Scholarlink

<p>Diabetic Retinopathy (DR) is an eye disorder in patients with diabetes that occurs when high blood sugar damages the retina, the light-sensitive tissue of the eye. It is becoming increasingly common, and the condition worsens when not detected early. This study investigates five major classes of DR and analyzes several existing works, implementing 19 state-of-the-art models. However, in our experiments, these models did not achieve satisfactory performance on the chosen dataset, with validation accuracies remaining below 80%. As a result, we propose DR-RetinaNet, a model based on the DenseNet201 architecture, modified through selective freezing of early convolutional layers, freezing of batch normalization layers to stabilize training, and fine-tuning of deeper layers for DR-specific feature extraction. These customizations reduce training time, improve stability, and enhance task-specific learning, making the model more effective. Additionally, the dataset used in this study was preprocessed to optimize it for DR-RetinaNet. With the preprocessed dataset, DR-RetinaNet achieved impressive training and validation accuracies of 99% and 94%, respectively. By offering a practical method for the early detection of DR, this research contributes to efforts in improving disease detection, particularly eye disorders, which will benefit the medical sector and individuals with diabetic retinopathy.</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

DR-RetinaNet: A Deep Learning Approach for Early Screening of Diabetic Retinopathy

  • 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
  • Samia Yasmin,
  • Mehedi Hasan,
  • Fariha Jahan,
  • M. F. Mridha

摘要

Diabetic Retinopathy (DR) is an eye disorder in patients with diabetes that occurs when high blood sugar damages the retina, the light-sensitive tissue of the eye. It is becoming increasingly common, and the condition worsens when not detected early. This study investigates five major classes of DR and analyzes several existing works, implementing 19 state-of-the-art models. However, in our experiments, these models did not achieve satisfactory performance on the chosen dataset, with validation accuracies remaining below 80%. As a result, we propose DR-RetinaNet, a model based on the DenseNet201 architecture, modified through selective freezing of early convolutional layers, freezing of batch normalization layers to stabilize training, and fine-tuning of deeper layers for DR-specific feature extraction. These customizations reduce training time, improve stability, and enhance task-specific learning, making the model more effective. Additionally, the dataset used in this study was preprocessed to optimize it for DR-RetinaNet. With the preprocessed dataset, DR-RetinaNet achieved impressive training and validation accuracies of 99% and 94%, respectively. By offering a practical method for the early detection of DR, this research contributes to efforts in improving disease detection, particularly eye disorders, which will benefit the medical sector and individuals with diabetic retinopathy.