Scholarlink

<p>Accurate optic disc segmentation in fundus images is critical for the diagnosis and monitoring of ocular diseases such as glaucoma, diabetic retinopathy, and age-related macular degeneration. Manual annotation is challenging due to significant variability in image quality, acquisition devices, and pathological manifestations, which complicates systematic screening. To overcome these challenges, we propose a fully automated segmentation framework based on a customized U-Net architecture that effectively integrates localization and context features for accurate optic disc boundary detection. The model was trained on the IDRiD dataset and validated on six datasets, which offer diverse imaging features. Quantitative evaluation shows an average Dice coefficient of 0.9728, an average IoU of 0.9473, an accuracy of 0.9989, a precision of 0.9722, a sensitivity of 0.9739, and a specificity of 0.9994, demonstrating high and consistent performance. A qualitative analysis conducted on DRIVE, DIARETDB1, PAPILA, and E-ophtha EX further illustrates the model’s visual portability. In future work, we plan to perform additional tests on annotated datasets to objectively assess large-scale robustness.</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

Customized U-Net architecture for automated optic disc segmentation in retinal fundus images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月23日

查看全文
  • Ayoub Skouta,
  • Abdelali Elmoufidi,
  • Said Jai-Andaloussi,
  • Ouail Ouchetto

摘要

Accurate optic disc segmentation in fundus images is critical for the diagnosis and monitoring of ocular diseases such as glaucoma, diabetic retinopathy, and age-related macular degeneration. Manual annotation is challenging due to significant variability in image quality, acquisition devices, and pathological manifestations, which complicates systematic screening. To overcome these challenges, we propose a fully automated segmentation framework based on a customized U-Net architecture that effectively integrates localization and context features for accurate optic disc boundary detection. The model was trained on the IDRiD dataset and validated on six datasets, which offer diverse imaging features. Quantitative evaluation shows an average Dice coefficient of 0.9728, an average IoU of 0.9473, an accuracy of 0.9989, a precision of 0.9722, a sensitivity of 0.9739, and a specificity of 0.9994, demonstrating high and consistent performance. A qualitative analysis conducted on DRIVE, DIARETDB1, PAPILA, and E-ophtha EX further illustrates the model’s visual portability. In future work, we plan to perform additional tests on annotated datasets to objectively assess large-scale robustness.