Scholarlink

Diabetic retinopathy (DR) stands as a severe complication of diabetes mellitus, ranking among the leading causes of blindness worldwide. Early detection of DR from retinal images holds immense importance for timely intervention. However, class imbalance within DR datasets presents a formidable obstacle to effective diagnosis. This paper presents a novel strategy to detect the DR effectively in class-imbalanced datasets. It uses the balanced weighted categorical loss function and transfer learning with VGG-19, DenseNet-201, and ResNet-50 architectures. Initially trained on an unbalanced dataset, our approach incorporates sophisticated data balancing techniques, including up-sampling and down-sampling, to investigate the effects of class imbalance. Experiments are conducted over the publicly available APTOS 2019 dataset. Experimental results show that addressing the challenge of class imbalance can improve the overall efficiency of these models.

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

A Novel Approach for Diabetic Retinopathy Detection with Class Imbalanced Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vikas Kumar Jain,
  • Sanchali Das,
  • Avni Gupta,
  • Ashwin Sharma,
  • Atul Rai,
  • Lavanya Bharani

摘要

Diabetic retinopathy (DR) stands as a severe complication of diabetes mellitus, ranking among the leading causes of blindness worldwide. Early detection of DR from retinal images holds immense importance for timely intervention. However, class imbalance within DR datasets presents a formidable obstacle to effective diagnosis. This paper presents a novel strategy to detect the DR effectively in class-imbalanced datasets. It uses the balanced weighted categorical loss function and transfer learning with VGG-19, DenseNet-201, and ResNet-50 architectures. Initially trained on an unbalanced dataset, our approach incorporates sophisticated data balancing techniques, including up-sampling and down-sampling, to investigate the effects of class imbalance. Experiments are conducted over the publicly available APTOS 2019 dataset. Experimental results show that addressing the challenge of class imbalance can improve the overall efficiency of these models.