Scholarlink

<p>Optical Coherence Tomography (OCT) is essential for early retinal disease detection, enabling timely intervention and preventing vision loss. However, the rapid growth of OCT data has increased diagnostic workload, motivating automated, accurate, and clinically interpretable computer-aided diagnostic systems. In this study, we propose a Cross-Attention CNN–ViT hybrid framework combining DenseNet201-based local feature extraction with ViT-Base/16 global contextual modeling via cross-attention, capturing both fine-grained lesions and long-range retinal dependencies. For interpretability, we introduce a dual explainability framework that fuses Grad-CAM and transformer attention into a unified Grad-Attention Map (GAM), visualizing both localized pathological biomarkers and global retinal context. The model is evaluated on the Kaggle OCT C8 dataset (24,000 images, eight classes: AMD, CNV, CSR, DME, DR, drusen, macular hole, normal), split into training, validation, and independent test sets, with five-fold cross-validation applied only on training data under strict image-level separation. It achieves 97.42% accuracy, 99.63% specificity, and 97.41% F1-score, outperforming CNN, transformer, and hybrid baselines. GAM visualizations align with clinically relevant biomarkers, indicating strong interpretability. The framework demonstrates a robust and scalable system for automated retinal screening, triage, and diagnosis.</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

Clinically explainable cross-attention CNN–ViT framework for automated multi-class retinal disease screening using OCT

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Sabib Ahmed,
  • Jannatul Ferdus Aspia,
  • Md Tanver Rana Sobur,
  • Md. Khabir Uddin Ahamed

摘要

Optical Coherence Tomography (OCT) is essential for early retinal disease detection, enabling timely intervention and preventing vision loss. However, the rapid growth of OCT data has increased diagnostic workload, motivating automated, accurate, and clinically interpretable computer-aided diagnostic systems. In this study, we propose a Cross-Attention CNN–ViT hybrid framework combining DenseNet201-based local feature extraction with ViT-Base/16 global contextual modeling via cross-attention, capturing both fine-grained lesions and long-range retinal dependencies. For interpretability, we introduce a dual explainability framework that fuses Grad-CAM and transformer attention into a unified Grad-Attention Map (GAM), visualizing both localized pathological biomarkers and global retinal context. The model is evaluated on the Kaggle OCT C8 dataset (24,000 images, eight classes: AMD, CNV, CSR, DME, DR, drusen, macular hole, normal), split into training, validation, and independent test sets, with five-fold cross-validation applied only on training data under strict image-level separation. It achieves 97.42% accuracy, 99.63% specificity, and 97.41% F1-score, outperforming CNN, transformer, and hybrid baselines. GAM visualizations align with clinically relevant biomarkers, indicating strong interpretability. The framework demonstrates a robust and scalable system for automated retinal screening, triage, and diagnosis.