What question did this study set out to answer?

This research aims to improve brain tumor classification from MRI images while preserving patient privacy.

June 5, 2026Open Access

Federated MobileNetV2 with ensemble meta-learning for privacy-preserving brain tumor classification

Key Points

This research aims to improve brain tumor classification from MRI images while preserving patient privacy.
Utilized federated learning with MobileNetV2 as the feature extractor.
Trained local meta-learners (MLP, SVM, ELM) for personalized predictions without sharing data.
Evaluated on the Brain Tumor MRI Dataset with 7023 slices, ensuring privacy during training.
Achieved 99.57% maximum accuracy and 99.29% ± 0.20% accuracy across four runs, 95% CI 98.97% to 99.61%.
Macro-average ROC-AUC of 0.998690, indicating excellent model performance.
Required 149.89 MB per round for five clients, achieving 83.34% less communication than ResNet-50.

Abstract

Abstract The identification of brain tumors from MRI images is very crucial for the selection of an appropriate treatment. However, the existing solution has issues with privacy and data sharing. To address this challenge, this paper proposes the use of federated learning. The proposed solution employs a light convolutional backbone and some adaptive local meta-learners. The proposed solution employs MobileNetV2 as the feature extractor. This is fine-tuned for many clients using a combination of FedAvg and FedProx regularization. Each client also trains a few meta-learners (MLP, SVM, and ELM) using the local feature embeddings, enabling people to obtain personalized predictions without sharing their private information. For inference, the framework supports both probability-level averaging across client ensembles and deployable single-client prediction using only the local meta-learners of one client. On the Brain Tumor MRI Dataset, containing 7023 image slices across glioma, meningioma, no tumor, and pituitary classes, the proposed framework achieved a maximum observed accuracy of 99.57%. Across four repeated runs, it achieved 99.29% +/- 0.20% accuracy with a 95% confidence interval of 98.97% to 99.61%, while maintaining strong macro-F1 performance and a macro-average ROC-AUC of 0.998690. Under the same preprocessing and split protocol, it outperformed internally reimplemented CNN+FedAvg and CNN+FedAvg+FedProx baselines and preserved near-centralized ROC-AUC performance. Communication analysis showed that exchanging the MobileNetV2 backbone required 149.89 MB per round for five clients, corresponding to an 83.34% reduction relative to a ResNet-50 backbone.

Bookmark

View Full Paper