Scholarlink

<p>Federated learning (FL) combined with privacy-preserving techniques provides a potential solution to the problems of collaborative neuroimaging analysis among multiple institutions, initial critical challenges of privacy, heterogeneity, and scalability of centralized solutions. This systematic review surveys the neurological imaging data distinct features of high dimensionality, clinical sensitivity, and inter-institutional variability that require the application of FL with privacy-preserving methods protocols specifically to medical imaging in general. We survey the principles, threat models, and privacy-protective mechanisms specific to neuroimaging tasks, which point to the progress in Differential Privacy, Homomorphic Encryption, and Secure Multi-Party Computation. A multi-axis taxonomy of FL architectures, privacy strategies and neuroimaging settings identifies unexplored areas and gaps in the methodology. We evaluate domain adaptation policies, resistance to the heterogeneity of data, and the effects of privacy assurances on clinical informatics through a critical review of applications in Alzheimer disease, brain tumor analysis, lesion segmentation, and multiple sclerosis. We also address the technical issues, such as statistical heterogeneity, paucity of annotation, efficiency of communication, interpretability, fairness, and ethical implementation. New directions like personalized FL, foundation model pretraining, synthetic data generation, and benchmarking efforts are investigated. Lastly, we find open questions and roadmap milestones that are needed to scale, standardize, align governance, multimodal integrate, and translate neuroimaging FL systems to clinical practice. The purpose of this review is to notify researchers and clinicians on how to build the powerful, privacy-assuring, and clinically feasible federated neuroimaging solutions.</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 and privacy-preserving learning in neurological imaging: a systematic review of frameworks, applications, and evidence gaps

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Faisal Sajjad,
  • Rana M. Tahir,
  • Wei Qi Yan,
  • Wei Zhiyuan,
  • Yang Xia Zhou

摘要

Federated learning (FL) combined with privacy-preserving techniques provides a potential solution to the problems of collaborative neuroimaging analysis among multiple institutions, initial critical challenges of privacy, heterogeneity, and scalability of centralized solutions. This systematic review surveys the neurological imaging data distinct features of high dimensionality, clinical sensitivity, and inter-institutional variability that require the application of FL with privacy-preserving methods protocols specifically to medical imaging in general. We survey the principles, threat models, and privacy-protective mechanisms specific to neuroimaging tasks, which point to the progress in Differential Privacy, Homomorphic Encryption, and Secure Multi-Party Computation. A multi-axis taxonomy of FL architectures, privacy strategies and neuroimaging settings identifies unexplored areas and gaps in the methodology. We evaluate domain adaptation policies, resistance to the heterogeneity of data, and the effects of privacy assurances on clinical informatics through a critical review of applications in Alzheimer disease, brain tumor analysis, lesion segmentation, and multiple sclerosis. We also address the technical issues, such as statistical heterogeneity, paucity of annotation, efficiency of communication, interpretability, fairness, and ethical implementation. New directions like personalized FL, foundation model pretraining, synthetic data generation, and benchmarking efforts are investigated. Lastly, we find open questions and roadmap milestones that are needed to scale, standardize, align governance, multimodal integrate, and translate neuroimaging FL systems to clinical practice. The purpose of this review is to notify researchers and clinicians on how to build the powerful, privacy-assuring, and clinically feasible federated neuroimaging solutions.