Applications of artificial intelligence-assisted retinal imaging in systemic diseases: A literature review

Abstract

The retina is a vulnerable structure that is frequently affected by different systemic conditions. The main mechanisms of systemic retinal damage are either primary insult of neurons of the retina, alterations of the local vasculature, or both. This vulnerability makes the retina an important window that reflects the severity of the preexisting systemic disorders. Therefore, current imaging techniques aim to identify early retinal changes relevant to systemic anomalies to establish anticipated diagnosis and start adequate management. Artificial intelligence (AI) has become among the highly trending technologies in the field of medicine. Its spread continues to extend to different specialties including ophthalmology. Many studies have shown the potential of this technique in assisting the screening of retinal anomalies in the context of systemic disorders. In this review, we performed extensive literature search to identify the most important studies that support the effectiveness of AI/deep learning use for diagnosing systemic disorders through retinal imaging. The utility of these technologies in the field of retina-based diagnosis of systemic conditions is highlighted.

INTRODUCTION

Systemic diseases encompass a wide range of medical conditions affecting various organ systems in the body. Some of these diseases, such as metabolic disorders (e.g., diabetes) and autoimmune disorders (e.g., lupus), can involve the retina, leading to retinal manifestations that can significantly impact visual function.[1,2] The retina, being an extension of the central nervous system, is highly susceptible to the effects of systemic diseases. Retinal manifestations of systemic diseases can occur earlier in the disease course compared to symptoms affecting other organs, making the retina a valuable window for early disease detection.[3] In some cases, retinal manifestations may remain subclinical. This subclinical nature poses a challenge for the timely diagnosis and management of systemic diseases. However, the retina often exhibits specific or pathognomonic signs that are characteristic of the underlying systemic condition, providing valuable diagnostic clues.[4]

The advent of artificial intelligence (AI) and deep learning algorithms has revolutionized medical image analysis, including the analysis of retinal imaging.[5] AI techniques can effectively extract and analyze intricate details from retinal images, enabling the identification of subtle retinal manifestations that may not be discernible to the naked eye.[5] This advanced imaging analysis using AI can prompt the screening and diagnosis of systemic diseases at their subclinical phase, even before the onset of noticeable symptoms or signs.[6,7] Thus, AI has the potential to facilitate risk prediction of systemic diseases by analyzing retinal changes.[8,9,10] Moreover, by integrating the data from patients with systemic diseases such as diabetes mellitus, including mainly data of their retinal imaging findings, AI algorithms can generate predictive models to identify individuals at higher risk of developing retinal manifestations.[11,12] This early detection can enable timely interventions, therefore, can improve patient's outcomes.

The early detection of retinal manifestations using AI and deep-learning algorithms holds profound implications for the prognosis and treatment of both ophthalmological and systemic aspects of the disease. From an ophthalmological perspective, early identification of retinal involvement allows for timely intervention, preventing irreversible damage to retinal structures and preserving visual function.[13] On a systemic level, the prompt diagnosis of systemic diseases through retinal imaging can lead to early commencement of treatment, targeting not only ocular manifestations but also other organs involved in the disease.[10] This timely initiation of therapy may positively impact both functional (vision-related) and organic (histological) outcomes, optimizing patient care and potentially improving long-term prognosis.

While several reviews have explored the applications of AI-assisted retinal imaging (AI-ARI) in retinal diseases, synthetic articles that addressed its potential applications in systemic diseases are scarce. The present review explored the various applications of AI-ARI in the management of systemic diseases. By synthesizing the current knowledge and future perspectives, we aim to promote research and clinical applications of AI-based retinal imaging analysis beyond the retinal diseases, specifically in the diagnosis and management of systemic diseases affecting the retina. Such an insight fosters multidisciplinary care practices and enhances the role of the ophthalmologist in a holistic and patient-centered approach. Ultimately, the integration of AI technologies has the potential to improve patient outcomes and shape the future of systemic disease management.

METHODOLOGY

An extensive literature search was conducted mainly via the databases “PubMed,” “Google Scholar,” and “Google.” The following search terms as well as their derivatives were entered in different combinations: AI, deep learning, neural network, retinal imaging, retinal abnormalities, diabetic retinopathy (DR), hypertensive retinopathy (HR), cardiovascular disease (CVD), coronary artery disease, neurodegenerative disease, neuropsychiatric disease, multiple sclerosis (MS), Alzheimer's disease, Parkinson's disease, chronic kidney disease (CKD), hematological diseases, and systemic vasculitis. Studies were included if they showed relevant results.

RETINA AS A VULNERABLE STRUCTURE TO SYSTEMIC DISEASES

Retina, the innermost layer in the eye, is formed by ten distinct layers of neurons interconnected by synapses. The retinal cells have the main job of capturing photons by photoreceptors cells (rods and cones) and transmit it into action potentials that the brain's cortex processes into the three-dimensional vision.[14] The retina is supplied by two vascular systems: the central retinal artery and the choriocapillaris. The former originates from the ophthalmic artery, itself branch of the internal carotid artery, and further divides into superior and inferior arcades (which form the blood-retina barrier) to supply the inner retina. Whereas the choriocapillaris a richly anastomotic, fenestrated group of capillaries which supplies the retinal pigment epithelium and the layer of photoreceptors (outer retina) is originated from the long and short ciliary arteries, issued from the ophthalmic artery.[14,15,16] During systemic diseases, the retina can be damaged through multiple mechanisms. First, it can be affected by primary neurodegeneration as a result of metabolic stress such as in DR, neuroinflammation such as in MS or infection by a pathogen with possible retinal tropism such as in tuberculosis. Second, the retinal tissue can be prone to hypoxia and subsequent ischemia as a result of disruption of the local vasculature which can be the site of inflammation such as in systemic vasculitides or obstruction/remodeling such as the case of atherosclerosis and chronic hypertension. Third, it may result from the two mechanisms combined such as in DR which regroups both anomalies of cellular metabolism due to sorbitol pathway dysregulation and vascular damage due chronic hyperglycemia.

CURRENT RETINAL IMAGING TECHNIQUES

The abnormalities of the retinal tissue can be initially visualized by 2-Dfundus imaging techniques including fundus photography, color fundus photography (CFP), stereo fundus photography, hyperspectral imaging, scanning laser ophthalmoscopy (SLO), and adaptive optics SLO. However, optical coherence tomography (OCT) provides more refined 3-D views of the retinal layers with great estimation of their depth; therefore, it has become the technical imaging of choice.[17] The retinal circulation fundus can be visualized by fluorescein angiography, nonetheless, this is an invasive diagnostic procedure that has widely been replaced by OCT angiography (OCTA). The latter offers a less invasive option for detecting different anomalies of microvasculature during perfusion-deficit states such as sepsis, systemic hemodynamic disturbances such as ischemic CVDs, and systemic inflammatory conditions such as inflammatory bowel disease.[17,18]

IMPLICATIONS OF ARTIFICIAL INTELLIGENCE IN RETINAL IMAGING

The main idea of involving AI in retinal imaging is to transfer a data containing a high number of fundus photography or OCT images obtained from patients with retinopathies into an AI or deep learning system. The system will then develop its own algorithm in a way that whenever it is exposed to similar images can discriminate by itself the images containing the prelearned retinopathies features. Dong et al. developed a deep learning system to identify 10 retinal diseases using data from 120,002 ocular fundus photographs. For prospective evaluation, the system was exposed to a total of 208,758 images obtained from 110,784 individuals. The results showed a sensitivity of 89.8% to detect all 10 retinal diseases and diagnostic accuracies ranging from 95.3% to 99.9%. Moreover, the system reached higher or comparable sensitivity in identifying 7 of 10 retinal diseases including, glaucoma, macular hole, epiretinal macular membrane, HR, myelinated fibers, and retinitis pigmentosa.[19] Additionally, in a large national real-world evidence study, deep learning system named Comprehensive AI Retinal Expert (CARE) system was trained to detect 14 most common retinal abnormalities (such as DR, HR, Glaucomatous optic neuropathy, retinal vein occlusion, pathological myopia and others) through its exposition to 207,228 color fundus photographs, derived from different clinical centers across China with different disease distributions. After evaluation, the CARE system had a similar performance to that of ophthalmologists in different regions with varying experience.[20] Similarly, deep learning system could interpret 71,896 images of 14,880 patients with different retinal conditions including DR with a sensitivity of 90.5% and specificity of 91.6%, possible glaucoma with a sensitivity of 96.4% and specificity of 87.2%, age-related macular degeneration with a sensitivity of 93.2% and specificity of 88.7%.[21] Therefore, based on promising previous results, the automatic diagnostic potential of AI is being increasingly investigated in the field of ophthalmology in order to fully determine its utility that seems to be high and validate the use of this technique in the clinical practice.

ARTIFICIAL INTELLIGENCE APPLICATIONS IN SYSTEMIC DISEASE WITH RETINAL MANIFESTATIONS

Diabetic retinopathy

DR is one of the major chronic complications that can significantly compromise the visual function of diabetic patients, affect their quality of life, and is associated with increased risk of mortality by CVD.[22,23] According to a population-based study, the prevalence of DR after 15 years of diabetes reaches over 77.8%.[24] Hence, since DR is the most common and leading cause of blindness among patients with diabetes, ongoing research focuses on providing novel disease insights, preventive and therapeutic strategies, as well as diagnostic tools. Novel cost-effective methods of diagnosis are needed to enhance the effectiveness of screening and early diagnosis of this condition as even with sophisticated imaging techniques such as OCTA, the proportion of DR patients that can be missed remains significant.[25] On the other hand, a growing body of evidence is supporting the ability of AI inaccurately diagnose DR and reduce the risk of false-negative patients. Thus, a large study of 4997 adults with diabetes showed the high sensitivity and specificity of deep learning machine in identifying referable DR after accurately interpreting 9963 images of retinal fundus photographs.[26] Moreover, recent study by Lim et al. tested the superiority of AI in detecting more than mild DR (mtmDR) in comparison to ophthalmologists or retina specialists among 521 diabetic patients. They found that the AI system (EyeArt System) displayed higher sensitivity in interpreting the dilated ophthalmoscopy findings significantly outperforming the included physicians.[27] Another prospective multicenter cross-sectional diagnostic study carried by Ipp et al. including 893 diabetic participants reported that EyeArt system is a safe and accurate tool for the automated detection of both mtmDR and vision-threatening DR, without physician assistance.[28] Similarly, the safety and reliability of the AI technology CARE was observed by Dong et al. in a multicenter study involving the Chinese community.[29] In an additional prospective evaluation, Heydon et al. studied the sensibility of EyeArt algorithm in screening for retinopathy among 30,000 diabetic patients. Thus, the sensitivity was 98.3% (97.3%–98.9%) for mild-to-moderate nonproliferative retinopathy, while it reached 100% (98.7%, 100%) for moderate-to-severe nonproliferative retinopathy and 100% (97.9%, 100%) for proliferative retinopathy.[30] Besides potential of accurate diagnosis, AI can also allow the early detection of DR. Bora et al. demonstrated the ability of deep-learning system in predicting the risk of developing DR using color fundus photographs.[11]

Hypertensive retinopathy

HR is believed to be directly caused by vascular changes in retinal vasculature as response to systemic elevated blood pressure. Hypertension can lead first to vasospasm and increased vascular tone of the retinal arterioles which manifests as generalized narrowing of the local arteriolar vasculature, then, in advanced stages, the intimal wall of the retinal arterioles becomes thick and hyaline degeneration starts to occur, resulting in definitive narrowing and arteriovenous nipping.[31] This causes multiple pathological features including opacification of the vessel wall (copper or silver wiring), areas of retinal tissue ischemia, edematous and nonedematous exudation, intraretinal hemorrhages, and microaneurysms.[31] The diagnosis of HR can be challenging for ophthalmologists as it requires certain tiresome tasks such as the analysis of retinal vasculature during fundus imaging through manual segmentation.[32] Moreover, as HR can progress insidiously, diagnosing its earlies features at preclinical stage is difficult. In two studies conducted by Arsalan et al. AI systems have shown high sensitivity, specificity and diagnostic accuracy in identifying HR with potential of detecting pathological changes in the minor vessels which ensures the least false negatives results.[32,33] Krismono Triwijoyo and Pradipto suggested a deep learning-based method for early detection of HR. The approach uses Deep Neural Networks and Boltzmann Machines to analyze the features of artery and vein diameter ratio combined with changes position with Optic Disk in retinal images, and classify the severity of the HR.[34] Recently the automatic detection of HR was demonstrated via transfer learning-based network. The latter after receiving information from more than 6000 retinal images from MESSIDOR and ODIR datasets, achieved the classification accuracy of 98.72%.[35] In a preventive approach, several other authors demonstrated the utility of AI-assisted ocular imaging in predicting systemic systolic and diastolic blood pressure; however, the accuracies were disappointing, with area under the Receiver Operating Characteristic curve (AUROC) values ranging from 0.17 to 0.40 for systolic blood pressure and 0.16 to 0.35 for diastolic blood pressure.[9]

Cardiovascular disease

Retinal microvasculature can provide insights into an individual's overall cardiovascular health. Mechanistically, changes in retinal microvasculature features, such as changes in vessel caliber, can be used as biomarkers for cardiovascular conditions. By applying AI to analyze retinal images, these changes can be identified and quantified more accurately and efficiently than by human experts alone.[36] This can contribute to the prediction of cardiovascular risk factors, such as blood pressure, heart failure, and diabetes, as well as the prediction of direct cardiovascular events, such as coronary artery abnormalities, major adverse cardiovascular events (MACE) and mortality.[9,36,37] This could supplement the current risk stratification approaches with more accurate estimates and prediction, leading to more efficient strategies and improved patient outcomes.[37]

While evidence showed low utility of AI-ARI in predicting systolic and diastolic blood pressures, the prediction of hypertension, MACE, and carotid artery calcification showed promising results with AUROC values up to 0.77, 0.70, and 0.83 respectively.[9] Further applications were reported in a review that synthesized evidence on the utility of AI-aided retinal imaging in both ischemic (such as stroke and coronary artery disease) and nonischemic CVD (such as heart failure and hypertension). It also demonstrated other applications in CVD such as the automated methods for biomarkers extraction and detection of cardiovascular risk factors. These investigations utilized diverse datasets from multiple geographical locations, such as Singapore, Sydney, Melbourne, China, and the UK, and employed a variety of machine learning architectures with various preprocessing techniques, including data augmentation and mean-filtering. Notably, AUROC values achieved in these studies were quite promising, ranging from 0.80 to 0.97. Additionally, the accuracy metrics varied between 0.75 and 0.91. Overall, these findings underscore the increasing diversity and potential efficacy of applying AI algorithms to retinal imaging for CVD prediction and management.[38]

Additionally, retinal images interpretation by deep learning was proposed as a valid prognosis marker for hypertension and ischemic heart disease.[39] Deep learning model was trained by Rim et al. to assess the presence of coronary artery calcifications by analyzing retinal photographs. Remarkably, the predictive value of the tested model was comparable to that of computed tomography (CT) scan-measured coronary artery calcifications in stratifying the risk of cardiovascular events.[40] A recent meta-analysis of 26 studies has revealed the ability of deep learning-ARI to predict 42 CVD risk-related outcomes. These included CVD risk factors, cardiac imaging biomarkers, CVD risk scores, the presence of CVD, and incident CVD.[40,41]

Nonetheless, there exist several technical and economic challenges that must be overcome to fully leverage this technology and derive widespread public health benefits.

Neurodegenerative and neuropsychiatric diseases

Over recent years, advancements in ocular imaging, particularly with tools like OCT, have not only facilitated better patient care in cases of optic neuropathies but also offered valuable insights into the progression of neurological and neurodegenerative diseases. The fusion of AI and deep learning systems with these state-of-the-art imaging modalities promises to further elevate the precision and efficiency of disease detection and prognosis.[42]

MS is a neurodegenerative disorder that leads to loss of myelin sheath in the central and peripheral nervous system neurons with subsequent neuronal death. The neuronal structures in retina are degenerated during MS through multiple mechanisms including primary retinal degeneration, trans-synaptic degeneration related to lesions in the posterior afferent visual pathway and retrograde degeneration which is the consequence of optic nerve inflammation.[43] In a study, the learning algorithm model, Support Vector Machine, was able to identify structural neurodegeneration in the retina of MS patients from swept-source OCT (SS-OCT) data with sensitivity of 89%, specificity of 92%, and accuracy of 91%.[44] In a similar study, the use of machine learning allowed the prediction of MS relevant retinal anomalies including thickness loss in the retinal nerve fiber layer (RNFL) and the complex ganglion cell layer–inner plexiform layer (+), by analyzing SS-OCT data.[45] AI-assisted detection of RNFL loss in MS patients was also demonstrated by applying artificial neural network technique.[46]

The detection of symptomatic Alzheimer's disease was also possible by deep learning using convolutional neural networks to multimodal retinal images obtained from affected and healthy individuals.[47] In a retrospective multicenter case–control study, a deep learning model was trained by 12949 retinal photographs from 648 patients with Alzheimer's disease and 3240 healthy controls. Interestingly, authors reported 83.6% accuracy, 93.2% sensitivity, and 82.0% specificity in detecting Alzheimer's disease dementia.[48] Another retrospective study including 353,157 patients supported the utility of deep learning models in helping the screening for all-cause dementia by retinal OCT imaging. Thus, in this study, the combination of OCT layer thicknesses (interpreted by deep learning) with nonretinal risk factors significantly increased the AUROC for dementia prediction.[49] Similarly, deep learning assessment for retinal age gap was shown to be a potential predictor of Parkinson's disease among high-risk individuals. In particular, Hu et al. showed in their study on 35,834 participants the ability of deep learning-based assessment of retinal age gap in identifying at risk individuals for Parkinson's disease as 1 year increase in retinal age gap was linked with 10% higher risk to develop the disease. Interestingly, retinal age exhibited a predictive value comparable to that of the known Parkinson's disease risk factors.[50]

Psychiatric disorders are another potential field of interest for AI and deep learning-mediated diseases prediction from retinal imaging.[51,52,53] It has been recently shown that convolutional neural networks, when trained to interpret retinal fundoscopy images, accurately classified schizophrenia patients and normal subjects with an AUROC of 0.98.[51] Moreover, the results of a small study indicated the utility of AI-aided discrimination of retinal features (particularly those of the inner nasal region and the outer nasal and inner inferior regions) in diagnosing patients with bipolar disorder.[52] Automatic detection of autism spectrum disorders by retinal imaging and machine learning was also demonstrated to be a possible approach yielding a sensitivity of 95.7%, a specificity of 91.3% and an AUROC of 0.974.[53]

Chronic kidney disease

CKD is associated with a higher risk of retinal disease and vice versa.[54] This bidirectional relationship between retinopathy and CKD, suggests that presence of common pathogenic pathways including cardiovascular risk factors, diabetes, inflammation, oxidative stress, endothelial dysfunction, and microvascular dysfunction.[54] One study showed that 16.7% of patients with stage 3-5 CKD develop retinal anomalies including retinal hemorrhage, microvascular and DR, and macular degeneration,[55] and another study revealed that retinovascular pathology reflects renal disease.[56] Thereby, current studies support the use of retinal imaging to screen for renal function alterations. To do so, AI machine was employed to detect CRD from retinal images. Sabanayagam et al. found deep learning algorithm able to estimate CRD from retinal images and suggested that it can be an effective screening tool for CKD in community populations.[57] Moreover, applying deep learning to process retinal photographs had led to an accurate prediction of creatinine levels.[10] Retina-based deep learning was further shown to be able to execute riskstratification for CRD with a superior effectiveness than that of conventional eGFR-based methods.[58]

Hematological diseases

The potential of AI to enhance the screening, diagnosis, and treatment of hematological malignancies has been recognized, offering prospects of improved patient outcomes and decreased mortality rates. Despite this promise, certain limitations in the current AI applications within hematology have been observed, including inadequate data availability.[59] In order to address this limitation, an effective strategy could be the utilization of retinal images to supplement the data fed into machine learning algorithms, thereby bolstering the data pool and improving model performance.

Retinal imaging can reveal subtle changes indicative of hematological disorders. For instance, several alterations in the retinal vascular system, such as altered vessel density, diameter, or complexity have been observed in OCTA of patients with acute leukemia, and seemed to have a prognostic value.[60] Other abnormalities that can be reflective of hematological diseases are optic nerve and retinal infiltrations, which were observed in one-third of the patients on conventional imaging techniques.[61] Moreover, both hemoglobin concentration and anemia diagnosis can be accurately and effectively screened by deep learning-mediated analysis of fundus photographs.[62,63] This approach might be useful among diabetic patients who undergo routine retinal imaging and who have particular morbidity and mortality risks linked to anemia.[63]

The use of AI and machine learning algorithms represents an opportunity to enhance the quantification and classification of these abnormalities and establish their clinical value. Furthermore, AI and machine learning can be trained to create predictive models, learning from retinal image patterns associated with different stages of hematological diseases. This predictive analysis could enable the identification of individuals at risk of disease progression, thereby facilitating early interventions and potentially improving prognosis. Regular AI-ARI can also play a crucial role in monitoring the effectiveness of treatments. By providing real-time data on the impact of therapeutic interventions on the retina's microvasculature, this technology offers a noninvasive method to assess and adjust treatment plans.

Other systemic diseases

A part from the above-mentioned disorders, the utility of AI/deep learning models has been investigated in diagnosing several other systemic diseases. For instance, a cross-sectional analysis involving 61,718 individuals from the UK demonstrated that the use of deep learning algorithms to interpret retinal CFP and OCT images has enabled the identification of multiple significant associations between retinal biomarkers and a wide range of systemic diseases. Among these, renal, cardiovascular, and endocrine diseases were the most common.[64] Predicting Behcet's disease by AI-based algorithms was also shown in a previous study carried by Güler and Ubeyli, reporting a diagnostic accuracy <90%.[65] Moreover, AI-assisted retina analysis of SS-OCT from patients with fibromyalgia showed high discriminative capacity in automatic detection of retinal structures (retinal thicknesses) anomalies due to this disorder.[66]

Further applications in systemic context

In addition to diagnosing systemic diseases, AI-ARI can serve a valuable role in quantifying and predicting physiological parameters. A pioneering study showed that deep learning could quantify composition indices including muscle mass, height, and bodyweight from retinal photographs based on data of 236 257 retinal photographs from seven diverse Asian and European cohorts.[10]

Nonetheless, one challenge remains regarding the applications of AI in retinal imaging, and in imaging in general, which is data poverty. The insufficiency of diverse, representative datasets impairs the generalizability and reliability of AI and machine learning algorithms, risking diagnostic inaccuracies and suboptimal treatment plans. For instance, if the training data primarily represent one gender, AI algorithms may underperform in detecting retinal signs of systemic diseases like diabetes or hypertension in the other gender. This perpetuates and exacerbates existing health disparities, creating a digital divide in healthcare access and outcomes. Gender-related and racial disparities in health data serve as obstacles to achieving equitable AI-based solutions.[67] Addressing this data poverty through the collection of inclusive, representative datasets, coupled with robust governance and transparency, is pivotal for leveraging AI's full potential in retina imaging for systemic disease management.

CONCLUSION AND FUTURE DIRECTIONS

AI-based systems are noninvasive, effective tools that can significantly optimize the diagnosis and management of a wide range of systemic diseases with retinal involvement, as well as retinal complications in the context of systemic disorders [Figure 1]. Beyond diagnosing early retinal changes with precision, the primary advantages of AI/deep learning algorithms include the early identification of individuals at risk for systemic diseases and the establishment of novel biomarkers and risk scores. However, as with any emerging technology, especially one that is proliferating as rapidly as AI, it is imperative to deepen our understanding of the reliability and clinical applicability these digital tools offer in detecting systemic disorders through retinal imaging changes. In addition, the notion that these systems can outperform ophthalmologists might lead to differing opinions and acceptance among ophthalmology specialists, as well as trust and privacy concerns among patients. Thus, it is crucial to thoroughly evaluate their advantages and disadvantages before fully integrating them into clinical settings.

Figure 1.

Applications of artificial intelligence-assisted retinal imaging in systemic diseases

Financial support and sponsorship

Nil.

Conflicts of interest

There are no conflicts of interest.