Scholarlink

Diabetic retinopathy (DR), one of the leading causes of blindness detection worldwide, necessitates early detection to prevent vision loss. In this work, we developed a system to classify the severity of DR using deep learning techniques. We evaluated three convolutional neural network (CNN) architectures—Inception V3, Visual Geometry Group 16 (VGG16), and EfficientNet B3—on the APTOS 2019 blindness detection dataset, which consists of retinal fundus images labelled by DR severity. To improve model performance, preprocessing techniques such as resizing, normalization, contrast enhancement using CLAHE, class balancing, and data augmentation techniques like rotation and zooming were applied to enhance image quality and address data imbalance. Among the three architectures, Inception V3 achieved the highest performance, with a training accuracy of 87.81% and a testing accuracy of 78.14%.

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

Evaluating CNN-Based Approaches for Diabetic Retinopathy Detection and Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Trushita Deshmukh,
  • Sanjana Patil,
  • Saanvi Surve,
  • Swarangi Kale,
  • Dhanashri Bhosale,
  • Saguna Ingle

摘要

Diabetic retinopathy (DR), one of the leading causes of blindness detection worldwide, necessitates early detection to prevent vision loss. In this work, we developed a system to classify the severity of DR using deep learning techniques. We evaluated three convolutional neural network (CNN) architectures—Inception V3, Visual Geometry Group 16 (VGG16), and EfficientNet B3—on the APTOS 2019 blindness detection dataset, which consists of retinal fundus images labelled by DR severity. To improve model performance, preprocessing techniques such as resizing, normalization, contrast enhancement using CLAHE, class balancing, and data augmentation techniques like rotation and zooming were applied to enhance image quality and address data imbalance. Among the three architectures, Inception V3 achieved the highest performance, with a training accuracy of 87.81% and a testing accuracy of 78.14%.