<![CDATA[Deep learning algorithms effectively work with, a significant amount of data. trained on small datasets tend to have poor generalization. Data augmentation techniques can be used to make better use of existing training data, improving the applicability of deep learning methods. However, traditional data augmentation methods often produce limited additional credible data. The deep learning approach's performance can be enhanced by generating new data by employing generative adversarial networks (GANs). Although GANs have been extensively used to improve the performance of convolutional neural networks (CNNs), there has been relatively less research on data augmentation methods specifically for GAN training. This study focuses on using a Wasserstein GAN (WGAN) architecture for generating synthetic optical coherence tomography angiography (OCTA) images of diabetic retinopathy to aid in the detection of different types of diabetic retinopathy diseases, including proliferative diabetic retinopathy (PDR), Severe non-PDR (NPDR), Moderate NPDR, and Mild NPDR. WGAN, provides the generator with a more informative learning signal, making training more stable, particularly in high-dimensional spaces. The trained WGAN model is saved in .h5 file format (HDF), converted to portable network graphics (PNG) image format, and then classified into different categories of diabetic retinopathy using a ResNet50 model with various fine-tuning methods. The proposed model has demonstrated better results than the previous study. 99.95% accuracy is exhibited.]]>
Copyrights © 2024
