Classification of diabetic retinopathy using ensemble convolutional neural network architectures

Abstract

Background Diabetic retinopathy (DR) constitutes a primary cause of blindness across all age groups. Ophthalmologists examine fundus images (FI) to detect and classify stages of DR. Development of deep learning can help clinicians to attain a larger volume in screening and diagnosing diabetic retinopathy, thereby decreasing the burden of visual impairment caused by DR. This study aimed to classify DR using ensemble convolutional neural networks (CNN) architectures. Methods We used data from the Indian Diabetic Retinopathy Image Dataset which consist of typical diabetic retinopathy lesions at pixel level. The dataset contains typical diabetic retinopathy structures as well as normal retinal structures and is divided into three parts: segmentation, classification, and location. There are 516 original color fundus images in the classification used as training set (413 images) and testing set (103 images). We used ensemble CNN architectures to classify diabetic retinopathy as no DR, mild non-proliferative DR (NPDR), moderate NPDR, severe NPDR and proliferative DR (PDR) based on fundus image. Results In this study we successfully created a model with ensemble CNNs to detect DR based on fundus images with area-under-the-curve, sensitivity, and specificity of 0.88, 0.89, and 0.90, respectively, which is on par with the most modern methods. Conclusion Based on the results, this model performs quite well in early detection of diabetic retinopathy and can be used to develop a more accurate model for detecting and classifying diabetic retinopathy. This model can also be used in assisting mass screening at lower cost without reducing diagnostic effectiveness.

Keywords

• Diabetic retinopathy
• deep learning
• convolutional neural networks
• early detection