 This paper presents a novel approach for automatic screening of diabetic retinopathy, DR. The proposed method uses an asymmetric deep learning feature for segmentation of the optic disc and blood vessels from the retina images. After segmentation, a convolutional neural network, CNN, with a support vector machine, SVM, is used for classification of the lesions as either normal, microaneurisms, hemorrhages or exudates. The proposed method achieves high accuracy of 98.6% and 91.9% for non-diabetic retinopathy detection in the aptis and mesodore datasets, respectively. Additionally, the accuracy of exudate detection is also improved by the precise retinal image segmentation. This article was authored by Pradeep Kumar Jena, Bonamali Kuntia, Charal Tapalli, and others.