-known deep learning models, namely, DenseNet121, EfficientNetB0, and MobileNetV2 by training them on both real and synthetic images. We observed that DenseNet121 consistently performed the best. For example, it achieved an accuracy of 98% when trained on both real and synthetic images for colon cancer. Thus, synthetic data can be effectively used to train deep learning models via FL when real data are limited.