Scholarlink

<p>Diabetic Retinopathy (DR), the leading cause of vision loss in diabetics, requires early and accurate diagnosis to ensure appropriate treatment and prevent irreversible vision damage. This work highlights an automated and noninvasive approach to detect and classify diabetic retinopathy using Optical Coherence Tomography Angiography (OCTA) images. This study addresses several issues including limited dataset size, image quality, and efficient feature extraction through a novel integration of data augmentation, preprocessing, and deep learning techniques. Conditional Generative Adversarial Networks (cGANs) were employed and improved augmentation techniques were implemented to address the problems posed by limited datasets, increasing the dataset size to 3000 high-resolution images. A comprehensive preprocessing pipeline was used to highlight important retinal features, including contrast enhancement via the Regional Dynamic Histogram Equalization (RDHE) algorithm, noise reduction via the Generalized Gauss-Markov Random Field (GGMRF) method, and vessel segmentation using a Markov-Gibbs Random included Field (MGRF) model. The proposed Convolutional Neural Network (CNN) demonstrated exceptional performance with 99.5% accuracy, 99.67% precision, 99.33% recall/sensitivity, 99.67% specificity, 99.5% F1 score and an AUC ROC of 0.997. With a low false negative rate, these measures demonstrate how well the system can differentiate between DR and non-DR cases. By providing early and accurate diagnosis, our work provides a reliable and useful solution for DR screening, improving the diagnosis of diabetic eyes.</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

A novel deep learning approach for diabetic retinopathy classification using optical coherence tomography angiography

  • 期刊论文

  • 发表日期：

    2025年3月10日

查看全文
  • Abini M. A.,
  • S. Sridevi Sathya Priya

摘要

Diabetic Retinopathy (DR), the leading cause of vision loss in diabetics, requires early and accurate diagnosis to ensure appropriate treatment and prevent irreversible vision damage. This work highlights an automated and noninvasive approach to detect and classify diabetic retinopathy using Optical Coherence Tomography Angiography (OCTA) images. This study addresses several issues including limited dataset size, image quality, and efficient feature extraction through a novel integration of data augmentation, preprocessing, and deep learning techniques. Conditional Generative Adversarial Networks (cGANs) were employed and improved augmentation techniques were implemented to address the problems posed by limited datasets, increasing the dataset size to 3000 high-resolution images. A comprehensive preprocessing pipeline was used to highlight important retinal features, including contrast enhancement via the Regional Dynamic Histogram Equalization (RDHE) algorithm, noise reduction via the Generalized Gauss-Markov Random Field (GGMRF) method, and vessel segmentation using a Markov-Gibbs Random included Field (MGRF) model. The proposed Convolutional Neural Network (CNN) demonstrated exceptional performance with 99.5% accuracy, 99.67% precision, 99.33% recall/sensitivity, 99.67% specificity, 99.5% F1 score and an AUC ROC of 0.997. With a low false negative rate, these measures demonstrate how well the system can differentiate between DR and non-DR cases. By providing early and accurate diagnosis, our work provides a reliable and useful solution for DR screening, improving the diagnosis of diabetic eyes.