Early Diabetic Retinopathy Detection using Data Augmentation Techniques and DenseNet - 121

A medical condition that affects the eyes and is often associated with diabetes is known as Diabetic Retinopathy. It may lead to surgery and can cause blindness. The success of the treatment depends on the early detection of the disease. The correct identification of the retinopathy stage, however, is challenging and necessitates competent human interpretation of human fundus pictures. Therefore, in this research paper, we propose a new detection mechanism that can effectively and efficiently detect the 4 stages of retinopathy. To implement this, we considered the fundus images dataset available on the Kaggle website. Moreover, our approach involves Ben Graham’s preprocessing and modern data augmentation techniques. Furthermore, we employ the DenseNet-121 model, a hybrid artificial neural network (ANN) and convolutional neural network (CNN) architecture, for the classification of the four stages of retinopathy. As a result, we were able to surpass the current state-of-the-art with an accuracy of 99.04%, a quadratic weighted kappa of 0.922, and a macro-average specificity of 0.9496.