Brain Tumors Classification in MRIs Based on Personalized Federated Distillation Learning With Similarity-Preserving

Owing to legal restrictions and privacy preservation, it is impractical to consolidate medical data across multiple regions for model training, leading to difficulties in data sharing. Federated learning (FL) methods present a solution to this issue. However, traditional FL encounters difficulties in handling non-independent identically distributed (Non-IID) data, where the data distribution across clients is heterogeneous and not uniformly distributed. Although personalized federated learning (PFL) can tackle the Non-IID issue, it has drawbacks such as lower accuracy rates or high memory usage. Furthermore, knowledge-distillation-based PFL exhibits shortcomings in model learning capabilities. In this study, we propose FedSPD, a novel federated learning framework that integrates similarity-preserving knowledge distillation to bridge the gap between global knowledge and local models. FedSPD reduces discrepancies by aligning feature representations through cosine similarity at the feature level, enabling local models to assimilate global knowledge while preserving personalized characteristics. This approach enhances model performance in heterogeneous environments while mitigating privacy risks by sharing only averaged logits, in line with stringent medical data security requirements. Extensive experiments were conducted on three datasets: MNIST, CIFAR-10, and brain tumor MRI, comparing FedSPD with nine state-of-the-art FL and PFL algorithms. On general datasets, under the IID setting, FedSPD achieved performance comparable to existing methods. In Non-IID scenarios, we employed the Dirichlet distribution to control the data distribution across clients, allowing us to model and assess non-uniform data partitions in our FL settings. FedSPD demonstrated exceptional performance, with accuracy improvements of up to 77.77% over traditional FL methods and up to 4.19% over PFL methods. On the brain tumor MRI dataset, FedSPD outperformed most algorithms under the IID condition. In Non-IID settings, it exhibited even greater advantages, with accuracy improvements of up to 78.41% over traditional FL methods and up to 10.55% over PFL methods. Additionally, FedSPD significantly reduced computational overhead, shortening each training round by up to 67.25% compared to other PFL methods and reducing parameter size by up to 49.34%, thereby improving scalability and efficiency. By effectively integrating global and personalized features, FedSPD not only enhanced model generalization across heterogeneous medical datasets but also strengthened clinical decision-making, contributing to more accurate diagnoses and better patient prognosis. This scalable and privacy-preserving solution meets the practical demands of healthcare applications.