Abstract
Diabetes is a widespread disease in the world and can lead to diabetic retinopathy, macular edema, and other obvious microvascular complications in the retina of the human eye. This study attempts to detect diabetic retinopathy (DR), which has been the main reason behind the blindness of people in the last decade. Timely or early treatment is necessary to prevent some DR complications and control blood glucose. DR is very difficult to detect in time-consuming manual diagnosis because of its diversity and complexity. This work utilizes a deep learning application, a convolutional neural network (CNN), in fundus photography to distinguish the stages of DR. The images dataset in this study is obtained from Xiangya No. 2 Hospital Ophthalmology (XHO), Changsha, China, which is very large, little and the labels are unbalanced. Thus, this study first solves the problem of the existing dataset by proposing a method that uses preprocessing, regularization, and augmentation steps to increase and prepare the image dataset of XHO for training and improve performance. Then, it takes the advantages of the power of CNN with different residual neural network (ResNet) structures, namely, ResNet-101, ResNet-50, and VggNet-16, to detect DR on XHO datasets. ResNet-101 achieved the maximum level of accuracy, 0.9888, with a training loss of 0.3499 and a testing loss of 0.9882. ResNet-101 is then assessed on 1787 photos from the HRF, STARE, DIARETDB0, and XHO databases, achieving an average accuracy of 0.97, which is greater than prior efforts. Results prove that the CNN model (ResNet-101) has better accuracy than ResNet-50 and VggNet-16 in DR image classification.