Scholarlink

<p>Federated learning, as a privacy-preserving distributed learning paradigm, enables multiple medical institutions to collaboratively train deep models without sharing raw data. In recent years, it has garnered significant attention in medical image classification. However, substantial data heterogeneity across hospitals often leads to performance degradation in traditional federated learning. To address these challenges, this paper proposes a novel federated learning framework, FedGA, designed to enhance semantic consistency and global optimization under heterogeneous data distributions. Specifically, a trainable global prototype network (GPN) is introduced on the server side, enforcing consistency between clients’ local features and global prototypes via a prototype alignment loss. This promotes semantic collaboration across clients and stabilizes the shared feature space. In addition, a gradient-norm-based aggregation mechanism is devised, dynamically adjusting aggregation weights according to the L2 norms of gradients obtained during each client’s local training. This strategy improves overall optimization and mitigates the impact of non-representative client updates. Extensive experiments were conducted on two public medical image datasets, OCTMNIST and Dermoscopic, where FedGA achieved classification accuracies of 95.71% and 85.74%, respectively. The results demonstrate that FedGA outperforms eight state-of-the-art methods, highlighting the effectiveness and robustness of the trainable GPN and gradient-norm-based aggregation mechanism in addressing data heterogeneity.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated learning for medical image classification based on prototype alignment

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Mingjun Wei,
  • Rongyang Xu,
  • Qian Zhang,
  • Hui Li,
  • Jinyun Liu

摘要

Federated learning, as a privacy-preserving distributed learning paradigm, enables multiple medical institutions to collaboratively train deep models without sharing raw data. In recent years, it has garnered significant attention in medical image classification. However, substantial data heterogeneity across hospitals often leads to performance degradation in traditional federated learning. To address these challenges, this paper proposes a novel federated learning framework, FedGA, designed to enhance semantic consistency and global optimization under heterogeneous data distributions. Specifically, a trainable global prototype network (GPN) is introduced on the server side, enforcing consistency between clients’ local features and global prototypes via a prototype alignment loss. This promotes semantic collaboration across clients and stabilizes the shared feature space. In addition, a gradient-norm-based aggregation mechanism is devised, dynamically adjusting aggregation weights according to the L2 norms of gradients obtained during each client’s local training. This strategy improves overall optimization and mitigates the impact of non-representative client updates. Extensive experiments were conducted on two public medical image datasets, OCTMNIST and Dermoscopic, where FedGA achieved classification accuracies of 95.71% and 85.74%, respectively. The results demonstrate that FedGA outperforms eight state-of-the-art methods, highlighting the effectiveness and robustness of the trainable GPN and gradient-norm-based aggregation mechanism in addressing data heterogeneity.