Diabetes is a critical chronic health disease that is characterized by elevated blood sugar levels. Diabetes patients are more widespread today than ever before, and their numbers are skyrocketing. Inadequate treatments might result in visual deterioration and possibly permanent vision loss. Sadly, determining the exact level of diabetic retinopathy is extremely  difficult and needs experienced human analysis of retinal images. Streamlining the diagnostic  process is significant and can benefit millions of individuals. This study present a deep learning approach to automatic screening of diabetic retinopathy.

Introduction of Diabetes

Diabetes is a serious chronic health condition that is caused due to high blood sugar levels. Obesity in young individuals, as well as a family history of diabetes, are also to blame. Although there is no cure for diabetes, regular exercise, eating healthily, and losing weight can assist in coping with the illness. Changing lifestyle is very helpful in managing the diabetes. Diabetic patients are more common than ever before, and their numbers are increasing massively. The world had merely 108 million diabetes patients in 1980 and now, there are 539 million adults all over the globe living with Diabetes which is expected to rise to  783 million in the next 20 years. In the United Kingdom, there are about 4.9 million diabetic  patients. 13.6 million individuals in the UK, and 541 million people worldwide are at risk of  developing diabetes. One in every six hospital patients has diabetes. Around 9600 patients get their legs, toes, or feet removed each year due to diabetes. Over 1,700 people in the UK lose their sight to diabetes each year, making it one of the leading causes of blindness. Diabetic retinopathy is an eye-related complication that is caused directly due to diabetes. Diabetes retinopathy (DR) is one of the four leading causes of blindness globally. Diabetic retinopathy leads to visual distortion because of fluid leaking from the retina’s blood vessels, which can lead to the creation of lesions in the retina. If not diagnosed or treated on time, it can be a cause of blindness, but it takes several years to reach a stage where it can damage a patient’s retina. An increasing number of diabetic patients worldwide are also a cause of increased cases of diabetic retinopathy. One of the main causes of vision loss and blindness in the UK, particularly among working-class adults, is diabetic retinopathy. Despite such frightening figures, research suggests that appropriate and comprehensive eye treatment and monitoring can eliminate at least 90% of new occurrences. The longer an individual may have diabetes mellitus, the more likely it is that he or she may develop diabetic retinopathy. Nearly two-thirds of Type 2 diabetics in the UK develop some extent of retinopathy within 20 years of diagnosis of diabetes. It has three main stages which are background retinopathy, pre-proliferate retinopathy, and proliferate retinopathy. The most important thing a person can do to decrease the risk of getting diabetic retinopathy or to halt its progression is to maintain normal blood sugar and blood pressure level. It is one of the most important aspects of patient treatment. For manual detection of diabetic retinopathy, we need expert clinicians, advanced equipment, khttps://onlinelibrary.wiley.com/doi/10.1111/j.1755-3768.2016.0635

. Diabetes is more prevalent in low and middle-income nations than in other ones. These countries do not have the financial resources, skilled health specialists, or equipment to examine such a big number of diabetes patients. These patients must be examined once a year on a regular basis, necessitating a large setup to do this sort of frequent examination. Retinal photographs of the eye are used to diagnose the condition. But even so, it only reveals .

Mild symptoms until it has become too late for an effective therapy. The medical examination procedure which is used to detect diabetic retinopathy these days is very effective, but it requires a specialized ophthalmologist who can accurately examine the retinal images of the patients. It is a time-consuming and difficult job for a clinician to examine a complete image and find out all lesions. An automatic screening procedure using deep learning techniques will help in reducing the burden on health systems of low and middle-income countries. It will also help hospitals to focus on other patients. Deep learning techniques are quite famous in healthcare these days and many areas of research are being explored with them.

Use of machine learning in healthcare and medicine

The use of machine learning in healthcare and medicine in the modern era has significantly increased the effectiveness of digital and real-time treatment approaches. The process of dynamic learning and the constantly evolving neural network model allows it to become more accurate and precise with its findings. Diabetes has grown to be a radical chronic disease  that impacts blood sugar levels as well as multiple distinct human bodily organs, including  kidneys, heart and eyes especially the retina. Diabetic Retinopathy (DR) stems from diabetes  targeting the retina of the eye causing impaired vision, blurriness and ultimately leading to blindness if not detected early and cured. In diabetic retinopathy, the eye’s blood vessels are affected and the retina’s sensitive and vulnerable tissue doesn’t receive enough blood to function; this ongoing phenomenon causes mild visual abnormalities before blindness sets in. Starting with Stage 0, DR progresses through a set of five stages. Although the patient’s diabetic retinopathy has started at this point, the retina has not yet exhibited any clear indications or symptoms of the condition. After some time, the retina develops a micro aneurysm (MA), an abnormality that can weaken the walls of the retinal capillaries and cause damage to them. This subtle irregularity is the onset for Stage 1. The condition known as non-proliferative diabetic retinopathy, or NPDR, develops when the MA from the first stage spreads to additional retinal quadrants. Stage 2 is a moderate NDPR that has expanded to several retinal corners. Patients in stage 3 have serious inter-retinal hemorrhages that extend deep into several quadrants and result in neovascularization, or perhaps the growth of numerous more abnormal blood vessels. These blood veins extend far into the retina and occur simultaneously in the choroid. A transition from non-proliferative diabetic retinopathy to proliferative diabetic retinopathy occurs in Stage 4, the last stage of diabetic retinopathy. Before they experience total vision loss, the patients at this point urgently need laser therapy. The Preventative measures for DR can be taken in order to control the growth of micro aneurysms by constantly monitoring your A1c . A recommended method is to take hemoglobin A1c tests regularly for early detection of abnormalities in the retina. One of the best ways to control your A1c levels is to consume more fruits, less fats, cutting down on salt. intake and regularly exercising. Stress is another direct factor that causes tension in the blood vessels leading to limited supply of oxygen and blood to the optical nerves. A well-rested and energized body is able to function efficiently and helps lower blood sugar levels. Numerous other indirect factors, including as excessive alcohol consumption and cigarette use, can contribute to the development of micro aneurysms. Partial blindness affects 1.93 million individuals worldwide, or 4% of the population of the United Kingdom. Given these facts, diabetic retinopathy may be reduced if correctly identified at an early stage. The actions that may be performed to identify diabetic retinopathy in its early stages will be covered in depth in this research project. Although manual detection methods are quite good and are 98% accurate, it requires experienced and highly competent doctors as well as a large medical infrastructure, which are expensive. Typically, places where diabetes is spreading like wildfire cannot afford such medical infrastructure. Covid-19 has already placed a significant strain on health-care systems across the world. This will help reduce the strain on local governments and medical institutions. In the current world situation, the number of diabetic patients is increasing day by day and it will be resulting in more frequent cases of diabetic retinopathy. Health systems cannot bear the burden of such big populations. Therefore, there is a need for a reliable automatic screening procedure, it can be designed with 100 times less cost and is more reliable and time-saving. Hospital appointments are also time taking and costly. It also saves time for disease treatment since spending more time on disease identification causes the disease to worsen. Major health concerns, such as the detection of diabetic retinopathy, demand the use of intelligent automated systems. Therefore, in this study, we present an automated system for both binary and multi-label classification of diabetic retinopathy.

Methods

To solve the challenging task of detecting DR in fundus images, we need to use certain approaches that can extract information from the images and train on the different pixels of the images. Machine learning can be used to tackle this problem by extracting relevant information from images with a three-dimensional data matrix, which can then assist us in evaluating the scale of the images. Machine learning models are further subdivided into two branches: deep neural networks (DNN), which can learn from images like a human brain, and artificial neural networks, which act similarly to human brains and are made up of many neurons. DNN consists of an input layer, a hidden layer, and an output layer, with the number of hidden layers ranging from one to many. DNN are dense like human brains and are the product of the connection of millions of neurons. DNN is trained in a feed forward manner, with inputs from the input layer being fed throughout the model. All neurons are triggered by applying the activation function to the weights and the corresponding linked neuron. This process is repeated until the neural network reaches the output layer. In this manner, the model has been validated by calculating the error on the predicted output and comparing it to the expected output, and then a backward process is performed on the model to improve the weights of the model while considering some other hyper parameters such as lambda, learning rate, and momentum. This way, the total loss of the models is reduced to achieve higher accuracy on the data, and eventually better weights are preserved, which may subsequently be used to evaluate the model.

DNN is further subdivided into convolutional neural network (CNN) algorithms, which are particularly built for images and have the capacity to detect hidden information in images and then learn from it. Deep learning models are built using shallow neural networks. As a result, they have many similarities, such as layer-wise design, parameter tuning, calculating the gradient of the cost function via back propagation, and using same optimization techniques. The presence of various learning convolutional layers that are automatically extracting features, is the major difference in both architectures. Deep learning algorithms for categorizing images are provided with the features retrieved by previous layers as input. Similarly, we have pretrained models derived from CNN called VGG16 and Alexines, which are further discussed below but classed as pre trained models since they are highly complicated DNN models that can outperform a customized CNN. These are commonly referred to as pre-trained models as they can pass information across multiple layers to obtain better results. These pre-trained models are already being tested on larger datasets of images, demonstrating their capacity to analyse larger datasets and generate higher accuracy. These pre-trained models are the best feature extraction models since they consider a huge number of parameters for training and testing.