Scholarlink

<p>Diabetic retinopathy (DR), a progressive microvascular complication of diabetes, remains a leading cause of preventable blindness worldwide. Its clinical progression—from mild non-proliferative to severe proliferative stages—demands early and accurate detection to enable timely therapeutic intervention. Traditional diagnostic practices rely on the manual analysis of retinal fundus images, yet these approaches face significant limitations due to the scarcity of ophthalmic specialists and restricted access to screening, particularly in low-resource settings. This underscores the critical need for automated, scalable, and high-performing diagnostic solutions. In response, this study introduces SCGNet, an advanced deep learning framework for automated DR classification that integrates graph-based relational reasoning, residual convolutional architectures, and a meta-learning ensemble strategy. The system is supported by a comprehensive preprocessing pipeline featuring adaptive Master-Slave artifact filtering and progressive image enhancement to optimize visual feature quality. SCGNet was trained exclusively on the APTOS 2019 dataset and rigorously evaluated on two external datasets—IDRiD and a clinically curated private dataset—to assess its generalizability across diverse populations and imaging conditions. On APTOS, the model achieved an accuracy of 99.45%, precision of 99.55%, recall of 99.12%, F1-score of 99.12%, Cohen’s Kappa of 99.13%, and MCC of 99.14%. Validation on the IDRiD dataset yielded an accuracy of 94.95%, while the private Moroccan dataset produced an accuracy of 95.85%, with consistently high scores across all performance metrics. These findings highlight the model’s capacity to maintain robust diagnostic performance under domain shifts, making it a viable candidate for integration into real-world DR screening pipelines. By enabling precise, early-stage detection across heterogeneous clinical environments, SCGNet offers promising utility in reducing the burden of diabetic vision loss and enhancing long-term patient outcomes across global healthcare systems.</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

Robust multiclass diabetic retinopathy classification via hybrid graph attention and convolutional ensemble learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Assma Ihiya,
  • Fouad Yakoubi,
  • Bahia Ouazzani Chahdi,
  • Ibtissam Benchaji,
  • Samira Douzi,
  • Bouabid El Ouahidi

摘要

Diabetic retinopathy (DR), a progressive microvascular complication of diabetes, remains a leading cause of preventable blindness worldwide. Its clinical progression—from mild non-proliferative to severe proliferative stages—demands early and accurate detection to enable timely therapeutic intervention. Traditional diagnostic practices rely on the manual analysis of retinal fundus images, yet these approaches face significant limitations due to the scarcity of ophthalmic specialists and restricted access to screening, particularly in low-resource settings. This underscores the critical need for automated, scalable, and high-performing diagnostic solutions. In response, this study introduces SCGNet, an advanced deep learning framework for automated DR classification that integrates graph-based relational reasoning, residual convolutional architectures, and a meta-learning ensemble strategy. The system is supported by a comprehensive preprocessing pipeline featuring adaptive Master-Slave artifact filtering and progressive image enhancement to optimize visual feature quality. SCGNet was trained exclusively on the APTOS 2019 dataset and rigorously evaluated on two external datasets—IDRiD and a clinically curated private dataset—to assess its generalizability across diverse populations and imaging conditions. On APTOS, the model achieved an accuracy of 99.45%, precision of 99.55%, recall of 99.12%, F1-score of 99.12%, Cohen’s Kappa of 99.13%, and MCC of 99.14%. Validation on the IDRiD dataset yielded an accuracy of 94.95%, while the private Moroccan dataset produced an accuracy of 95.85%, with consistently high scores across all performance metrics. These findings highlight the model’s capacity to maintain robust diagnostic performance under domain shifts, making it a viable candidate for integration into real-world DR screening pipelines. By enabling precise, early-stage detection across heterogeneous clinical environments, SCGNet offers promising utility in reducing the burden of diabetic vision loss and enhancing long-term patient outcomes across global healthcare systems.