Grading Diabetic Retinopathy Severity Using Modern Convolution Neural Networks (CNN)

Abstract

Diabetic Retinopathy is an ophthalmic complication eventuating with impaired vision or even blindness if left unmanaged. In modern society, ophthalmologists are responsible for diagnosing diabetic retinopathy to prevent such outcomes. However, medical costs and the availability of clinicians are just some of the barriers of entry to these services. Portable and more automated solutions could find immediate effectiveness in remote areas and developing countries lacking necessary medical infrastructure. Over time, various computer vision-based techniques have been proposed to automatically diagnose diabetic retinopathy. However, grading diabetic retinopathy in its different stages is still yet to reach the required clinical precision. In this paper, we developed a solution to this problem by image processing followed by ensembling state of the art Convolution Neural Networks (CNNs). We demonstrate the effectiveness of the developed method on publicly available datasets and show that the method outperforms previous studies in multi-classification metrics, achieving accuracies for 5-classes of up to 88.71 % and quadratic weighted kappa scores of up to 0.9256. These outcomes provide promising validation for the clinical relevance and applicability of modern CNN architectures as automated, portable and accurate solutions for the grading of diabetic retinopathy severity.