Diabetic Retinopathy is a degenerative retinal disease caused primarily by Diabetes Mellitus, a disease of high blood sugar that damages the retina. In developing countries, DR is one of the main causes of blindness in adults. The treatment for DR will always focus on maintaining the patient’s existing level of vision, since the disease is irreversible. In order to improve the efficacy of treatment, early diagnosis and prompt intervention are necessary to slow down the progress of eyesight loss. Traditional diagnosis necessitate the ophthalmologists to manually review the retinal fundus images, which is generally time-consuming and expensive. This study uses fundus images to explore the efficacy of deep learning approaches for identifying DR in early, moderate, and advanced phase. The diabetic retinopathy classification involves caegorizing diabetic patient’s fundus images based on the scale of severity 0 to 4. Some of the recently introduced technologies are capsule networks, which have shown efficient performance compared with traditional machine learning techniques in DR diagnosis. This paper proposes an optimized capsule network for the detection and classification of diabetic retinopathy. It uses convolutional and primary capsule layers to extract features from fundus images, while class capsule and softmax layers calculate the probability of each image belonging to a specific class. In the APTOS2019 dataset, the accuracy of classification reached 88.96%, which proves that the modified capsule network is very effective in classifying DR images of multiple classes.