Analysis of Deep Learning Performance for Diabetic Retinopathy Severity Classification

Diabetic retinopathy (DR) is a common diabetic complication that affects the retina of the eye. The severity of DR is determined by the number and type of lesions, such as microaneurysms, haemorrhages, and exudates that appear on the surface of the retina. However, DR is hard to be detected in the initial stages and may vary from expert to expert. This may lead to a serious side effect in giving the patient a suitable treatment, which can cause a significant impact commonly among high-risk patients. Thus, the demand for advanced DR diagnosis and treatment has drawn the attention of researchers. The main contribution of this work is to develop an automated grading system using deep learning architecture in the urge to help experts in identifying the severity level of DR. The process that has been addressed in this work begins with the pre-processing step, followed by the segmentation of features using local entropy thresholding. In the classification stage, three different deep learning classifier architectures, namely CNN, ResNet152v2, and Inception-v3 convolution neural networks were used to differentiate the category between the normal, mild non-proliferative diabetic retinopathy (NPDR), moderate NPDR, and severe NPDR. Overall, the classification performance results show that the ResNet152v2 model is a better classifier than the other two models with a testing accuracy of 90%, 93%, 97%, and 93% for normal, mild NPDR, moderate NPDR, and severe NPDR, respectively.