| Summary: | Data privacy and heterogeneity among healthcare settings present fundamental challenges to machine learning (ML) brain tumor classification (BTC) model development based on local data. In this paper, we outline the need to develop an effective brain tumor diagnosis model while ensuring data security. We propose a distributed learning framework based on Federated Learning (FL) to build such a system. This framework combines Federated Averaging (FedAvg) and Federated Proximal (FedProx) to train Convolutional Neural Networks (CNN) on data hosted by multiple clients. FedAvg saves these locally trained models from the patients to make a global model while keeping sensitive patient data decentralized and private. Nevertheless, traditional FL suffers from model instability and divergence due to the disparity of data distributions among clients. FedProx addresses this challenge by introducing a proximal regularization term that penalizes divergence from the global model to stabilize local training considering data heterogeneity. In this work, we utilized the Brain Tumor MRI Dataset from Kaggle and trained the model to classify four brain tumors (BT), namely glioma, meningioma, pituitary tumor, and no-tumor class. Experiments were conducted for ten communication rounds, with each client training the model independently based on its local dataset and without sharing data. Results revealed that the method consistently outperformed accuracy, recall and F1-score, further validating the capabilities of the proposed framework in effectively handling data heterogeneity while conserving data privacy. The global model outperformed models that had been centralized and trained on a smaller, common dataset observed in FL scenarios to improve the potential for collaborative medical diagnosis. This method is especially suitable for medical applications with strict privacy regulations, as it provides an approach to linking heterogeneous data from multiple healthcare institutions.
|
|---|
