| 摘 要: 联邦学习是一种分布式训练方法,允许医疗机构在不交换数据的情况下联合训练模型,保护数据隐私。然而,由于不同地区疾病种类和数据分布的差异,数据通常呈现非独立同分布,影响全局模型的聚合,导致模型性能下降。为此,提出了一种面向医学图像的联邦学习混合权重聚合方法 HybridFed,该方法通过利用差分隐私的数据异质性和分层模型一致性,综合考虑客户端权重比例,更好地聚合全局模型。实验结果表明,HybridFed在3个医学图像数据集上优于其他方法,将准确率分别提升至93.61%、94.38%和74.27%,由此证明了该方法的有效性和可行性。 |
| 关键词: 联邦学习 医学图像 权重聚合 数据异构 |
|
中图分类号: TP399
文献标识码: A
|
|
| Hybrid Weight Aggregation Method for Federated Learning in Medical Imaging |
|
WANG Ruiqi, LI Feng, FAN Jiayi
|
(School of Computer Science and Communication Engineering, Jiangsu University, Zhenjiang 212013, China)
wangrq_0087@stmail.ujs.edu.cn; fengli@ujs.edu.cn; fanjy_8536@stmail.ujs.edu.cn
|
| Abstract: Federated Learning ( FL) is a distributed training approach that enables medical institutions to collaboratively train models without exchanging data, thereby preserving data privacy. However, due to regional
variations in disease types and data distributions, medical data often exhibit a non-independent and identically distributed (non-IID) nature. This heterogeneity impairs global model aggregation and degrades performance. To
address this, we propose HybridFed, a hybrid weight aggregation method for federated learning in medical imaging.HybridFed integrates differential privacy-based data heterogeneity measurement and hierarchical model consistency to
comprehensively weigh client contributions, achieving superior global model aggregation. Experimental results demonstrate that HybridFed outperforms existing methods across three medical image datasets, improving accuracy to
93.61% , 94.38% , and 74.27% , respectively. This validates the method’s effectiveness and feasibility. |
| Keywords: Federated Learning medical imaging weight aggregation data heterogeneity |
