Scholarlink

<p>The retinopathy of diabetic patients can be detected at earlier stages to prevent additional deterioration of vision in the patient. CAD-based methods of detection are rather popular today due to their high accuracy and their ability to help a doctor. It was a work that proposed a new approach to image detection of retinopathy using transfer learning. The novelty is in the creation of the modified ResNet-18 architecture (RetinaNet) along with a hybrid preprocessing and segmentation architecture designed to work on a proprietary clinical dataset of more than 2000 fundus images. The images can be subjected to some form of noise. In order to normalize all the input images, pre-processing of all the images is first done through cropping, reshaping, and contrast enhancement. Moreover, the images are passed to the proposed transferring learning model to accomplish the feature extraction for the classification of the images. The accuracy of RetinaNet is 97.09, which is 2–10% higher in clinical noise than 2025 SOTA (e.g., 99.36% of RSG-Net on public data).</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

Classification of retinopathy images using augmentation and transfer learning

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Manjot Kaur,
  • Karthick Panneerselvam,
  • Akhil Sohal,
  • Arshiya

摘要

The retinopathy of diabetic patients can be detected at earlier stages to prevent additional deterioration of vision in the patient. CAD-based methods of detection are rather popular today due to their high accuracy and their ability to help a doctor. It was a work that proposed a new approach to image detection of retinopathy using transfer learning. The novelty is in the creation of the modified ResNet-18 architecture (RetinaNet) along with a hybrid preprocessing and segmentation architecture designed to work on a proprietary clinical dataset of more than 2000 fundus images. The images can be subjected to some form of noise. In order to normalize all the input images, pre-processing of all the images is first done through cropping, reshaping, and contrast enhancement. Moreover, the images are passed to the proposed transferring learning model to accomplish the feature extraction for the classification of the images. The accuracy of RetinaNet is 97.09, which is 2–10% higher in clinical noise than 2025 SOTA (e.g., 99.36% of RSG-Net on public data).