Scholarlink

<p>Diabetic retinopathy is an illness associated with diabetes mellitus and is the major cause of blindness among the working-age population. The most effective way to prevent this serious consequence of vision loss is through early and accurate detection. In this regard, deep learning approaches, particularly convolutional neural networks, have shown significant potential in computer vision, medical imaging, and computer-aided diagnosis. This study presents the implementation of a modified U-Net architecture model, trained on IDRiD (Indian Diabetic Retinopathy Image Dataset) for the semantic segmentation of exudates, one of the earliest abnormalities of diabetic retinopathy, using fundus images. The process begins with pre-processing the images and masks, followed by training the model using the IDRiD dataset. The application of data augmentation techniques significantly enhanced the model’s performance, making it robust and accurate. The modified U-Net model consistently outperformed other methods in terms of accuracy. The results show that our model, modified U-Net, surpasses other methods, with remarkable accuracy of 0.9963 and 0.9980 in specificity, 0.9963 in sensitivity 0.9970, and 0.9980 in the Dice coefficient.</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

Implementation of a Modified U-Net Architecture for Hard Exudate Semantic Segmentation using Fundus Images

  • 期刊论文

  • 发表日期：

    2025年6月10日

查看全文
  • Amine El hossi,
  • Abdelali Elmoufidi,
  • Mourad Nachaoui

摘要

Diabetic retinopathy is an illness associated with diabetes mellitus and is the major cause of blindness among the working-age population. The most effective way to prevent this serious consequence of vision loss is through early and accurate detection. In this regard, deep learning approaches, particularly convolutional neural networks, have shown significant potential in computer vision, medical imaging, and computer-aided diagnosis. This study presents the implementation of a modified U-Net architecture model, trained on IDRiD (Indian Diabetic Retinopathy Image Dataset) for the semantic segmentation of exudates, one of the earliest abnormalities of diabetic retinopathy, using fundus images. The process begins with pre-processing the images and masks, followed by training the model using the IDRiD dataset. The application of data augmentation techniques significantly enhanced the model’s performance, making it robust and accurate. The modified U-Net model consistently outperformed other methods in terms of accuracy. The results show that our model, modified U-Net, surpasses other methods, with remarkable accuracy of 0.9963 and 0.9980 in specificity, 0.9963 in sensitivity 0.9970, and 0.9980 in the Dice coefficient.