Privacy-Preserving CNN Frameworks for Brain Tumor Diagnosis Using Federated Learning

Abstract

Diagnosis of brain tumors is complicated, considering issues of data privacy, imaging from multiple sources, and the demand for reliable automated tools. In this context, the Federated Learning (FL) has emerged as a privacy-preserving paradigm that enables collaborative training of Convolutional Neural Networks (CNNs) across institutions without sharing raw medical data. In this paper, this paper reviews recent works on FL in CNNs for the classification and segmentation of brain tumors. Indeed, there is much to be said about advances related to FedAvg-based ResNet-18, weight-sharing optimization frameworks, and personalized distillation approaches. Other strategies, such as the use of multi-encoders and privacy-preserving mechanisms, deserve consideration. Moving forward, related to data heterogeneity, communication costs, privacy risks, interpretability, and scalability, the discussed challenges present corresponding future directions. The analysis indicates that integrating elements of compression, robust privacy, explainability, and large-scale validation into one element will be relevant to pursuing FL research into clinical practice.