Scholarlink

Diagnosing critical illness from x-ray or MRI images is a crucial and challenging task in medical field. Hospitals or data centers hesitate to give the medical data of their patients due to privacy concern. Federated learning allows us to do training in the edge devices without collecting the data to a centralized repository. However traditional FL techniques having limitations such as high communication cost and limited scalability. In this study, a Hierarchical Federated Learning (HFL) architecture that addresses those limitations is been provided. HFL uses a hierarchical structure with cluster-based communication, which improves scalability and lowers communication overhead. Using a COVID-19 dataset, we show that HFL surpasses standard FL models in disease prediction accuracy. The proposed model is implemented using different deep learning models (CNN, VGGNet, ResNet). The introduced HFL architecture achieves nearly same accuracy, Precision, Recall and f1-score as that of traditional federated learning systems while reducing the communication overhead.

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

A Hierarchical Federated Learning Architecture for Medical Diagnosis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • B. Neha,
  • Devika S. Menon,
  • R. Anagha,
  • Greeshma Sarath

摘要

Diagnosing critical illness from x-ray or MRI images is a crucial and challenging task in medical field. Hospitals or data centers hesitate to give the medical data of their patients due to privacy concern. Federated learning allows us to do training in the edge devices without collecting the data to a centralized repository. However traditional FL techniques having limitations such as high communication cost and limited scalability. In this study, a Hierarchical Federated Learning (HFL) architecture that addresses those limitations is been provided. HFL uses a hierarchical structure with cluster-based communication, which improves scalability and lowers communication overhead. Using a COVID-19 dataset, we show that HFL surpasses standard FL models in disease prediction accuracy. The proposed model is implemented using different deep learning models (CNN, VGGNet, ResNet). The introduced HFL architecture achieves nearly same accuracy, Precision, Recall and f1-score as that of traditional federated learning systems while reducing the communication overhead.