Scholarlink

Background <p>Diabetic retinopathy (DR) is a long-term eye condition associated with diabetes, caused by prolonged hyperglycemia, that damages retinal blood vessels and may result in progressive vision impairment or blindness. In the early stages, DR often shows few or no clear symptoms, making timely detection clinically challenging. Accurate assessment of disease severity is therefore important for guiding appropriate clinical management and reducing the risk of vision loss. Conventional diagnosis relies on manual fundus examination by ophthalmologists. However, large-scale screening with these methods can be labor-intensive and time-consuming.</p> Objective <p>To address these challenges, this research presents a deep learning framework to classify DR severity into five categories: no DR, mild, moderate, severe, and proliferative DR.</p> Methods <p>The proposed method for detecting and grading DR relies on Convolutional Neural Networks (CNNs). The models were trained and evaluated using publicly available retinal fundus image datasets, with APTOS 2019 (Kaggle) serving as the primary dataset. For further evaluation and analysis, STARE, DRIVE, DIARETDB0, and CHASE were also utilized.</p> Results <p>Model performance was evaluated using five-fold stratified cross-validation to provide reliable estimates of classification performance. Among the evaluated architectures, InceptionResNetV2 demonstrated the best performance, achieving an overall accuracy of 93% in classifying DR severity. Model performance was further examined using receiver operating characteristic (ROC) analysis, statistical validation, and Grad-CAM visualizations to support interpretability.</p> Conclusion <p>The results highlight the potential of deep learning approaches for DR severity classification, providing a methodological basis for future research on automated DR screening systems.</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

Diabetic retinopathy detection and categorization using deep learning

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Parvathi Subramanian,
  • Santhi Durairaj

摘要

Background

Diabetic retinopathy (DR) is a long-term eye condition associated with diabetes, caused by prolonged hyperglycemia, that damages retinal blood vessels and may result in progressive vision impairment or blindness. In the early stages, DR often shows few or no clear symptoms, making timely detection clinically challenging. Accurate assessment of disease severity is therefore important for guiding appropriate clinical management and reducing the risk of vision loss. Conventional diagnosis relies on manual fundus examination by ophthalmologists. However, large-scale screening with these methods can be labor-intensive and time-consuming.

Objective

To address these challenges, this research presents a deep learning framework to classify DR severity into five categories: no DR, mild, moderate, severe, and proliferative DR.

Methods

The proposed method for detecting and grading DR relies on Convolutional Neural Networks (CNNs). The models were trained and evaluated using publicly available retinal fundus image datasets, with APTOS 2019 (Kaggle) serving as the primary dataset. For further evaluation and analysis, STARE, DRIVE, DIARETDB0, and CHASE were also utilized.

Results

Model performance was evaluated using five-fold stratified cross-validation to provide reliable estimates of classification performance. Among the evaluated architectures, InceptionResNetV2 demonstrated the best performance, achieving an overall accuracy of 93% in classifying DR severity. Model performance was further examined using receiver operating characteristic (ROC) analysis, statistical validation, and Grad-CAM visualizations to support interpretability.

Conclusion

The results highlight the potential of deep learning approaches for DR severity classification, providing a methodological basis for future research on automated DR screening systems.