Scholarlink

<p>Diabetic Retinopathy (DR) is a vision-threatening disease that affects individuals with diabetes, leading to ocular lesions resulting from the rupture of retinal blood vessels. The automation of DR detection systems offers a promising solution for early screening and treatment. Moreover, many existing methods demonstrate limited effectiveness in accurately classifying the various stages of diabetic retinopathy (DR), especially at early stages. Adapting these models for DR classification requires fine-tuning hyperparameters, which traditionally involves retraining parts of the model, a time-consuming process. In this paper, we propose three transfer learning based models (AtRD : Auto-tuned Retino Detection for binary Detection, AtR3C for Three Class classification, and AtR5C for Five Class classification) integrating an automated Bayesian optimization process for hyperparameter fine tuning. Our pipeline includes image preprocessing, data imbalance correction, feature extraction using pre-trained models, and a customized multilayer classification stage. We evaluate several pre-trained architectures (InceptionV3, VGG16, ResNet50, Xception) under automated tuning to identify the best performing configuration. The ResNet50 based models consistently achieve superior results. AtRD reaches a detection accuracy of 99.22% and precision of 99.61% on the APTOS dataset. AtR3C yields high precision for “No DR” (93%) and “Early DR” (82%), while AtR5C maintains strong performance at the early stages with 98% for “No DR,” 77% for “Mild DR,” and 81% for “Moderate DR.” These results demonstrate that our optimized models surpass recent state-of-the-art approaches in early DR detection and fine-grained classification, offering a promising solution for clinical screening and disease progression monitoring.</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

Early Detection and Severity Grading of Diabetic Retinopathy Using Fine-Tuned Deep Learning Models with Automated Hyperparameter Optimization

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Samira Ait Kaci Azzou,
  • Djamila Boukredera,
  • Yacine Abiche,
  • Akram Amokrane,
  • Achour Achroufene

摘要

Diabetic Retinopathy (DR) is a vision-threatening disease that affects individuals with diabetes, leading to ocular lesions resulting from the rupture of retinal blood vessels. The automation of DR detection systems offers a promising solution for early screening and treatment. Moreover, many existing methods demonstrate limited effectiveness in accurately classifying the various stages of diabetic retinopathy (DR), especially at early stages. Adapting these models for DR classification requires fine-tuning hyperparameters, which traditionally involves retraining parts of the model, a time-consuming process. In this paper, we propose three transfer learning based models (AtRD : Auto-tuned Retino Detection for binary Detection, AtR3C for Three Class classification, and AtR5C for Five Class classification) integrating an automated Bayesian optimization process for hyperparameter fine tuning. Our pipeline includes image preprocessing, data imbalance correction, feature extraction using pre-trained models, and a customized multilayer classification stage. We evaluate several pre-trained architectures (InceptionV3, VGG16, ResNet50, Xception) under automated tuning to identify the best performing configuration. The ResNet50 based models consistently achieve superior results. AtRD reaches a detection accuracy of 99.22% and precision of 99.61% on the APTOS dataset. AtR3C yields high precision for “No DR” (93%) and “Early DR” (82%), while AtR5C maintains strong performance at the early stages with 98% for “No DR,” 77% for “Mild DR,” and 81% for “Moderate DR.” These results demonstrate that our optimized models surpass recent state-of-the-art approaches in early DR detection and fine-grained classification, offering a promising solution for clinical screening and disease progression monitoring.