Scholarlink

This study introduces a deep learning framework for the multi-class classification of diabetic retinopathy (DR) severity, using a fine-tuned Xception architecture. By integrating customized preprocessing techniques and comprehensive data augmentation strategies, the proposed pipeline improves classification accuracy across various DR stages. The results demonstrate robust performance in early-stage detection (No DR), while highlighting opportunities for further refinement in identifying Mild and Proliferative DR cases. This methodology holds promise for enabling automated screening, especially in resource-limited environments. The approach includes a detailed description of the APTOS 2019 dataset, the image normalization process, and various augmentation techniques. The Xception model pre-trained on ImageNet is fine-tuned using transfer learning and hyperparameter optimization. The performance of model’s, evaluated through confusion matrix, classification reports, ROC curves and accuracy curve shows its effectiveness in classifying DR stages. It demonstrates strong performance, particularly in identifying No DR cases (97.28% precision), and reasonable results for Mild DR (80.56% precision) and Proliferative DR (82.14% precision). The study concludes by summarizing the key findings, discussing their broader implications, and proposing future research directions, with a focus on improving accuracy for detecting subtle and advanced stages of DR.

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

Multi-class Classification for Diabetic Retinopathy Detection Using Fine-Tuned Xception Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hajar Karkri,
  • Ali El Moussati,
  • Yassine Ayat,
  • Oumayma Rachdi,
  • Mohammed Abdelouahab

摘要

This study introduces a deep learning framework for the multi-class classification of diabetic retinopathy (DR) severity, using a fine-tuned Xception architecture. By integrating customized preprocessing techniques and comprehensive data augmentation strategies, the proposed pipeline improves classification accuracy across various DR stages. The results demonstrate robust performance in early-stage detection (No DR), while highlighting opportunities for further refinement in identifying Mild and Proliferative DR cases. This methodology holds promise for enabling automated screening, especially in resource-limited environments. The approach includes a detailed description of the APTOS 2019 dataset, the image normalization process, and various augmentation techniques. The Xception model pre-trained on ImageNet is fine-tuned using transfer learning and hyperparameter optimization. The performance of model’s, evaluated through confusion matrix, classification reports, ROC curves and accuracy curve shows its effectiveness in classifying DR stages. It demonstrates strong performance, particularly in identifying No DR cases (97.28% precision), and reasonable results for Mild DR (80.56% precision) and Proliferative DR (82.14% precision). The study concludes by summarizing the key findings, discussing their broader implications, and proposing future research directions, with a focus on improving accuracy for detecting subtle and advanced stages of DR.