Scholarlink

<p>Federated learning (FL) has become a key approach for collaboratively training deep learning models across multiple institutions while preserving sensitive patient data. In this research, we develop nine deep federated learning (DFL) models for brain tumor classification using MRI scans, encompassing a total of 23,317 images distributed across four participating institutions. The work utilizes several publicly available datasets for federated training and cross-evaluation sets, ensuring heterogeneous and non-IID data distributions. We propose a secure federated learning with weighted model aggregation (FedSec-WtAvg) framework that enhances model robustness by integrating secure aggregation and Fernet encryption mechanisms. The scale and complexity of federated training, combined with hyperparameter optimization and large-scale cross-dataset evaluation on 3332 images, require substantial computational resources. To meet these demands, parallel GPU clusters and the Supercomputing Mission’s Param Porul facility were employed, demonstrating the necessity of high performance computing (HPC) for efficient model training and evaluation. Performance assessment across independent datasets confirms the effectiveness of the proposed approach in generalizing across diverse MRI sources, achieving a maximum accuracy of 98.35%. These results highlight the robustness, efficiency, and suitability of the method for HPC-driven, large-scale, privacy-preserving medical imaging applications.</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

Cross-dataset performance evaluation and secure federated learning with weighted model aggregation for MRI brain tumor image classification

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • G. Appasami,
  • Nickolas Savarimuthu

摘要

Federated learning (FL) has become a key approach for collaboratively training deep learning models across multiple institutions while preserving sensitive patient data. In this research, we develop nine deep federated learning (DFL) models for brain tumor classification using MRI scans, encompassing a total of 23,317 images distributed across four participating institutions. The work utilizes several publicly available datasets for federated training and cross-evaluation sets, ensuring heterogeneous and non-IID data distributions. We propose a secure federated learning with weighted model aggregation (FedSec-WtAvg) framework that enhances model robustness by integrating secure aggregation and Fernet encryption mechanisms. The scale and complexity of federated training, combined with hyperparameter optimization and large-scale cross-dataset evaluation on 3332 images, require substantial computational resources. To meet these demands, parallel GPU clusters and the Supercomputing Mission’s Param Porul facility were employed, demonstrating the necessity of high performance computing (HPC) for efficient model training and evaluation. Performance assessment across independent datasets confirms the effectiveness of the proposed approach in generalizing across diverse MRI sources, achieving a maximum accuracy of 98.35%. These results highlight the robustness, efficiency, and suitability of the method for HPC-driven, large-scale, privacy-preserving medical imaging applications.