Scholarlink

Background <p>The most serious complications of diabetes include Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME),&#xa0;potentially leading to impairment of vision or complete loss of sight if not detected earlier. Manual diagnosis by ophthalmologists is&#xa0;often time-consuming, costly, and subject to variability.</p> Objective <p>This study proposes an optimized deep learning (DL) framework for the early detection and severity grading of DR and&#xa0;DME.</p> Methods <p>The system enhances retinal image quality through preprocessing, extracts key features using a novel multi-head self-attention&#xa0;Xception (MSAXCP) network, selects important features using Support Vector Machine-centered Recursive Feature&#xa0;Elimination (SVMRFE), and classifies them with a Gannet Optimized Long Short-Term Memory (GOLSTM) model.</p> Results <p>The system was evaluated using 5-fold cross-validation on the IDRiD and Messidor datasets, achieving 99.7% accuracy for&#xa0;both DR and DME, surpassing existing methods.</p> Conclusion <p>The proposed model offers early, accurate, and automated diagnosis of diabetic eye diseases. However, this study&#xa0;presents a proof-of-concept evaluated on benchmark datasets, and substantial real-world clinical validation remains essential before&#xa0;clinical adoption.</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

Severity level prediction of diabetic retinopathy and diabetic macular edema using deep and transfer learning approaches from fundus images

  • 期刊论文

  • 发表日期：

    2025年11月13日

查看全文
  • Bala Subramanian Chokkalingam,
  • Gayathri Vairam

摘要

Background

The most serious complications of diabetes include Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME), potentially leading to impairment of vision or complete loss of sight if not detected earlier. Manual diagnosis by ophthalmologists is often time-consuming, costly, and subject to variability.

Objective

This study proposes an optimized deep learning (DL) framework for the early detection and severity grading of DR and DME.

Methods

The system enhances retinal image quality through preprocessing, extracts key features using a novel multi-head self-attention Xception (MSAXCP) network, selects important features using Support Vector Machine-centered Recursive Feature Elimination (SVMRFE), and classifies them with a Gannet Optimized Long Short-Term Memory (GOLSTM) model.

Results

The system was evaluated using 5-fold cross-validation on the IDRiD and Messidor datasets, achieving 99.7% accuracy for both DR and DME, surpassing existing methods.

Conclusion

The proposed model offers early, accurate, and automated diagnosis of diabetic eye diseases. However, this study presents a proof-of-concept evaluated on benchmark datasets, and substantial real-world clinical validation remains essential before clinical adoption.