Automated Systems for Calculating Arteriovenous Ratio in Retinographies: A Scoping Review

There is evidence of an association between hypertension and retinal arteriolar narrowing. Manual measurement of retinal vessels comes with additional variability, which can be eliminated using automated software. This scoping review aims to summarize research on automated retinal vessel analysis systems. Searches were performed on Medline, Scopus, and Cochrane to find studies examining automated systems for the diagnosis of retinal vascular alterations caused by hypertension using the following keywords: diagnosis; diagnostic screening programs; image processing, computer-assisted; artificial intelligence; electronic data processing; hypertensive retinopathy; hypertension; retinal vessels; arteriovenous ratio and retinal image analysis. The searches generated 433 articles. Of these, 25 articles published from 2010 to 2022 were included in the review. The retinographies analyzed were extracted from international databases and real scenarios. Automated systems to detect alterations in the retinal vasculature are being introduced into clinical practice for diagnosis in ophthalmology and other medical specialties due to the association of such changes with various diseases. These systems make the classification of hypertensive retinopathy and cardiovascular risk more reliable. They also make it possible for diagnosis to be performed in primary care, thus optimizing ophthalmological visits.


Introduction
Fundus photography, also known as retinography, is a popular imaging technique used to visualize changes in the retinal vessels through the pupil. It can capture changes in vascular caliber and the global geometric patterns of the retina [1]. It is also able to detect signs of retinopathy-such as microaneurysms, hemorrhages, cotton wool spots and hard exudates, and symptoms of the retinal arteriolar wall (e.g., generalized and focal arteriolar narrowing and arteriovenous nicking)-all of which are often observed in patients with systemic diseases, such as diabetes and hypertension.
Regarding diabetes, the 1989 Saint Vincent Declaration set the goal of reducing diabetes-related blindness by one-third over the next five years. This was restated in the 2005 Liverpool Declaration's objective of establishing systematic screening programs to reach at least 80% of the population with diabetes by 2021 [2]. The increased demand for diabetic retinopathy (DR) screening resulting from systematic programs could be met using automated retinal image analysis systems. Such systems can be used in different DR screening scenarios and offer relatively high sensitivity and a substantial reduction in the workload of the health system. Moreover, they are now mature enough to be safely used in DR screening [3,4]. Automated tools improve the quality of DR screening and accessibility to medical care while reducing the cost of the disease by promoting early detection and treatment, which is essential to stop progression [5].
Regarding hypertension, there is evidence that it is associated with retinal arteriolar narrowing. Retinal vessel diameter is expressed as arteriovenous ratio (AVR). According to the Keith-Wagener-Barker classification, AVR values of less than 0.66 reflect hypertensive retinopathy [6].
Arteriolar narrowing is associated with more severe coronary heart disease, stroke, and mortality [6][7][8][9]. There has also been a recent increase in evidence showing that retinal arteriolar narrowing, retinal venular widening, and a suboptimal retinal vascular network are associated with poorer cognitive performance [10][11][12]. Retinal imaging techniques provide unique information about the state of the microvasculature and neuronal structure, different from current neuroimaging markers, such as brain magnetic resonance imaging (MRI), and systemic markers, such as blood pressure. While retinal imaging cannot fully replace PET scans or MRIs in the diagnosis of disease, it does offer a complementary approach to these brain imaging techniques and has considerable potential in clinical and research settings [13].
For all these reasons, retinal imaging can be used as a risk stratification tool because studies suggest that the addition of retinal measures improves the prediction of stroke (an improvement on established risk factors of approximately 10%) [14,15]. Although just a modest improvement in prediction, these findings suggest that adding a combination of various retinal features and/or retinal functional parameters (i.e., a "multimarker approach") may further improve the prediction of dementia and stroke. It might also enable the identification of a more specific subgroup of patients who could benefit from more intensive and expensive examinations, such as brain MRI.
Retinal vascular imaging has also been used to examine the effects of antihypertensive therapy, showing that lowering blood pressure leads to the regression of retinal vascular signs [16,17]. While there have been no significant intervention studies using changes in retinal images as alternative outcome measurements in dementia and stroke, this approach has substantial potential. In addition to its clinical value, retinal imaging may also be a worthwhile research tool in major brain and neurological diseases, such as multiple sclerosis [18][19][20], depression [21,22], and schizophrenia [23,24].
Evaluating retinographies manually implies additional variability in retinal vessel measurements, even when following a standardized protocol. This variability is eliminated if fully automated software is used to measure retinal vascular caliber and other anomalies, although there may be other additional sources of variation in the measurements, such as retinal pigmentation, pupil dilation, the presence of cataracts and other media opacities, photographic technique, type of camera (mydriatic/non-mydriatic or desktop/portable), and image quality (brightness, focus, and contrast) [25]. Manually segmenting vessels, labeling arteries and veins, and localizing the optic disc is a time-consuming task that decreases process efficiency. However, over the past two decades, multiple software systems have been developed to measure and semi-automatically assess the retinal vessel caliber from fundus photographs using artificial intelligence (AI) algorithms [26].
Research question: What is the current stage of implementation of automated retinal vessel analysis systems retinographies? Aim: This scoping review aims to summarize the research available on automated retinal vessel measurement systems so they may be considered in future research and introduced into clinical practice.

Materials and Methods
This review followed the PRISMA extension checklist for scoping reviews [27].

Search Strategy, Data Sources, and Selection
Searches were performed in the Medline, Scopus, and Cochrane electronic databases to locate studies published between 1 january 2004 and 1 september 2022 examining automated systems for the diagnosis of retinal vessel alterations caused by hypertension. The following keywords were used: diagnosis; diagnostic screening programs; image processing, computer-assisted; artificial intelligence; electronic data processing; hypertensive retinopathy; hypertension; retinal vessels; arteriovenous ratio (no MeSH); retinal image analysis (no MeSH).

Selection Criteria
Articles were included in the review if they met the following inclusion criteria: (1) Automated systems were used to partially or totally analyze photographic images of the retina. Only papers written in English were selected. Studies using automated systems for diabetes screening were excluded. Qualitative studies and gray literature were excluded.

Selection of Studies
Abstracts and articles were independently reviewed by two members of the research team based on predetermined inclusion and exclusion criteria. When it was unclear whether an article should be included or some discrepancy appeared, the coordinating researcher of the study also reviewed it.

Data Extraction
Initial data extraction elements included: author, country, research aim, study design, study setting, interpretation system, degree of software automation (semi-automatic, automatic), lesions that the system was able to detect, focus of the photograph and area analyzed, sensitivity, specificity, diagnostic precision, economic evaluation, time savings, local management or the possibility of electronically sending the image to a repository, possibility of comparison for patient follow-up.
Data from each article were independently extracted by two of the authors and then verified by two others.

Search Process
The PRISMA flow diagram ( Figure 1) describes the steps taken to select the articles [28]. The search strategies generated 433 articles, of which 58 full-length articles were evaluated for eligibility. Of these, 25 articles were included in the scoping review.
The PRISMA flow diagram ( Figure 1) describes the steps taken to select the article [28]. The search strategies generated 433 articles, of which 58 full-length articles were eva uated for eligibility. Of these, 25 articles were included in the scoping review.

Characteristics of the Articles
The articles included in the review were published between 2010 and 2022, and th number of retinographies analyzed ranged from 20 to 95,716. These retinographies wer provided by international databases, as well as real scenarios. Table 1 presents a summar of the characteristics of the studies, including first author, year of publication, the countr where the study was conducted, study aim, sample, number of retinographies, name o the software used, degree of automation, and, lastly, the scenario in which it was tested The median number of retinographies analyzed in the included studies was 180, wit a maximum of 54,714 and a minimum of 40. In relation to the degree of automation, in 1 articles they used automatic software and in 15 semi-automated systems.

Interpretation Procedures
The analysis systems included in these articles detect various alterations of the ves sels, including tortuosity, arteriolar and venular caliber, and even AVR calculation. Tabl

Characteristics of the Articles
The articles included in the review were published between 2010 and 2022, and the number of retinographies analyzed ranged from 20 to 95,716. These retinographies were provided by international databases, as well as real scenarios. Table 1 presents a summary of the characteristics of the studies, including first author, year of publication, the country where the study was conducted, study aim, sample, number of retinographies, name of the software used, degree of automation, and, lastly, the scenario in which it was tested.
The median number of retinographies analyzed in the included studies was 180, with a maximum of 54,714 and a minimum of 40. In relation to the degree of automation, in 10 articles they used automatic software and in 15 semi-automated systems.

Interpretation Procedures
The analysis systems included in these articles detect various alterations of the vessels, including tortuosity, arteriolar and venular caliber, and even AVR calculation. Table 2 shows a description of the retinal lesions detected in each study, the focus of the photograph, and the area of the retina that was analyzed. The focus of the image was mainly on the macula and optic disc, and the analyzed area ranged between 2 to 3 radii from the optic disc.    Lastly, the sensitivity, specificity, and diagnostic accuracy were recorded, although this information was only reported in seven articles.
The researchers were initially interested in studying additional data, such as economic evaluation, time savings, local management or the possibility of electronically sending the image to a repository, and the possibility of comparison for patient follow-up. However, the articles selected did not provide this information; therefore, such data were not collected.

Summary of the Results
Most of the articles reviewed fall into two main categories. The first includes publications dealing with automated or semi-automated systems that measure retinal vessels as a diagnostic method for other pathological processes. Retinal vessel measurement can be useful for diagnosing pathologies related to cardiovascular risk [26,34,41] and hypertension [26,33,34,36], dementia and stroke [35], glaucoma [37], chronic kidney disease [11], glycemic control in children [40], myopia [39], and severity of diabetic macular edema [45].
The second category comprises articles aimed at developing automated retinal vessel measurement systems. These articles cover several levels of development: those that use systems limited to vessel segmentation [46,49,51]; those that include vessel labeling and creation of the vascular tree [31,42,45,47,52]; and those that calculate AVR in order to grade hypertensive retinopathy [43,44,49,53].

Discussion
This scoping review aimed to summarize the research available on automated retinal vessel analysis systems in order to determine where automated AVR calculation systems are currently at in terms of implementation. The results indicate that interest in developing technology that facilitates the analysis of the retinal microvascular network has increased over the past eight years. Publications from 2014 and earlier refer to experiments to test the algorithms developed. As of 2015, the software developed from these algorithms has been introduced into clinical practice, demonstrating advantages in real-life scenarios, even though it is not yet widely nor systematically employed. The results of this review confirm that automated AVR calculation systems have not just been introduced as a diagnostic tool for retinal vascular disorders in the field of ophthalmology, but they have been extensively applied in other medical specialties as an accessible and efficient diagnostic tool for other pathologies. Numerous findings indicate that retinal vascular caliber is associated with various systemic diseases, such as hypertension, obesity, diabetes, chronic kidney disease, and stroke. Moreover, since AVR is associated with the development of cardiovascular disease, examining it in this way offers a non-invasive view into the systemic microvasculature.
Hypertensive retinopathy is an indicator of damage to other target organs. However, it is difficult for ophthalmologists to study hypertensive retinopathy in the early stages. Therefore, further research should be conducted on computer-assisted diagnoses that use AVR calculation to automatically detect hypertensive retinopathy and grade it in its early stages [54].
The automation of processes is a prerequisite to improving the affordability, efficiency, and accessibility of these procedures [55] and reducing the high subjectivity of manually assessing AVR [56]. Deep learning methods arise to compute AVR. Convolutional Neural Networks (CNN, Atlanta, GA, USA) obtain a good approximation of AVR value by applying a sequence of spatial filters, subsampling, and non-linear operations.
The clinical applications of artificial intelligence in automated AVR reading could cover a wide range of tasks, including automating hypertensive retinopathy screening, supporting treatment decision-making, assessing systemic vascular status and cardiovascular mortality [57], prescribing medications and diagnostic tests, and creating prognostic models of different diseases [58] to provide more efficient, precise, and sensitive methods in the interpretation of clinical data.
Nonetheless, using artificial intelligence to analyze retinal microvasculature does present some limitations. Firstly, the images used to validate and train the AI may not have enough ethnic variability to provide high external validity because there must be enough, but not too many, images for the processes to be efficient. Secondly, the data should be restricted to those criteria of greatest prognostic relevance, maintaining maximum diagnostic accuracy and minimum algorithm complexity. Consequently, if these processes were automated, they could be implemented in primary care for use by trained health professionals and in rural settings, thus facilitating the classification of cardiovascular risk and reducing the need to refer cases for evaluation by an ophthalmologist. Overall, this would result in the optimization of available health resources.

Conclusions
In recent years, there has been increased interest in developing technology that facilitates the analysis of the retinal microvascular network. Software has been developed and is being introduced into clinical practice not just as a diagnostic tool in the field of ophthalmology, but also in other medical specialties, because there is an established association between various diseases and retinal vessel alterations.
If automated processes for retinal vessel measurement were implemented in primary care for use by trained health professionals, fewer cases would need evaluation by an ophthalmologist, thus optimizing the available health resources.
Moreover, such processes improve the reliability of vasculature measurements, which, in turn, leads to better classification of hypertensive retinopathy by eliminating observer subjectivity and taking cardiovascular risk into account. Also, the more reliable the measurements, the better the early diagnosis of other pathologies, such as dementia and stroke. Further research on the evaluation and implementation of these technologies is needed to recommend their use.