Scholarlink

Diabetic Retinopathy (DR) has become one of the major reasons for vision loss worldwide. It is a retinal complication resulting from diabetes. DR can lead to ultimate blindness if not detected in the early stages. Since the symptoms don’t show till the late stages, it is very important to detect the disease early on to prevent vision loss. Ophthalmologists use fundus images to find internal scars or lesions. For the past few years, machine learning specifically deep learning has played an important role in the automation of medical screening. One of the limitations of using machine learning for medical imaging is a shortage of large medical datasets. Data augmentation plays a vital role in avoiding overfitting problems. This study used six pre-trained models MobileNet, DenseNet201, InceptionV3, VGG16, Xception, and ResNet50 with suitable data augmentation techniques and compared the results. The results showed that the deep learning models trained on augmented DR images performed better compared to the models without prior augmentation. Inception-V3 in particular trained on an augmented dataset achieved the highest 96.02% validation accuracy and 98.06% AUC hence out-performed other pre-trained models.

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

Leveraging Data Augmentation to Improve Deep Transfer Learning Models for Diabetic Retinopathy Severity Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kinza Nazir,
  • Jisoo Kim,
  • Yung-Cheol Byun

摘要

Diabetic Retinopathy (DR) has become one of the major reasons for vision loss worldwide. It is a retinal complication resulting from diabetes. DR can lead to ultimate blindness if not detected in the early stages. Since the symptoms don’t show till the late stages, it is very important to detect the disease early on to prevent vision loss. Ophthalmologists use fundus images to find internal scars or lesions. For the past few years, machine learning specifically deep learning has played an important role in the automation of medical screening. One of the limitations of using machine learning for medical imaging is a shortage of large medical datasets. Data augmentation plays a vital role in avoiding overfitting problems. This study used six pre-trained models MobileNet, DenseNet201, InceptionV3, VGG16, Xception, and ResNet50 with suitable data augmentation techniques and compared the results. The results showed that the deep learning models trained on augmented DR images performed better compared to the models without prior augmentation. Inception-V3 in particular trained on an augmented dataset achieved the highest 96.02% validation accuracy and 98.06% AUC hence out-performed other pre-trained models.