Deep Learning Based Method for Computer Aided Diagnosis of Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is a retinal disease caused by the high blood sugar levels that may damage and block the blood vessels feeding the retina. In the early stages of DR, the disease is asymptomatic; however, as the disease advances, a possible sudden loss of vision and blindness may occur. Therefore, an early diagnosis and staging of the disease is required to possibly slow down the progression of the disease and improve control of the symptoms. In response to the previous challenge, we introduce a computer aided diagnosis tool based on convolutional neural networks (CNN) to classify fundus images into one of the five stages of DR. The proposed CNN consists of a preprocessing stage, five stage convolutional, rectified linear and pooling layers followed by three fully connected layers. Transfer learning was adopted to minimize overfitting by training the model on a larger dataset of 3.2 million images (i.e. ImageNet) prior to the use of the model on the APTOS 2019 Kaggle DR dataset. The proposed approach has achieved a testing accuracy of 77% and a quadratic weighted kappa score of 78%, offering a promising solution for a successful early diagnose and staging of DR in an automated fashion.