Scholarlink

<p>Diabetic retinopathy grading from retinal fundus photographs remains challenging because subtle lesion patterns, inter-class visual similarity, imaging variability, and class imbalance limit the robustness of single-backbone models. We propose a compact multi-backbone retinal representation framework that integrates local lesion-sensitive CNN features, hierarchical DenseNet121 features, and global Vision Transformer embedding. The fused representation is refined using a Multi-Objective Evolutionary Algorithm based on Decomposition, which jointly optimizes classification error and feature compactness before LightGBM-based prediction. The framework was evaluated on five public fundus image datasets, including APTOS 2019, Messidor-2, DDR, Diabetic Balanced, and DR Dataset, for both binary DR detection and five-class severity grading. Across datasets, the proposed method achieved consistent improvements over single-backbone baselines, with the best binary accuracy of 99.09% and best multiclass accuracy of 84.72%. Ablation experiments further demonstrated the complementary value of local, hierarchical, and global representations, as well as the contribution of MOEA/D-based feature selection to reducing redundancy while preserving discriminative power. Code, trained configurations, and reproducible data split files are publicly available at: <a href="https://github.com/NoushinSaba/DR-MOEAD-Fusion">https://github.com/NoushinSaba/DR-MOEAD-Fusion</a><Emphasis Type="Underline">.</Emphasis> These results suggest that compact, multi-scale visual representation fusion can improve the efficiency and reproducibility of retinal image grading.</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

Tri-scale retinal feature fusion with multi-objective selection for efficient diabetic retinopathy grading

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Afia Zafar,
  • Muhammad Attique Khan,
  • Noushin Saba,
  • Samar Ahmad,
  • Saleha Masood,
  • Fatimah Alhayan

摘要

Diabetic retinopathy grading from retinal fundus photographs remains challenging because subtle lesion patterns, inter-class visual similarity, imaging variability, and class imbalance limit the robustness of single-backbone models. We propose a compact multi-backbone retinal representation framework that integrates local lesion-sensitive CNN features, hierarchical DenseNet121 features, and global Vision Transformer embedding. The fused representation is refined using a Multi-Objective Evolutionary Algorithm based on Decomposition, which jointly optimizes classification error and feature compactness before LightGBM-based prediction. The framework was evaluated on five public fundus image datasets, including APTOS 2019, Messidor-2, DDR, Diabetic Balanced, and DR Dataset, for both binary DR detection and five-class severity grading. Across datasets, the proposed method achieved consistent improvements over single-backbone baselines, with the best binary accuracy of 99.09% and best multiclass accuracy of 84.72%. Ablation experiments further demonstrated the complementary value of local, hierarchical, and global representations, as well as the contribution of MOEA/D-based feature selection to reducing redundancy while preserving discriminative power. Code, trained configurations, and reproducible data split files are publicly available at: https://github.com/NoushinSaba/DR-MOEAD-Fusion. These results suggest that compact, multi-scale visual representation fusion can improve the efficiency and reproducibility of retinal image grading.