+44-3308224832Rachit Katiyar*
This paper aims to review and propose some techniques for the early detection of diabetic retinopathy. Diabetic retinopathy is an eye complication that mostly happens in older age people which are affected with diabetes. Diabetes occurs when there is a high sugar level in our blood. When a person is affected with diabetes then there is a huge chance for diabetic retinopathy. In diabetic retinopathy, light sensitive tissue of the blood vessels at the back side of retina gets affected by diabetes. In reversible Vision loss, blindness may occur when necessary treatment has not been taken on time. Early detection of diabetic retinopathy may lead to stop these complications Various deep learning models have been used to detect the diabetic retinopathy at early stage which are Vgg19, Inception, ResNet50, Vision transformer, AlexNet, DenseNet. In this paper, the author proposes a transfer learning model of CNN by using the Resnet50 architecture in which fully connected layer used to detect high level features. The preprocessing of the fundus images done through Contrast Limited Histogram Equalization (CLAHE), RGB to Grayscale conversion, RGB to HSI conversion, remove the noise by applying gaussian filtering mean and median filtering, and augmented the image through geometric transformation and colour transformation. Segmentation of the given model have been done by a U-Net architecture and multilayer thresholding and adaptive thresholding. The given model extracts the feature of different lesions which are microaneurysms, hamorrhage, exudates by feature extraction of retinal image by applying morphological operation, thresholding, clustering. The proposed model uses ResNet50 model for classification of different severity levels such as normal, early diabetic retinopathy, mild NPDR, moderate NPDR, severe NPDR, PDR and neovascularization. The proposed model has accuracy of 96.56% for successfully classification of different severity levels of DR which is much improved as compared to other deep learning models present. ODIR dataset have been used here where evaluation parameters are accuracy, precision, specificity, sensitivity and F1 score.
Share this article