Scholarlink

<p>Glaucoma, a major contributor to irreversible blindness, often progresses silently, making early detection vital. Manual interpretation of retinal fundus images requires expertise and is prone to variability across clinicians. To solve the issue, a novel automated classification framework is proposed that leverages handcrafted features extracted from retinal zones segmented through a Fibonacci-based annular ring division technique. Concentric regions centered on the optic disc enhance regional representation, especially in clinically significant zones. The classification pipeline employs a hybrid deep learning model combining 1D CNN layers, bidirectional LSTM units, and an attention mechanism, finalized by dense and softmax layers. Learned representations benefit from spatial and sequential modelling, with attention guiding focus toward informative regions. Extensive evaluation across publicly available datasets validates the effectiveness of the architecture. The model achieved classification accuracies of 98.48% on Drishti-GS, 97.98% on Origa, and 97.76% on RimOneV2. Subsets of HVD yielded 98.89% on HVD-Advance, 97.86% for HVD-Early, 97.28% for HVD-Binary class, and 95% for the multi-class classification. Consistent performance across diverse datasets highlights the reliability and clinical potential of the proposed glaucoma detection pipeline.</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

Attention‑Guided Hybrid Deep Learning with Fibonacci and Annular Ring Handcrafted Features for Glaucoma Detection in Retinal Images

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Dip Das,
  • B Ramachandra Reddy

摘要

Glaucoma, a major contributor to irreversible blindness, often progresses silently, making early detection vital. Manual interpretation of retinal fundus images requires expertise and is prone to variability across clinicians. To solve the issue, a novel automated classification framework is proposed that leverages handcrafted features extracted from retinal zones segmented through a Fibonacci-based annular ring division technique. Concentric regions centered on the optic disc enhance regional representation, especially in clinically significant zones. The classification pipeline employs a hybrid deep learning model combining 1D CNN layers, bidirectional LSTM units, and an attention mechanism, finalized by dense and softmax layers. Learned representations benefit from spatial and sequential modelling, with attention guiding focus toward informative regions. Extensive evaluation across publicly available datasets validates the effectiveness of the architecture. The model achieved classification accuracies of 98.48% on Drishti-GS, 97.98% on Origa, and 97.76% on RimOneV2. Subsets of HVD yielded 98.89% on HVD-Advance, 97.86% for HVD-Early, 97.28% for HVD-Binary class, and 95% for the multi-class classification. Consistent performance across diverse datasets highlights the reliability and clinical potential of the proposed glaucoma detection pipeline.