Convolutional Neural Network for Classification of Diabetic Retinopathy Grade

Diabetic Retinopathy (DR) represents an important group of lesions found in the retina of patients who suffer from diabetes mellitus, affecting around one out of three patients and presenting a global prevalence of approximately 34.6%. Besides, DR is characterized as being the leading cause of vision loss in adults. Its diagnosis consists on a series of screening tests to obtain digital photographs of the retina, to find the grade of the evolution of the disease, which can be classified into four grades. The early detection and diagnosis of DR are fundamental to prevent its evolution. In this paper it is proposed the implementation of the Convolutional Neural Network (CNN), VGGNet-like, which is a model focused in the classification of images based on object recognition and detection. The main objective is the classification of a set of images containing the four different grades in the evolution of DR. The datasets used are the Indian Diabetic Retinopathy Image Dataset and the Diabetic Retinopathy Detection. The performance of the CNN proposed is evaluated through a statistical analysis based on accuracy, the loss function and area under the curve (AUC). The results present statistically significant values, obtaining 0.81 of accuracy, 0.49 of loss function and, 0.71 of micro-average and 0.72 of macro-average in the AUC. According to the results, it is possible to conclude that the CNN implemented can classify DR into its different grades in patients with presence of diabetes mellitus, obtaining a preliminary Computer-Aided Diagnosis tool that could be supportive for the diagnosis of the evolution of DR.