Detection of Diabetic Retinopathy Using Deep Learning Techniques

Abstract

Diabetic retinopathy is one of the most compromising complexities of diabetes that can lead to vision impairment and even irreversible blindness if left untreated. For the treatment to be a success early detection is one of most important challenges. Unfortunately, physical diagnose by trained eye physician would require large amount of time to diagnose this disorder individually furthermore, the specific recognizable proof of the diabetic retinopathy stage can be famously precarious from the fundus pictures. Convolutional neural systems (CNN) have been applied effectively in numerous neighboring activities and for discovery of diabetic retinopathy itself. In this paper, we propose an automatic deep-learning-based method for identifying the phases of diabetic retinopathy from the human fundus images. Here, the use of three models i.e. VGG, DenseNet and EfficientNet for the comparison of the efficiencies. Our system utilizes CNN alongside denoising to distinguish highlights like micro-aneurysms and hemorrhages on the retina. We prepared this system utilizing a top of the line GPU on the openly accessible Kaggle dataset utilized for preparing and approval. The standard exactness measurements and quadratic weighted Kappa score is utilized to assess grouping capacity of utilized models we utilized that is determined between the anticipated scores and scores provided in the dataset. The finest tested model achieving an accuracy approximately 97% in detecting the retinopathy and assessing its stage is the DenseNet model with size 300x300. Moreover, system obtained decent Kappa scores for EfficientNet, DenseNet and VGG models respectively over a dataset of 3500 images. The accomplished outcomes demonstrated that deep learning algorithms can be effectively utilized to take care of this difficult issue.