Scholarlink

Diabetic retinopathy is an incurable ocular disease stemming from elevated blood glucose levels, typical of diabetes mellitus. This ailment affects the retina’s blood vessels, potentially leading to blindness if not treated promptly. Computer-aided diagnosis (CAD) systems based on artificial intelligence are designed to assist specialists in the early diagnosis of such diseases, thereby facilitating prognosis. In this study, we utilized the DDR and IDRiD datasets for the binary classification of Diabetic Retinopathy using the K-Fold method for cross-validation and transfer learning. The first dataset was used for model training, while the second was reserved exclusively for external validation. The VGG16, ResNet50, and MobileNetV2 networks were evaluated in accuracy, precision, F1 score, AUC ROC score, recall, and specificity. MobileNetV2 exhibited the best performance, achieving values of 78.68%, 76.68%, 86.15%, 93.38%, 98.28%, and 38.10% for accuracy, precision, F1 score, AUC ROC score, recall, and specificity, respectively, when tested with images from IDRiD, which had never been seen by the model before.

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

Convolutional Neural Networks for Diabetic Retinopathy Fundus Image Classification Using Transfer Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Carlos Victor G. Moura,
  • Maria Cecília A. Castro,
  • Antonio Ermeson P. Alves,
  • Pedro Crosara Motta,
  • Débora F. de Assis,
  • Bruno Riccelli dos Santos Silva,
  • Paulo César Cortez

摘要

Diabetic retinopathy is an incurable ocular disease stemming from elevated blood glucose levels, typical of diabetes mellitus. This ailment affects the retina’s blood vessels, potentially leading to blindness if not treated promptly. Computer-aided diagnosis (CAD) systems based on artificial intelligence are designed to assist specialists in the early diagnosis of such diseases, thereby facilitating prognosis. In this study, we utilized the DDR and IDRiD datasets for the binary classification of Diabetic Retinopathy using the K-Fold method for cross-validation and transfer learning. The first dataset was used for model training, while the second was reserved exclusively for external validation. The VGG16, ResNet50, and MobileNetV2 networks were evaluated in accuracy, precision, F1 score, AUC ROC score, recall, and specificity. MobileNetV2 exhibited the best performance, achieving values of 78.68%, 76.68%, 86.15%, 93.38%, 98.28%, and 38.10% for accuracy, precision, F1 score, AUC ROC score, recall, and specificity, respectively, when tested with images from IDRiD, which had never been seen by the model before.