Is it time to consider teleophthalmology as a game-changer in the management of diabetic retinopathy?

ABSTRACT Currently the “pandemic” of diabetes mellitus is noted. The incidence and prevalence of diabetes and diabetic retinopathy, the most common microvascular complications of diabetes, are exponentially growing due to increased life expectancy in many parts of the world. The increasing number of people suffering from diabetic retinopathy not only highlights medical issues, but also an economic burden, representing a medical and social challenge. It is extremely important to identify a disease as soon as possible and successfully treat it. Technological progress results in developing Artificial Intelligence systems capable of detecting diabetic retinopathy. Current screening will be cost effectively based on the use of advanced digital technologies, in particular teleretinal screening systems. At present, we may consider teleophthalmology and Artificial Intelligence with automatic analysis of fundus photos as a Millennium-minded impactful tool for increasing discoverability and manageability of diabetic retinopathy, especially in filling the gap of inaccessibility to hard-to-reach areas, which enforces highly professionally effective time- and cost-saving care everywhere to provide the best possible care for the patients.


INTRODUCTION
Diabetes is "a pandemic of unprecedented magnitude". (1)The incidence and prevalence of diabetes and diabetic retinopathy (DR), as the most common microvascular complication of diabetes, (2) is exponentially growing due to increased life expectancy in many parts of the world.According to The International Diabetes Federation, currently 537 million adults (10.5% of the world's adult population) have diabetes, reflecting a 16% increase comparing to findings in 2019, (3) estimated to rise to 783.2 (12.2%) million by 2045. (4)DR is the most challenging cause of visual impairment and blindness worldwide. (5)he forecast for the year 2045 indicates that the number of patients with DR will increase more than 50%, reaching 160.5 million with an alarming surge of vision-threatening diabetic retinopathy (VTDR) prevalence by 57.0% to 44.82 million from 2020 to 2045. (6)It must be taken into consideration that these "pandemic" numbers will cause a medical and social challenge with an economic burden, which underscores the significance of screening (7 ), providing a window of opportunity for the prevention of vision loss due to DR.
Current screening will be cost effectively based on the use of teleophthalmology-telemedicine in ophthalmology, specifically "asynchronous type of telemedicine with the 'store-and-forward' approach where investigations are captured and uploaded onto cloud servers, and can be analyzed by artificial intelligence (AI) algorithms, to determine diagnosis and management". (8)The term "artificial intelligence" (AI), which means machines capable of performing tasks that typically require human intelligence, was introduced by John McCarthy in 1956. (9,10)Machine learning (ML), as the ability of giving a computer program the skill to learn from examples without being explicitly trained to do so, have shown promising results in image evaluation starting from 1950, (11) but it is worth noting that early imperfect hardware restricts its clinical use for medical image analysis.A lot has changed since then. (12)In 2012, Deep Learning (DL), which is a subset of ML aimed at making decisions using specific algorithm-neural networks through automatically extracted useful pieces of information, was really introduced to the world. (13,14)Technological progress results in developing AI systems capable of detecting DR.
The objective of this review is to evaluate the evidence and discuss the rationale behind the recent suggestions on the role of telemedicine in the management of DR worldwide.based on the currently available findings.

METHODS
For this review, a literature search was conducted using PubMed®/Medline® and Google Scholar for studies published up to November 2022 inclusive.The following keywords were used in various combinations: "diabetic retinopathy", "teleophthalmology", "artificial intelligence", "machine learning", "deep learning".Articles with high or medium clinical relevance, including also manually searched from the references, were selected for this review.

RESULTS
We found that studies have demonstrated the suitability of AI in the management of DR.
In 1996, Gardner et al. (15) presented the use of an artificial neural network in DR detection based on the recognition of vessels, exudates, and hemorrhages with rates 91.7%, 93.1%, and 73.8% respectively, reaching 88% sensitivity and 83.5% specificity, as compared to a trained ophthalmologist.Researchers concluded that "The system could be used as an aid to the screening of diabetic patients for retinopathy", acting as a starting point for new researchers to work on the development of ML systems for DR detection.
In 2016, Abràmoff et al. (16) continued refinement of diagnostic algorithm incorporating a DL component and evaluated its performance on 874 subjects.The images have been graded as no DR in case of no DR or mild DR without diabetic macular edema (DME), and referable DR.Referable DR was diagnosed as moderate non-proliferative DR, severe non-proliferative DR, proliferative DR, and/or DME; vision-threatening DR was defined as severe non-proliferative DR, proliferative DR, and/or DME.It was shown that proposed enhanced algorithm was capable of detecting referable DR with 96.8% sensitivity and 87.0% specificity.The authors reported that no cases of severe non-proliferative DR, proliferative DR, and/or DME were missed.
Further assessment and validation of DL algorithm in DR and DME detection based on 128,175 retinal images were conducted by Gulshan et al. (17) It has been reported that tested algorithm achieved a high sensitivity and specificity in grading referable DR.Despite this, however, the authors of the study concluded that it requires further scientific investigation in clinical settings.
Another study initiated by Ting et al. (18) was focused on assessing DL system capability of detecting not only referable DR, but also VTDR, and at the same time eye diseases, as an age-related macular degeneration and glaucoma in patients with diabetes.This was a large scale study based on the analysis of 112,648 retinal images from multiethnic populations, which have evidenced a high sensitivity (90.5%) and specificity (91.6%) for identifying referable DR, and 100% and 91.1% for VTDR, respectively.The similar efficacy were reported for other eye comorbidities.Based on these findings it was also highlighted that it is necessary to have a system verification at healthcare settings.
It is noteworthy that Lee et al. (32) have validated seven DR screening algorithms based on 311,604 images evidencing wide fluctuations in sensitivity.The authors emphasized a need for meticulous testing of all algorithms before clinical implementation.The latest validation retrospective study conducted by Zhang et al. (37) underscored that developed automatic DL system for referable DR detection is suitable for real-word screening.
However, despite multiple credible studies, only in 2018 the Food and Drug Administration (FDA) approved the first autonomous AI-enabled medical device for the detection of DR-Digital Diagnostics' IDx-DR system, (38) connected to the Topcon NW400 non-mydriatic fundus camera and allowing to detect more-than-mild DR (mt-mDR) at primary healthcare setting without the eye specialist's image grading. (19)Diagnostic study in a cohort of 819 participants have evidenced 87.2% sensitivity, 90.7% specificity, with 96.1% imageability rate after enrichment correction.
For the first time, the IDx-DR system was prospectively evaluated in 310 youth (5-21 years old) patients with DR. (39) It has been reported that there is 85.7% sensitivity and 79.3% specificity in detection of mtmDR.The latest retrospective validity and reliability study based on the IDx-DR system (40) has shown 100% sensitivity (95% of confidence interval [95%CI] 90.8-100%) and 89.2% specificity (95%CI 87.0-91.1%)for referable DR, at the same time indicating gradeability decrease with ageing.
To date, the second FDA approved fully autonomous for mtmDR detection and the first for mtmDR and VTDR, commercially available system is the EyeArt system (Eyenuk, Inc). (33,41)According to a prospective multicenter cross-sectional study (33) the EyeArt system's version 2.1.0has shown 96% sensitivity, 88% specificity, and 97% imageability for detecting eyes with mtmDR, and for the first time for VTDR -95.1 and 89.0%respectively in nondilated cases.Despite this, however, the study authors reported that imageability increased after dilation in cases of ungradable results, reaching 97.4%.It should be emphasized that the system used is reliable in detecting, not only mtmDR, but also of VTDR, as was stated by researchers.
Another prospective study in three English diabetes screening programs based on the same software (31) evidenced 98.3% sensitivity for mild-to-moderate non-proliferative retinopathy with referable maculopathy, 100% for moderate-to-severe non-proliferative retinopathy and 100% for proliferative retinopathy respectively.Further supportive data for this notion were obtained by Lim et al., (34) where it was shown higher sensitivity of the EyeArt system comparing to general ophthalmologists or retina specialists in prospective, pivotal, multicenter trial conducted from April 2017 to May 2018.In contrast to the IDx-DR system, the EyeArt system could detect not only mtmDR, but also VTDR simultaneously.

DISCUSSION
Currently AI-based DR screening systems have been evaluated in multiple prospective studies with patients of different races and ethnicities (Table 2).The results obtained earlier by Kanagasingam et al. (42) in evaluating AI-based grading system in a primary care office in Australia with a small sample size revealed a high false-positive rate with a positive-predictive value of 12%, accordingly from 17 patients with severe DR, and 15 patients were false positives.In contrast, in another study from Australia, (30) also with a small sample size, high sensitivity (96.9%) and specificity (87.7%) were reported with a positive experience feedback from the patients and healthcare workers.These results supported the studies previously conducted by Gulshan et al., (24) Bellemo et al., (25) Sosale et al. (43) Another prospective study in three English diabetes screening programs was described above. (31)The latest largest prospective interventional cohort study was conducted by Ruamviboonsuk et al. (44) at nine primary care units in Thailand.This study adds more to the body of knowledge demonstrating DL capability of diagnosing DR similar to retina specialists.In another new prospective, multicenter study (45) in China, a high efficacy of referable DR detection was shown, except for maculopathy.The recent randomized trial (46) evidenced that patients with diabetes, who underwent AI-based screening, sought treatment faster comparing to the control group.likely ready for clinical applications".However, it is worth noting that recently Nakayama et al. (61) discussed the utilization of supervised ML algorithms for DR classification and have emphasized a need for the implementation of versatile comprehensive classification with equal referral criteria.
77)(78) Further supportive data for this notion were obtained by Malerbi et al. (79) in the retrospective study that evaluated the capability of semiautomated DR screening with mobile handheld retinal camera to diagnose mtm-DR and VTDR after pupil dilation in 824 patients.It has been reported that the system achieved 97.8% sensitivity and 61.4% specificity, maintaining a sufficient quality of images in more than 80% of the cases.The obtained results are promising, and they deserve further scientific investigation.
It is worth noting a prospective, open study initiated by Gobbi et al. (80) at the DR screening clinic of Hospital das Clínicas de Ribeirão Preto, in Brazil in an effort to assess suitability of smartphone-based DR photoscreening in ninety-nine diabetic patients (194 eyes) conducted by undergraduate medicine and nursery students unexperienced in retinal imaging, who simulated real-word screening at primary healthcare level.It has been postulated that specificity tends to be greater than sensitivity, 0.94 and 0.71 respectively detecting any level of DR; 0.99 and 0.76 respectively detecting proliferative DR; 0.94 and 0.72 respectively detecting macular exudates.The researchers concluded that "the smartphone-based device showed promising accuracy in the detection of DR (84.07%), making it a potential tool in the screening and early diagnosis of DR".
There is still room for improvement. (81,82)(85)(86)(87)(88)(89)(90)(91)(92)(93) The recent one presented by Zhang et al. (83) is based on Deep Graph Correlation Network for grading.Researchers postulated that it has an accuracy near to retinal specialists and more than trained graders.Further improvements are presented by Canayaz, (84) Hu et al., (85) Datta et al., (86) Hassan et al., (87) Venkaiahppalaswamy et al. (88) The efforts have been made to develop the hybrid DL system with a segmentation of optic disc and blood vessels. (88)It has been reported that the system achieved 94% accuracy All the reports described above and the findings from other prospective studies (19,21,27,33,39) underscore the potential utility of AI based system in DR screening.
Recent reviews and surveys as those of authors like Fenner et al., (47) Ting et al., (48) Asiri et al., (49) Grzybowski et al., (50) Stolte et al., (51) He et al., (52) Bilal et al., (53) Williamson, (54) Lalithadevi et al., (55) Iqbal et al., (56) Celard et al. (57) and Vujosevic et al. (58) cover a significant amount of works that apply DL to retinal image analysis in order to diagnose DR.A metaanalysis conducted by Wu and coworkers in 2019 (59) and 2021 (60) evidenced a high efficacy of ML algorithms in DR detection, specifically mtmDR.The authors concluded that "ML-based DR screening algorithms are with capability of detecting early DR.Further research directed to detect retinal vessels were continued by Jiang et al. (89) and Arsalan et al. (90) .According to developers (90) , the technology named pool-less residual segmentation network (PLRS-Net) showed to be promising in terms of detecting retinal vessels abnormalities in DR and hypertensive retinopathy.Researchers have found that it has a high accuracy, sensitivity, and specificity, achieving 96.82%, 82.69%, 98.17% respectively and stated that "These accuracies show exceptional segmentation performance of the proposed method compared to stateof-the-art approaches for automatic vessel detection for diagnosis purposes".
Tokuda et al. (91) have selected retinal hemorrhage as a marker for DL-based retinal image analysis in DR.The authors evaluated 70 fundus images of 70 patients categorizing DR as a mild-or-worse non-proliferative DR and moderate-or-worse non-proliferative DR.As was mentioned by investigators, the study has several limitations: hemorrhage is a nonspecific sign, which could accompany other retinal diseases, such as retinal vein occlusions, retinal vasculitis, age-related macular degeneration; signs other than hemorrhage, such as microaneurysms, exudates, and cotton-wool spots could be present in different stages of DR.The algorithm is analyzing single-field 45-degree photographs centered at the fovea.Ischemic proliferative DR is also noneligible for proposed algorithm.The cohort in this study was too small to evidence any benefit.Despite this, however, the study's authors concluded that this approach "could be used to diagnose DR requiring ophthalmologist intervention".
Further, the limitation of this review is the lack of comparability due to different algorithms and grading systems used, commonly separating only referable DR from nonreferable, which hinders the generalization of data.
Foreseen future research should be aimed at the development of portable, fast, tech-friendly ML-based DR screening system capable of detecting all stages of DR, starting from mild, thereby identifying microaneurysms, which possibly would become treatable.
Future innovations are required on versatile algorithms, capable of diagnosing not only DR, but also other eye diseases, such as glaucoma, age-related macular degeneration, cataract, etc., and at the same time to predict cardiovascular (92) and neurodegenerative diseases, (93) in best scenarios at home as a self-testing.

CONCLUSION
The continued assessment and refinement of diagnostic algorithms have the potential to enhance diabetic retinopathy manageability.
A growing body of evidence indicates that machine learning-based diabetic retinopathy screening system is capable of detecting diabetic retinopathy equally or better than professional levels.
Current diabetic retinopathy screening will be cost effectively based on the use of advanced digital technologies, in particular teleretinal screening systems.At present, we may consider teleophthalmology and artificial intelligence with automatic analysis of fundus photos as a Millennium-minded impactful tool for increasing discoverability and manageability of diabetic retinopathy, especially in filling the gap of inaccessibility to hard-toreach areas, which enforces highly professionally effective time-and cost-saving care everywhere to provide the best possible care for the patients.

Table 1 .
Summary of Deep Learning algorithms performance in diabetic retinopathy grading

Table 2 .
Summary of prospective studies worldwide on the efficacy of Artificial Intelligence -based screening systems in the management of diabetic retinopathy