Scholarlink

<p>Diabetic retinopathy (DR) is a major cause of vision impairment and requires highly accurate assessment while preserving patient data privacy. This study presents a confidential DR grading (CDRG) framework that addresses privacy concerns of centralized learning through federated learning integrated with an enhanced retinopathy severity grading network (ERSG-Net). The model is evaluated using the Mendeley DR dataset, where raw retinal images remain at local client sites to ensure data confidentiality. A federated contrast enhancement (FedCE) module standardizes illumination and improves lesion visibility across clients without sharing data. Retinal structures and pathological regions are segmented using a federated dual-attention segmentation network (FDASN), combining spatial and channel attention for consistent lesion localization. Vision transformers (ViT) extract high-level contextual features for severity grading, which are classified by ERSG-Net. The program manager optimization algorithm (PMOA) improves convergence stability and performance. The proposed CDRG framework was evaluated on the Mendeley diabetic retinopathy dataset comprising 35,126 retinal fundus images, distributed across 5 federated client nodes to emulate decentralized clinical institutions, with experiments conducted for five DR severity classes: No DR, Mild, Moderate, Severe, and Proliferative DR. Experimental results show superior accuracy, sensitivity, robustness, and privacy preservation, with ERSG-Net-PMOA improving accuracy and sensitivity by 7.66% and 6.18%, respectively.</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

An enhanced retinopathy severity grading network with federated learning for secure diabetic retinopathy assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • R. Ashok Kumar,
  • N. Thillaiarasu

摘要

Diabetic retinopathy (DR) is a major cause of vision impairment and requires highly accurate assessment while preserving patient data privacy. This study presents a confidential DR grading (CDRG) framework that addresses privacy concerns of centralized learning through federated learning integrated with an enhanced retinopathy severity grading network (ERSG-Net). The model is evaluated using the Mendeley DR dataset, where raw retinal images remain at local client sites to ensure data confidentiality. A federated contrast enhancement (FedCE) module standardizes illumination and improves lesion visibility across clients without sharing data. Retinal structures and pathological regions are segmented using a federated dual-attention segmentation network (FDASN), combining spatial and channel attention for consistent lesion localization. Vision transformers (ViT) extract high-level contextual features for severity grading, which are classified by ERSG-Net. The program manager optimization algorithm (PMOA) improves convergence stability and performance. The proposed CDRG framework was evaluated on the Mendeley diabetic retinopathy dataset comprising 35,126 retinal fundus images, distributed across 5 federated client nodes to emulate decentralized clinical institutions, with experiments conducted for five DR severity classes: No DR, Mild, Moderate, Severe, and Proliferative DR. Experimental results show superior accuracy, sensitivity, robustness, and privacy preservation, with ERSG-Net-PMOA improving accuracy and sensitivity by 7.66% and 6.18%, respectively.