Scholarlink

Ultra-widefield fundus imaging (UWF) enhances retinal diagnostics by offering comprehensive retinal views, crucial for detecting and managing diabetic retinopathy (DR). This study introduces a deep learning method utilizing self-supervised learning through an Autoencoder network to assess the quality of UWF images. The classification task distinguishes between gradable and ungradable images, a critical component for effective DR screening. The model was evaluated using the MICCAI UWF4DR 2024 dataset, which included 434 training images (with 201 images shared with Task 2) and 61 validation images. The system demonstrated excellent performance during the validation phase, achieving an AUROC of 0.8863, sensitivity of 0.9730, and specificity of 0.7500. In the test phase, the model achieved an AUROC of 0.8979, AUPRC of 0.9371, sensitivity of 0.7627, and specificity of 0.9000. These results highlight the model’s effectiveness in accurately classifying images as gradable or ungradable, demonstrating robust performance in UWF image quality assessment. This approach shows great potential for integration into large-scale DR screening programs, improving diagnostic efficiency and consistency in teleophthalmology.

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

Deep Self-supervised Learning for Ultra-widefield Fundus Image Quality Assessment

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ammar M. Okran,
  • Saif Khalid Musluh,
  • Saddam Abdulwahab,
  • Domenec Puig,
  • Hatem A. Rashwan

摘要

Ultra-widefield fundus imaging (UWF) enhances retinal diagnostics by offering comprehensive retinal views, crucial for detecting and managing diabetic retinopathy (DR). This study introduces a deep learning method utilizing self-supervised learning through an Autoencoder network to assess the quality of UWF images. The classification task distinguishes between gradable and ungradable images, a critical component for effective DR screening. The model was evaluated using the MICCAI UWF4DR 2024 dataset, which included 434 training images (with 201 images shared with Task 2) and 61 validation images. The system demonstrated excellent performance during the validation phase, achieving an AUROC of 0.8863, sensitivity of 0.9730, and specificity of 0.7500. In the test phase, the model achieved an AUROC of 0.8979, AUPRC of 0.9371, sensitivity of 0.7627, and specificity of 0.9000. These results highlight the model’s effectiveness in accurately classifying images as gradable or ungradable, demonstrating robust performance in UWF image quality assessment. This approach shows great potential for integration into large-scale DR screening programs, improving diagnostic efficiency and consistency in teleophthalmology.