Scholarlink

Diabetic macular edema (DME) is a significant complication of Diabetic retinopathy (DR) that can cause vision loss and even blindness. DME occurs when fluid escapes from the blood vessels in the macula or the retina thickens. The leakage appears as Hard Exudates (EX), which are yellow or white clusters that differ in shape, size, and location, serving as diagnostic indicators of DME in Color fundus color images (CFI). Detecting early DME can significantly improve treatment choices and patient quality of life. In this study, we propose evaluating three pretrained Convolutional Neural Network models to assess the risk of DME in CFI. This study employs transfer learning to leverage the convolutional basis of pretrained models to extract features from CFI. We apply model fine-tuning to improve higher-order feature representations in the base model, making them more relevant and improving performance. Two publicly available datasets, MESSIDOR and IDRID, were used to train and test the proposed method, achieving an F1-score of 0.89 (MESSIDOR) and 0.90 (IDRID).

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

Performance Evaluation of Pretrained Convolutional Neural Networks for Diabetic Macular Edema Diagnosis in Retinal Fundus Imaging

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • José Araque-Gallardo,
  • Eugenia Arrieta Rodríguez,
  • Margarita Gamarra,
  • Javier Sierra-Carrillo,
  • José Escorcia-Gutierrez

摘要

Diabetic macular edema (DME) is a significant complication of Diabetic retinopathy (DR) that can cause vision loss and even blindness. DME occurs when fluid escapes from the blood vessels in the macula or the retina thickens. The leakage appears as Hard Exudates (EX), which are yellow or white clusters that differ in shape, size, and location, serving as diagnostic indicators of DME in Color fundus color images (CFI). Detecting early DME can significantly improve treatment choices and patient quality of life. In this study, we propose evaluating three pretrained Convolutional Neural Network models to assess the risk of DME in CFI. This study employs transfer learning to leverage the convolutional basis of pretrained models to extract features from CFI. We apply model fine-tuning to improve higher-order feature representations in the base model, making them more relevant and improving performance. Two publicly available datasets, MESSIDOR and IDRID, were used to train and test the proposed method, achieving an F1-score of 0.89 (MESSIDOR) and 0.90 (IDRID).