Scholarlink

<p>Diabetic retinopathy (DR) is a major cause of preventable blindness worldwide with more than 146&#xa0;million people with diabetes worldwide. Automated screening systems are crucial for early detection and thus reduction of vision loss, especially in resource-limited settings where access to specialists is limited. This paper introduces a single-architecture with lightweight deep learning framework, EfficientNetB0 with Multi-Scale Spatial Attention Gates (MSAG), to detect and classify the severity of automated DR detection and severity. The preprocessing pipeline consists of green-channel extraction, contrast-limited adaptive histogram equalization (CLAHE) and massive data augmentation to deal with imaging variability and class imbalance. The model is trained using APTOS 2019 (3,662 training, 1,928 test images) and has a quadratic weighted kappa (QWK) of 0.923 +- 0.008 and macro-averaged AUC of 0.976, which is better than the ResNet50 (QWK: 0.901) and DenseNet121 (QWK: 0.908) baselines. The grad-cam visualization analysis results in interpretable attention maps containing pathologically important areas. The lightweight architecture can achieve sub-second inference using 5.3&#xa0;M parameters, which is ideal to be deployed in mobile screening as well as resource-constrained clinical settings.</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

AI-driven detection and classification of diabetic retinopathy stages using EfficientNetB0

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月21日

查看全文
  • Deepak Dharrao,
  • Madhuri Dharrao,
  • Shreyas Patil,
  • Sangeeth Salvin,
  • Prashant Ahire,
  • Yashwant Dongre

摘要

Diabetic retinopathy (DR) is a major cause of preventable blindness worldwide with more than 146 million people with diabetes worldwide. Automated screening systems are crucial for early detection and thus reduction of vision loss, especially in resource-limited settings where access to specialists is limited. This paper introduces a single-architecture with lightweight deep learning framework, EfficientNetB0 with Multi-Scale Spatial Attention Gates (MSAG), to detect and classify the severity of automated DR detection and severity. The preprocessing pipeline consists of green-channel extraction, contrast-limited adaptive histogram equalization (CLAHE) and massive data augmentation to deal with imaging variability and class imbalance. The model is trained using APTOS 2019 (3,662 training, 1,928 test images) and has a quadratic weighted kappa (QWK) of 0.923 +- 0.008 and macro-averaged AUC of 0.976, which is better than the ResNet50 (QWK: 0.901) and DenseNet121 (QWK: 0.908) baselines. The grad-cam visualization analysis results in interpretable attention maps containing pathologically important areas. The lightweight architecture can achieve sub-second inference using 5.3 M parameters, which is ideal to be deployed in mobile screening as well as resource-constrained clinical settings.