Diabetes is the leading cause of blindness among people under the age of 50 years. Diabetes Mellitus (DM) is a direct cause of Diabetic Retinopathy (DR) which is a complication of diabetes where prolonged hyperglycemia leads to endothelial damage in blood vessels and triggers vascular inflammation, known as microangiopathy, which, when affecting the retinal blood vessels, results in retinopathy due to reduced blood supply to the retina. In order to avoid complications associated with chronic diseases such as Diabetes, early detection is vital. 1 The main contributors to the vision loss include: Uncorrected refractive error (671 million), Cataract (100 million), Glaucoma (8 million), Age and related macular degeneration (8 million), DR (4 million). 1

The major two forms of Diabetes Mellitus include the following: Type-1 and Type-2 diabetes: Type-1 commonly manifests in children and adolescents. It is due to the combined interactive effects of immune and environmental factors, leading to a complete deficiency of insulin secretion. 2 Type-2 usually appears in middle-aged adults whose cells become resistant to insulin. The characteristics of DR appear in 60% of subjects with more than 15 years of disease with diabetes. DR can result in rapid loss of vision; the condition does not have symptoms in its early diagnosis. The improvement of no blindness and modification of disease in diabetic patients require regular monitoring and early detection with consistent treatments.

Artificial Intelligence has indeed given a new dimension to medical imaging, which enabled the automation of identification and classification of complex patterns within visual information. Machine Learning and Deep learning within AI have emerged as the most powerful tools that allow the computer to garner knowledge through big datasets and enhance their capability without explicit programming. 3 Deep learning, in particular, uses a type of artificial neural network composed of many interrelated layers, called Convolutional Neural Networks, which are specifically developed to perceive and analyze visual data such as images.

Basically, while comprising several layers to extract effective features from the image, a CNN comprises convolutional layers, pooling layers, and fully connected layers.

The convolutional layers apply filters on the input images to extract important features such as edges and textures, extracted at an increasingly higher level of complexity with a deeper network. The pooling layers decrease the spatial dimensions of data while keeping crucial information and reducing computational burdens. Finally, fully connected layers take the features learned from the previous layers and use them to make some kind of prediction, for example, whether any anomalies exist in the images.

Concretely speaking, a CNN updates its internal parameters during training by means of backpropagation on labelled datasets that may consist of both training and classification data. A training dataset teaches it to recognize certain features present in every class, while a classification dataset evaluates its performance on unseen data. This hierarchical learning enables CNNs to locate and classify objects in images with high accuracy.

These range from the detection and classification of various abnormalities, such as DR lesions in retinal fundus images, for which CNNs and other deep learning approaches are widely used in medical imaging. Diabetic retinopathy is one of the most common complications of diabetes. It affects the blood vessels of the retina and may lead to severe visual loss. The main types of DR lesions have been identified as microaneurysms, hemorrhages, exudates, and neovascularization, each having different visual characteristics, which may make their detection by humans difficult. 4 AI-powered algorithms significantly enhance the accuracy and speed of such lesion detection to assist the clinicians for early diagnosis and management of DR. Training of CNNs using large-scale annotated datasets helps them in the detection of subtle changes within the retinal images and enhances diagnostic capability while reducing the workload for health professionals. 5, 6

This paper is aimed at discussing the integration of AI, notably deep learning and CNNs in detecting DR lesions on fundus images; more importantly, how the inception of these technologies has transformed modern diagnostics.

Diabetic retinopathy (DR) has two main stages: Non-proliferative DR (NPDR) and Proliferative DR (PDR). In NPDR, damaged retinal blood vessels leak fluid, causing swelling and symptoms like microaneurysms, hemorrhages, exudates, and inter-retinal microvascular abnormalities. PDR is a more severe stage marked by the growth of abnormal blood vessels, potentially leading to complete blindness.

For an ophthalmologist, MAs are the first indication of DR, which result from leakage from the retina's small blood vessels.  They are red in colour, smaller in size, and circular in shape.

When the walls of MAs rupture, Hemorrhages (HMs) occur. Blot haemorrhages are larger red lesions, whereas dot haemorrhages resemble bright red dots. A clear sign of moderate DR and these outperform MAs with indistinct margins in size.

Hard exudates (yellow dots visible in the retina) and soft exudates (pale yellow or white areas with ill-defined edges) are two types of exudates. The lipid and proteinaceous components of the hard exudates, including albumin and fibrinogen, leak from the damaged blood-retinal barrier. They are typically deposited in the retina's outer plexiform layer.

Cotton wool spots (CWS) due to vascular occlusion, also known as soft exudates, occurs if the lipid buildup is on the macula or nearby, they can result in total blindness. Exudates (EXs) (Hard and Soft) are referred to as bright lesions and MAs and HMs as dark lesions.

Venous beading (VB), It shows damaged walls of major retinal vessels and is a delayed indication in non-proliferative DR. This is one of the best indicators that a person may develop proliferative DR (PDR).

Intra Retinal Microvascular Abnormality (IRMA), These are tiny blood arteries with unusual shapes that divert blood from arterioles to venules. ⁷ 

We primarily focus on NPDR lesions that are MAs, HMs, or EXs in this paper. According to the location and frequency of the lesions, ophthalmologists typically classify NPDR into three categories: Mild, Moderate, and Severe. Below we discuss about DR Level and Clinical features of DR level.

Few Microaneurysms.

At least one hemorrhage or Microaneurysms and/or at least one of the following: Retinal hemorrhages, Hard exudates, Cotton wool spots, Venous beading.

Use the 4-2-1 rule below. Only one of these criteria have to be met to be considered severe, 4 quadrants with microaneurysms or dot blot haemorrhages, ≥2 quadrants with venous beading, ≥1 quadrant with intraretinal microvascular abnormality (IRMA). ⁸

Microaneurysms (MA) are early indicators of DR, appearing as tiny red dots with sharp margins, usually not exceeding 125 micrometres, due to focal dilations in small retinal arteries. When MAs rupture, they result in haemorrhages. Retinal lesions like MAs, exudates, and haemorrhages occur in approximately 77-90% of diabetics who have had the disease for 15 years or more. This review aims to categorize these DR lesions and identify research gaps in DR detection for future studies.

Haemorrhages (HM), like microaneurysms, appear as large patches on the retina with irregular edges up to 125 micrometres wide. They are classified into two types- flame and blot based on their depth and surface area. The frequency and pattern of haemorrhages reflect the severity of diabetic retinopathy. Below, we provide an overview of literature using deep learning approaches for haemorrhage detection in DR.

Exudates (EXs), caused by fluid leaks from retinal blood vessels, are a significant contributor to vision loss in diabetic retinopathy. They are classified into- hard exudates yellow spots with sharp edges due to plasma leakage and soft exudates, which are swollen nerve fibres appearing as white, oval regions with blurred edges on the retina. Below, we summarize literature using deep learning methods for detecting these exudates in DR.

The following studies focuses on AI-based classification techniques applied to diabetic retinopathy, including those for both the detection of diabetic retinopathy versus non-diabetic retinopathy and nuanced stages. It also evaluates FDA-cleared algorithms that have demonstrated success in clinical settings, underlining their contribution to improving diagnostic precision and accessibility in the screening and management of diabetic retinopathy.

As artificial intelligence (AI) continues to advance, its role in diabetic retinopathy (DR) is becoming increasingly pivotal. Future perspectives on AI in DR should emphasize the practical clinical applications of FDA-approved algorithms, rather than focusing solely on in silico evaluations. AI systems such as IDx-DR, EyeArt, and AEYE have already received FDA approval, demonstrating their effectiveness in real-world settings. These algorithms are transforming DR screening by providing reliable, automated detection and stage classification, which enhances diagnostic efficiency and accessibility, especially in underserved areas. Looking ahead, the integration of these AI tools into routine clinical practice promises to improve patient outcomes through earlier and more accurate detection of diabetic retinopathy, ultimately facilitating better management and prevention of vision loss.

The field of Artificial Intelligence has made great strides in promoting early detection for diabetic retinopathy (DR) by offering significant advantages in clinical settings, particularly, in areas of the world with limited access to ophthalmologists. Therefore, AI-powered systems like IDx-DR, EyeArt, and AEYE enable the analysis of retinal fundus photographs with the intention of detecting and grading DR at various levels. This feature has great potential in peripheral and under-resourced areas where specialized eye care services might be short in supply, thus facilitating early detection and timely referral of cases for further evaluation and management. Artificial intelligence in community practices enhances early diagnostic capabilities and optimizes patient management to minimize the risk of visual impairment through early intervention.

Nonetheless, in spite of their progress, artificial intelligence systems possess certain constraints. The challenges encompass the necessity for extensive and varied datasets for training purposes to guarantee generalizability across distinct populations and imaging conditions. Moreover, artificial intelligence algorithms may encounter difficulties arising from discrepancies in image quality and patient demographics, which could adversely affect diagnostic precision. In addition, the deployment of AI tools necessitates meticulous evaluation regarding their incorporation into current clinical workflows, as well as continuous validation to uphold efficacy and dependability. Addressing these limitations is crucial for optimizing AI's role in DR detection and ensuring its equitable application in diverse healthcare settings.

Recent developments in medical image processing are facilitating quick and automated disease screening. Statistics show that DR affects 80% of diabetes patients who have had the disease for 15 to 20 years or longer. Worldwide, there is concern about diabetes and gestational diabetes (GDM). A serious case of DR during pregnancy or a worsening of an existing DR are potential outcomes of GDM. 27 Diabetes affects more than 171 million individuals globally. According to a poll by the World Health Organization (WHO), there will be 366 million cases of diabetes worldwide by 2030. 28

DR identification in the early stages of the illness must be improved by a multidisciplinary collaborative approach. With advancements in AI algorithms, this technique will enhance early detection of many additional retinal diseases in addition to DR screening. Automated screening techniques are not just useful for DR; they may also be used for other diseases, including as glaucoma and age-related macular degeneration, where early diagnosis would probably lead to better clinical outcomes. Such algorithms are widely available nowadays. All such algorithms should undergo thorough validation testing to verify their suitability for clinical application.

The initial sign of DR is blood vessel lesions with tiny, circular red patches. MAs, HEMs, EXs, and CWs are moderate indications of DR. To distinguish between mild and severe levels, it is crucial to consider the ratio of these symptoms. The visual resemblance of symptoms between No-DR, Mild-DR, and occasionally Moderate-DR makes it challenging to recognise in the early stages of diagnosis, which is a significant problem in identifying the level of DR severity. Finally, if DR worsens and reaches an advanced stage, may lead to vision loss.

Although DR cannot be reversed, it is crucial to identify it early to limit future suffering. For instance, early signs of DR will almost always be present in non-proliferative DR stages, and being able to recognise and categorise those stages using the right diagnostic technique may allow one to save their vision.

In this review paper, a major portion of the work focuses on the study of Haemorrhages, Microaneurysms and Exudates and the Figure 9 shows about the Current Research ratio papers localized by lesion types in DR detection.

These DL-based strategies could be incorporated into screening systems that are currently being developed to improve and categorise the DR stage using lesion detection methods on a variety of fundus images. The primary problem raised in the studies under examination is the weak NPDR lesions detection research ratio in comparison to other lesion DR detection research.

Additionally, the variations of fundus images that can be utilised to evaluate indications are constrained by dataset limits. The analysis of retinal scans has grown quicker, more inclusive, and more generalizable as a result of the effectiveness of Deep Learning techniques, yet the criteria employed to assess the outcomes and the corresponding datasets are still skewed and uneven between researches. These developments allow for the generalisation of DL-based models and the evaluation of a wider variety of symptoms and signs, which may aid in the discovery of the underlying pathologies underlying retina-based disorders. Due to the lack of publicly available datasets, DR lesions screening is still a problem. While most recent DL improvements achieve encouraging classification scores, some are still unable to distinguish impacted lesions.

Additionally, we believe that in the future, our assessment can be expanded to offer a comprehensive and current review of the difficult and rapidly expanding field of DR detection.

Artificial intelligence in the analysis of medical images has greatly enhanced early detection and screening for diabetic retinopathy, among other retinal diseases. Nonetheless, difficulties persist, including the need for extensive validation, a lack of significant datasets, and further refinement to detect stages of non-proliferative conditions. Continuing improvements in deep learning technologies and growing databases are likely to make diabetic retinopathy screening methods more robust and efficient. This includes a number of challenges that must be overcome if clinical outcomes are to be improved, and AI technologies fully utilized in the timely diagnosis and treatment of diabetic retinopathy.