A Deep Neural Network-Based Identification System for Diabetic Eye Diseases

Abstract

Diabetic Eye Disease occurs when blood vessels linked to light-sensitive tissue existing in the retina of the eye are damaged. Furthermore, based on the severity level of the disease, it can lead to full blindness and a variety of other visual problems. The present research work is based on the analysis of various Deep Neural Networks (DNN) that are applied on a dataset consisting of retinal images for the prediction of eye disease especially found in type-2 diabetic patients. This study validates that deep learning-based models such as Visual Geometric Group16 (VGG16), Visual Geometric Group 19 (VGG19), Efficient Network121(EfficientNet121), Residual Neural Network50(ResNet50), and Neural Architecture Search Network Large (NAS Net Large) can predict diabetic eye disease. Several image feature extraction techniques (Contour Feature Description, Segmentation, Color Conversion from BGR to RGB, Gaussian Blur, and Cropping) are used for the feature extraction of color retinal images. The dataset comprised 135930 training images whereas 45310 validation images fitted in five DR types such as No DR, Mild, Moderate, Severe and Proliferative, as a result of data split in ratios of 75% (train) and 25% (test). The accuracy based on training data is compared for all classification models considered in this research work and it has been observed VGG16 gives the highest accuracy. Similarly, the training data accuracy of other models used in this work is also considered (between 85%-99%). Likewise, VGG19 and VGG16 both had high validation data accuracy such as 89.01% and 88.27%, respectively, but ResNet50 had the lowest validation data accuracy of 89.01%.