Applications of Deep Learning in Diabetic Retinopathy Detection and Classification: A Critical Review

Abstract

AbstractDiabetes is the world’s fastest-growing illness, causing a slew of complications. One of these conditions is diabetic retinopathy (DR) which causes retinal lesions affecting the vision. There are several levels of DR, ranging from moderate to extreme. Early detection and treatment of DR can decrease the risk of loss of vision. Presently, DR detection is a laborious and blue-collar process that needs qualified ophthalmologists to inspect digital color retinal fundus photographs. As a result, various machine vision-based methods for automatically diagnosing diseases are explored in the literature. Deep learning has emerged in recent years achieving better performance than existing methods in many areas, especially in analyzing and examining the medical images. Deep CNNs are broadly used method in analysis of medical images. To highlight the role of CNNs in DR detection and classification, various recent papers have been studied and considered. The characteristics of various available DR Datasets of colored fundus images are also discussed. The paper suggests a set of steps that will form a process such that it will be promising at all the stages of DR detection/ classification. The explanation behind the suggested set is derived by comparing the results of the papers. Since the results of any method depend on the input, therefore, to see the effects of selected preprocessing method, i.e., CLAHE, was applied to images from most widely used dataset: Kaggle dataset. The processed images were noise-free also the lesions were clearly visible.KeywordsDiabetic retinopathyDeep learningTransfer learningEnsembleMicroaneurysms (MA)Haemorrhages (HM)Soft and hard exudates (EX)Cotton wool spotsNeovascularization (NV) and macular edema (ME)CLAHE