Enhancing Global Model Accuracy: Federated Learning for Imbalanced Medical Image Datasets

Federated learning (FL) is a deep-learning framework developed for cases where data privacy protection is essential due to security and privacy guidelines. Collaboration between independent institutions is critical to achieving better deep-learning model accuracy due to a single institution's shortage of information acquisition. This study focuses on medical image analysis of brain tumor dataset, using a federated learning framework. One of the drawbacks of FL is the global model accuracy impact due to the imbalanced distribution of dataset's samples and classes, called non-independent and identically distributed (non-IID). FL global model accuracy degradation can be optimized by various approaches, either by manipulating the dataset using augmentation techniques and data sharing or optimizing the global model aggregation algorithm. Our method uses two augmentation techniques: Generative adversarial network (GAN) and Synthetic Minority Over-Sampling Technique (SMOTE) to balance non-IID with the typical FedAvg aggregation algorithm to optimize global model accuracy. Global model accuracy using SMOTE and GAN outperforms IID and non-IID with a few numbers of global model iterations using local imbalance or globally imbalanced datasets.