A Fusion Based Federated Learning Approach for Multi Class Brain Tumor Classification with Enhanced Privacy

Brain Tumor (BT) leads to disability in cognitive, motor, and social skills, and therefore, early diagnosis should be a milestone for treatment. In this work, a novel Federated Learning-based Convolutional Neural Network (FL-CNN) model is proposed for Brain Tumor Classification (BTC) with FL serving as the framework for the model. The model is trained to distinguish between four classes of brain conditions: glioma, meningioma, pituitary adenoma, and non-neoplastic growth. Through the use of Federated Learning (FL), this method allows multiple Decentralized clients to cooperate in training the model without exchanging the raw medical data belonging to the patients. The provided dataset is derived from a training set containing 5707 images and a testing set containing 1311 images, and both sets are labeled among four categories. The fully trained 2D-CNN model deals with pre-processed MRI images in dimensions of 128×128 pixels and internalizes key attributes for identifying all forms of BT. As for understanding the model’s performance, we compute accuracy, precision, recall, and F1-score. The model achieved a peak validation accuracy of 97.48% with a precision, recall, and F1 score of 97.48%. Early stopping was applied at round 12 due to performance stagnation, preventing overfitting. The final global accuracy reached 97.48%, with a loss of 0.1483, demonstrating strong classification performance. The results exhibit that the federated strategy yields comparable classification accuracy with the conventional approach for distributed data and minimizes the violation of individual data privacy. Moreover, this work discusses the applicability of FL to medical image analysis, indicating that collaborative models in this area can provide a highly accurate performance while avoiding data aggregation. The following paper is intended to contribute to the improvement of privacy-preserving ML pertaining to medical diagnosis with regard to BTs.