Artificial Intelligence for Diabetic Retinopathy
Introduction
Diabetes is a chronic disease with a high prevalence globally. In 2019, approximately 463 million adults aged 20 to 79 years were living with diabetes, and this number is expected to rise to around 600 million by 2040. In China, the prevalence of diabetes among adults has increased from 9.7% in 2010 to 12.8% in 2018. Diabetes can cause damage to nerves, blood vessels, and multiple systems in the body. Diabetic retinopathy (DR) is one of the main complications of diabetes. Globally, it is estimated that 34.6% of diabetic patients suffer from DR, with 10.2% experiencing impaired vision. In China, 18.45% of the diabetic population has DR, and the prevalence increases with the duration of diabetes. Similarly, 17.6% of diabetic patients in India and 33.2% in the United States have developed DR. DR is progressive and is associated with the risk of vision loss and even blindness, making it the leading cause of blindness among people of working age.
DR is asymptomatic in its early stages, leading many diabetic patients to neglect regular fundus screening until the disease has progressed to a severe stage, at which point visual function is often difficult to recover. Therefore, early detection and timely treatment are crucial in preventing visual impairment caused by DR. The characteristic pathology of DR includes retinal vascular abnormalities such as microaneurysms, intraretinal hemorrhages, venous beading, exudates, and neovascularization. Based on severity, DR is classified into no apparent DR, mild non-proliferative DR (NPDR), moderate NPDR, severe NPDR, and proliferative DR (PDR). Diabetic macular edema (DME), characterized by the destruction of the blood-retinal barrier and fluid accumulation in the macular area, can occur at any stage of DR and threaten vision. DME is independently graded as mild, moderate, or severe.
The diagnosis of DR and DME is typically based on findings from direct and indirect ophthalmoscopy, slit-lamp biomicroscopy with front lenses, fundus photography (FP), optical coherence tomography (OCT), OCT angiography, fluorescein angiography, and B-ultrasound. Fundus photography is internationally recognized for DR screening and diagnosis, while OCT, although more expensive, is increasingly valued for its detailed imaging capabilities.
Current Status of DR Screening
The American Academy of Ophthalmology recommends that individuals with type 1 diabetes undergo annual eye examinations five years after the onset of diabetes, while individuals with type 2 diabetes should begin annual eye examinations at the time of diagnosis. Screening for DR involves assessing the existence or risk of the disease in asymptomatic individuals, with the goal of improving outcomes through early detection and treatment. In the United Kingdom, a nationwide DR screening program was initiated in 2003 and fully implemented by 2008. From 2015 to 2016, the program screened 82.8% of the 2,590,082 diabetic patients in the country, significantly reducing the incidence of blindness caused by DR.
However, compliance with DR screening recommendations remains poor. In the United States, nearly half of the 298,393 patients diagnosed with type 2 diabetes but without DR had never undergone an eye examination within the past five years, and only 15.3% followed the recommendation of having an eye examination once every two years. Similarly, one-third of patients with type 1 diabetes did not undergo eye examinations, and only 26.3% followed the American Diabetes Association’s screening recommendations. Reasons for poor compliance include a lack of understanding of the disease, poor accessibility of medical resources, and insufficient medical insurance coverage. Studies have shown that patients with more severe DR, impaired vision, and poor blood sugar control tend to have better compliance, indicating that many patients without visual impairment do not realize the importance of regular follow-up.
Telemedicine has emerged as a way to enhance the accessibility of fundus screening. For example, the Singapore Integrated Diabetic Retinopathy Program (SiDRP) allows patients to undergo fundus examinations remotely, with images evaluated by ophthalmologic professionals, significantly reducing medical costs. However, traditional telemedicine still relies on human resources to grade fundus images. The development of artificial intelligence (AI) offers a promising alternative to improve both patient compliance and the efficiency of telemedicine in DR screening.
AI in DR Screening
The concept of AI was first proposed by McCarthy et al. in 1956, followed by Arthur Samuel’s introduction of machine learning (ML) in 1959. Deep learning (DL), a branch of ML, is implemented using multi-layer neural networks and is particularly suited for processing images. Convolutional neural networks (CNNs) are commonly used DL models for image processing, consisting of convolutional layers, pooling layers, and fully connected layers. Common CNN architectures include AlexNet, VGGNet, Inception V1–V4, ResNet, and DenseNet. Transfer learning, another ML method, involves training a model in a source domain and fine-tuning it in a target domain, allowing the model to generalize effectively with a relatively small sample size.
AI has been applied to various image-based medical fields, including radiology, dermatology, pathology, and ophthalmology. In ophthalmology, AI can assist in the diagnosis of DR, glaucoma, age-related macular degeneration, and retinopathy of prematurity. Early AI software was developed to identify specific image features, but with advancements in technology, AI can now learn from large datasets of labeled images. In 2018, the U.S. Food and Drug Administration (FDA) approved the first AI software for DR, IDx-DR, which uses Topcon NW400 to capture fundus images and provides results based on image quality and severity of DR. Other AI systems, such as EyeArt, Retmarker DR, and Airdoc, have also been approved for DR screening.
AI-based diagnostic systems for DR offer several advantages, including high efficiency, high accuracy, and reduced demand for human resources. Studies have shown that AI systems can significantly reduce the workload of manual grading of DR images. For example, Retmarker can reduce the workload of manual grading by 48.42%. AI systems enable patients to undergo fundus screening at primary healthcare clinics, reducing the need for specialized hospitals and improving patient compliance.
Development of AI-Based Diagnostic Systems for DR
The development of AI-based diagnostic systems for DR requires the division of datasets into training, validation, and test sets, which should not overlap. The training set is used to train the algorithm, the validation set is used for parameter selection and tuning, and the test set is used to evaluate the algorithm’s performance in clinical settings. The training set should include a large number of high-quality images labeled by ophthalmologists. According to Chinese guidelines, the training set should consist of fundus photography (FP) images from at least two medical institutions and include images of other fundus diseases in addition to DR. The test set should include at least 5000 FP images, with a significant proportion representing different stages of DR.
Several AI systems have been developed for DR screening, including IDP, IDx-DR, EyeArt, and Google’s algorithm. IDP, an early AI system, identifies characteristic lesions without using DL techniques and has high sensitivity but low specificity. IDx-DR, which incorporates CNNs, improves specificity and has shown satisfactory sensitivity and specificity in clinical studies. EyeArt, the first AI system to detect DR on smartphones, has demonstrated high sensitivity and specificity in remote DR detection. Google’s algorithm allows for adjustable thresholds to achieve the desired sensitivity or specificity, making it suitable for screening when combined with manual grading.
Most AI-based diagnostic systems for DR are based on FP, which has limitations in detecting DME due to its two-dimensional nature. OCT, which provides more detailed imaging, has a higher detection rate for DME. Several AI systems combine OCT and AI techniques to identify DME, showing good sensitivity, specificity, and area under the curve (AUC) values. However, the accessibility of OCT equipment remains a challenge in areas with limited medical resources.
Limitations in Clinical Application
Despite the development of numerous AI-based diagnostic systems for DR, several challenges remain. First, many AI systems use online datasets such as Messidor and EyePACS, which may not be representative of real-world fundus images, leading to potential misdiagnosis. Second, the lack of a unified standard for evaluating AI algorithms makes it difficult to compare the performance of different systems. Third, the “black box” phenomenon of AI systems, where the decision-making process is not transparent, raises concerns about the explainability of AI-based diagnoses. Fourth, the issue of responsibility attribution in cases of misdiagnosis by AI systems remains unresolved. Fifth, AI systems are less reliable in cases of DR complicated by cataracts or other diseases that obscure the media. Sixth, information security is a critical concern when using AI systems for large-scale DR screening. Finally, most AI systems are designed to detect a single disease, limiting their utility in comprehensive eye examinations.
Conclusions and Prospects
AI holds great promise for the screening and diagnosis of DR, with several potential future directions. The development of AI systems based on portable devices such as smartphones could enable patients to undergo DR screening at home, reducing the need for trained medical workers and specialized equipment. This would be particularly beneficial during the COVID-19 pandemic, where telemedicine has become increasingly important. Additionally, AI systems could be expanded to cover more advanced imaging techniques such as multispectral fundus imaging and OCT, improving diagnostic accuracy. AI-assisted diagnosis systems could also support ophthalmologists in making more accurate and efficient diagnoses.
However, to fully realize the potential of AI in DR screening, it is essential to address the limitations of current datasets, establish unified standards for algorithm evaluation, and ensure the explainability and security of AI systems. Combining AI techniques with manual grading may be the most realistic approach in the initial stages of AI adoption in clinical practice.
doi.org/10.1097/CM9.0000000000001816
Was this helpful?
0 / 0