Scholarlink

This paper analyzes the problem of creating an automatic detection of diabetic retinopathy (DR) symptoms in patients treated in a nephrology clinic. The challenge was that the specialist initially classified all images obtained from Wrocław Medical University as images without DR symptoms (class 0). As a starting point, we have selected a new, relatively little-known but accurately classified by specialists database: ‘Dataset from fundus images for the study of diabetic retinopathy’ [3]. Based on this database, a CNN classifier with VGG16 architecture was trained using transfer learning. Our classifier afforded over \(84\%\) accuracy for the seven-class recognition problem, and in the case of two classes (no DR or DR decision), the classification accuracy was \(96\%\) . Unfortunately, despite some attempts to unify the photos from the database and the images from Wrocław, most of the images from the new set, treated as a test set, were classified incorrectly. Combining both sets and training a new classifier on that basis allowed us to achieve \(96\%\) accuracy in the case of two classes (no DR or DR). All images obtained from the Wrocław Medical University set that were randomly selected for the validation set were classified correctly.

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

Creating an Automatic Classifier for Detecting Diabetic Retinopathy Based on the Combining a New 7-Class Dataset Fundus Images and a Collection of Images from Another Source

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Michał Zmonarski,
  • Ewa Skubalska-Rafajlowicz,
  • Aleksandra Zgryźniak,
  • Sławomir Zmonarski

摘要

This paper analyzes the problem of creating an automatic detection of diabetic retinopathy (DR) symptoms in patients treated in a nephrology clinic. The challenge was that the specialist initially classified all images obtained from Wrocław Medical University as images without DR symptoms (class 0). As a starting point, we have selected a new, relatively little-known but accurately classified by specialists database: ‘Dataset from fundus images for the study of diabetic retinopathy’ [3]. Based on this database, a CNN classifier with VGG16 architecture was trained using transfer learning. Our classifier afforded over \(84\%\) accuracy for the seven-class recognition problem, and in the case of two classes (no DR or DR decision), the classification accuracy was \(96\%\) . Unfortunately, despite some attempts to unify the photos from the database and the images from Wrocław, most of the images from the new set, treated as a test set, were classified incorrectly. Combining both sets and training a new classifier on that basis allowed us to achieve \(96\%\) accuracy in the case of two classes (no DR or DR). All images obtained from the Wrocław Medical University set that were randomly selected for the validation set were classified correctly.