Scholarlink

<p>Automated diabetic retinopathy (DR) screening must capture both global retinal topology and fine-grained pathological details, which is difficult for standard grid-based Convolutional Neural Networks (CNNs). We propose DS-WaveGNN, a dual-path wavelet graph neural network. After contrast enhancement and superpixel segmentation, DS-WaveGNN builds a spatio-semantic graph that explicitly fuses deep semantic features with clinical-pathological priors, such as texture and microvascular density. An adaptive wavelet mechanism dynamically decouples the graph signal into low- and high-frequency paths, separating macroscopic vascular structures from microscopic lesion details. These paths are processed independently using lightweight Deep Separable Graph Convolutions (DSGraphConv), and an ordinal-aware hybrid loss penalizes large-span severity errors. On APTOS 2019, DS-WaveGNN achieves 86.01% accuracy and 0.9122 Quadratic Weighted Kappa (QWK) for five-class grading, and 99.56% accuracy for binary screening. After retraining on Messidor-2 with the same protocol, it achieves 80.25% accuracy and 0.8460 QWK, demonstrating robust multiscale graph representation for DR 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

DS-WaveGNN: a Dual-path wavelet graph neural network for fundus image classification

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Hailong Ma,
  • Haiyang Wang,
  • Pujie Jing

摘要

Automated diabetic retinopathy (DR) screening must capture both global retinal topology and fine-grained pathological details, which is difficult for standard grid-based Convolutional Neural Networks (CNNs). We propose DS-WaveGNN, a dual-path wavelet graph neural network. After contrast enhancement and superpixel segmentation, DS-WaveGNN builds a spatio-semantic graph that explicitly fuses deep semantic features with clinical-pathological priors, such as texture and microvascular density. An adaptive wavelet mechanism dynamically decouples the graph signal into low- and high-frequency paths, separating macroscopic vascular structures from microscopic lesion details. These paths are processed independently using lightweight Deep Separable Graph Convolutions (DSGraphConv), and an ordinal-aware hybrid loss penalizes large-span severity errors. On APTOS 2019, DS-WaveGNN achieves 86.01% accuracy and 0.9122 Quadratic Weighted Kappa (QWK) for five-class grading, and 99.56% accuracy for binary screening. After retraining on Messidor-2 with the same protocol, it achieves 80.25% accuracy and 0.8460 QWK, demonstrating robust multiscale graph representation for DR grading.