Federated Transfer Learning with Ensemble-Based Optimized Client Weighting for Binary Brain Tumor Detection from MRI Images

Abstract

Accurate detection of brain tumors from MRI scans is essential in clinical settings, yet training centralized models often faces obstacles due to privacy concerns and restrictions on data sharing. Federated learning (FL) offers a solution by enabling multiple centers to collaboratively develop models without sharing raw patient data. However, FL can encounter difficulties when client data is highly heterogeneous or imbalanced. In this study, we propose a federated transfer learning approach for binary brain tumor detection using MRI images from the Figshare dataset. Our method stands out by employing an ensemble-based strategy that assigns adaptive weights to each client’s contribution during model aggregation. We leverage a pre-trained backbone to extract meaningful features and estimate client-specific weights through an ensemble technique, which then guide the weighted averaging of parameters across three federated clients. Experimental results demonstrate that our approach outperforms conventional unweighted aggregation methods, achieving an accuracy of 96.33% along with improved F1-score, sensitivity, and specificity. These findings emphasize that optimized client weighting via ensemble methods can enhance the performance of federated transfer learning, providing a privacy-preserving solution well-suited for clinical decision support in brain tumor screening.