Scholarlink

Federated learning (FL) offers a groundbreaking approach to collaborative medical image analysis by enabling model training across distributed healthcare institutions without sharing sensitive patient data. This paper presents a comprehensive FL framework designed for privacy-preserving diagnostic applications, integrating adaptive aggregation strategies, differential privacy mechanisms, and advanced transfer learning architectures such as EfficientNetV2 and ResNet-RS. The framework supports iterative parameter exchange between a central server and multiple clients, ensuring robust model performance while maintaining strict data confidentiality. Evaluated on diverse medical imaging tasks—including TB detection in chest X-rays, brain tumor segmentation in MRI, and diabetic retinopathy grading—the proposed approach achieves diagnostic accuracies comparable to centralized methods, with privacy guarantees quantified by differential privacy parameters (e.g., ε = 2.90). The system addresses key challenges such as data heterogeneity, security vulnerabilities, and regulatory compliance, providing adaptive client-specific adaptations and mitigation against gradient inversion attacks. This paper indicates and proves that federated learning is an effective and active solution to the development of global healthcare analytics without jeopardizing patient privacy and facing numerous regulatory requirements.

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 Privacy-Preserving Medical Image Analysis Across Global Healthcare Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Saurabh Yadav,
  • M. R. M. Veeramanickam,
  • Ravi Kumar,
  • Shubneet,
  • Anushka Raj Yadav,
  • Navjot Singh Talwandi

摘要

Federated learning (FL) offers a groundbreaking approach to collaborative medical image analysis by enabling model training across distributed healthcare institutions without sharing sensitive patient data. This paper presents a comprehensive FL framework designed for privacy-preserving diagnostic applications, integrating adaptive aggregation strategies, differential privacy mechanisms, and advanced transfer learning architectures such as EfficientNetV2 and ResNet-RS. The framework supports iterative parameter exchange between a central server and multiple clients, ensuring robust model performance while maintaining strict data confidentiality. Evaluated on diverse medical imaging tasks—including TB detection in chest X-rays, brain tumor segmentation in MRI, and diabetic retinopathy grading—the proposed approach achieves diagnostic accuracies comparable to centralized methods, with privacy guarantees quantified by differential privacy parameters (e.g., ε = 2.90). The system addresses key challenges such as data heterogeneity, security vulnerabilities, and regulatory compliance, providing adaptive client-specific adaptations and mitigation against gradient inversion attacks. This paper indicates and proves that federated learning is an effective and active solution to the development of global healthcare analytics without jeopardizing patient privacy and facing numerous regulatory requirements.