Scholarlink

<p>The widespread use of digital devices has led to a surge in visual-system disorders, underscoring the importance of early and accurate diagnosis. This study introduces SSRetNet, a novel dual-branch deep learning model that integrates spatial and spectral feature extraction for the classification of retinal diseases using Optical Coherence Tomography (OCT) images. The architecture combines a ResNet-50 backbone for spatial features with a wavelet-enhanced GoogLeNet branch for spectral information, enabling complementary feature fusion to improve diagnostic accuracy. SSRetNet is evaluated on three publicly available OCT datasets: OCTDL, OCTID, and OCT-C8. It achieves classification accuracies of 94.20%, 99.36%, and 98.25%, respectively, outperforming several state-of-the-art models. Individual branch analyses reveal the unique diagnostic contributions of spatial and spectral domains, while their fusion leads to significant performance gains. To enhance interpretability, Grad-CAM is applied, highlighting class-relevant retinal structures and improving clinical trust in the model’s predictions. The proposed SSRetNet framework demonstrates strong generalizability, high performance, and explainability, making it suitable for real-world clinical deployment in ophthalmology and telehealth applications. This work contributes to the growing field of AI-assisted retinal diagnostics by providing an effective, interpretable, and dataset-robust solution for the early detection of retinal disorders.</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

SSRetNet: a dual-branch spatial–spectral feature fusion network for robust retinal disease diagnosis

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Abdulfattah Ba Alawi,
  • Tolga Aydın,
  • Ferhat Bozkurt

摘要

The widespread use of digital devices has led to a surge in visual-system disorders, underscoring the importance of early and accurate diagnosis. This study introduces SSRetNet, a novel dual-branch deep learning model that integrates spatial and spectral feature extraction for the classification of retinal diseases using Optical Coherence Tomography (OCT) images. The architecture combines a ResNet-50 backbone for spatial features with a wavelet-enhanced GoogLeNet branch for spectral information, enabling complementary feature fusion to improve diagnostic accuracy. SSRetNet is evaluated on three publicly available OCT datasets: OCTDL, OCTID, and OCT-C8. It achieves classification accuracies of 94.20%, 99.36%, and 98.25%, respectively, outperforming several state-of-the-art models. Individual branch analyses reveal the unique diagnostic contributions of spatial and spectral domains, while their fusion leads to significant performance gains. To enhance interpretability, Grad-CAM is applied, highlighting class-relevant retinal structures and improving clinical trust in the model’s predictions. The proposed SSRetNet framework demonstrates strong generalizability, high performance, and explainability, making it suitable for real-world clinical deployment in ophthalmology and telehealth applications. This work contributes to the growing field of AI-assisted retinal diagnostics by providing an effective, interpretable, and dataset-robust solution for the early detection of retinal disorders.