Scholarlink

<p>The analysis of fundus images is vital for early detection of retinal diseases such as diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD), but traditional methods are resource-intensive. We propose an automated and interpretable diagnostic framework that leverages novel feature representations to improve performance. Our main contribution is a topological feature extraction technique based on <i>Topological Data Analysis</i> (TDA), which captures geometric and structural patterns in fundus images. These features are computationally efficient and interpretable. We integrate them with pretrained CNN features (e.g., ResNet-50) into a hybrid deep model, <i>Topo-CNN</i>, combining global image context with topological structure. We evaluate Topo-CNN on three benchmarks: APTOS (binary and five-class DR), ORIGA (Glaucoma), and IChallenge-AMD. Our model achieves <b>98.7% accuracy/98.9 AUC</b> on binary DR, <b>95.5 AUC</b> on five-class DR, <b>93.8% accuracy/93.6 AUC</b> on AMD, and <b>82.3% accuracy/95.8 specificity</b> on glaucoma. Ablation studies confirm the added value of topological features, and our Topo-CNN consistently outperforms existing methods across tasks.</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

Topo-CNN: Retinal Image Analysis with Topological Deep Learning

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Faisal Ahmed,
  • Mohammad Alfrad Nobel Bhuiyan,
  • Baris Coskunuzer

摘要

The analysis of fundus images is vital for early detection of retinal diseases such as diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD), but traditional methods are resource-intensive. We propose an automated and interpretable diagnostic framework that leverages novel feature representations to improve performance. Our main contribution is a topological feature extraction technique based on Topological Data Analysis (TDA), which captures geometric and structural patterns in fundus images. These features are computationally efficient and interpretable. We integrate them with pretrained CNN features (e.g., ResNet-50) into a hybrid deep model, Topo-CNN, combining global image context with topological structure. We evaluate Topo-CNN on three benchmarks: APTOS (binary and five-class DR), ORIGA (Glaucoma), and IChallenge-AMD. Our model achieves 98.7% accuracy/98.9 AUC on binary DR, 95.5 AUC on five-class DR, 93.8% accuracy/93.6 AUC on AMD, and 82.3% accuracy/95.8 specificity on glaucoma. Ablation studies confirm the added value of topological features, and our Topo-CNN consistently outperforms existing methods across tasks.