Efficient diabetic retinopathy detection using convolutional neural network and data augmentation

Diabetic retinopathy (DR) has become a major disease which causes loss of vision worldwide. If the retinal disease is not treated in a timely manner, it might progress to severe stages. It is caused by diabetes, which raises the sugar levels in the retina's blood vessels. This paper discusses the novel proposed model which detects the stages of DR in patients with the help of fundus images. It will detect retinopathy disease and predict the stage in which the corresponding fundus image belongs. Generally, DR may be divided into non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). The proposed model has been trained to classify DR into five stages, namely no DR, mild DR, moderate DR, severe DR and PDR, respectively. Having the ability to identify people with DR quickly will be beneficial for both doctors and researchers. In the current situation, manually analysing each little nerve cell from fundus pictures is extremely time-consuming for clinicians. Hence, a CNN-based model with data augmentation to classify DR from fundus images is proposed in this paper. This augmented dataset is used to train different models like DenseNet121, DenseNet169, ResNet50 and InceptionV3 with high-end GPUs. These models achieved accuracy of 96.64, 95.95, 95.71 and 94.73%, respectively. The best accuracy reported is 96.64% with DenseNet121 which is best among other state-of-the-art (SOTA) models.