516
OPHTALMOLOGY
Zdrav Vestn | November – December 2022 | Volume 91 | https://doi.org/10.6016/ZdravVestn.3259
Copyright (c) 2022 Slovenian Medical Journal. This work is licensed under a
Creative Commons Attribution-NonCommercial 4.0 International License.
Artificial intelligence for retinal disease management
Umetna inteligenca v obravnavi bolnika z boleznijo mrežnice
Maša Koce,1 Polona Jaki Mekjavič,2,3 Mojca Urbančič,2 Manca Tekavčič Pompe,2,3 Mojca Globočnik Petrovič2,3
Abstract
Image recognition artificial intelligence represents a new milestone in modern medical technology since it has become 
a helpful tool in identifying suspicious changes and diseases, patient monitoring, and predicting treatment outcomes. 
Especially through the implementation of convolutional neural networks in the modelling of computer-based artificial 
intelligence systems, its clinical applicability has recently increased dramatically. Ophthalmology, particularly retinology, 
where diagnosis almost entirely relies on imaging, is highly technology-driven and as such uniquely positioned to bring 
AI innovations into clinical use. However, the integration of AI in the screening, diagnosis and treatment of ophthalmic 
diseases is however still limited, mainly due to the over-generalisation of lesion detection and the poor ability to identify 
different clinical entities simultaneously. This articleis focused on recent artificial intelligence algorithms and software, 
which are primarily aimed to support the detection of the most common retinal diseases and have been, or will be - with 
improved specificity and sensibility, introduced into clinical practice.
Izvleček
Umetna inteligenca predstavlja zaradi možnosti podpore pri odkrivanju in spremljanju bolezni ter napovedovanju izida 
novi mejnik v sodobni medicinski tehnologiji. Njena klinična uporabnost se je v zadnjih letih dramatično povečala pred-
vsem na račun uvedbe konvolucijskih nevronskih mrež v modeliranje računalniško podprtih sistemov umetne inteligence. 
Celotna oftalmologija, zlasti pa področje bolezni mrežnice, pri kateri diagnosticiranje sloni na slikovnih preiskavah, se 
vodi izrazito tehnološko. Zato je edinstvena stroka pri vpeljavi novosti s področja umetne inteligence v klinično uporabo. 
Kljub temu pa je integracija umetne inteligence v presejanje, diagnosticiranje in zdravljenje oftalmoloških bolezni pred-
vsem zaradi prevelikega posploševanja pri zaznavanju sprememb in slabih zmožnosti sočasnega prepoznavanja različnih 
kliničnih entitet zaenkrat še omejena. V članku se osredinjamo na nedavno predstavljene algoritme umetne inteligence, 
ki so prvenstveno namenjeni odkrivanju najpogostejših bolezni mrežnice in so, ali se še bodo, z morebitnimi izboljšavami, 
pridobljeno višjo občutljivostjo in specifičnostjo, vendarle uvedli v klinično prakso.
1 Division of Surgery, University Medical Centre Ljubljana, Ljubljana, Slovenija
2 Department of Ophthalmology, University Clinical Centre Ljubljana, Slovenia
3 Department of Ophthalmology, Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia
Correspondence / Korespondenca: Maša Koce, e: masa.koce@gmail.com
Key words: artificial intelligence; machine learning; age related macular degeneration; diabetic retinopathy; optical coherence 
tomography
Ključne besede: umetna inteligenca; strojno učenje; starostna degeneracija makule; diabetična retinopatija; optična koherentna 
tomografija
Received / Prispelo: 18. 4. 2021 | Accepted / Sprejeto: 15. 9. 2021
Cite as / Citirajte kot: Koce M, Jaki Mekjavič P, Urbančič M, Tekavčič Pompe M, Globočnik Petrovič M. Artificial intelligence for retinal 
disease management. Zdrav Vestn. 2022;91(11–12):516–24. DOI: https://doi.org/10.6016/ZdravVestn.3259
eng slo element
en article-lang
10.6016/ZdravVestn.3259 doi
18.4.2021 date-received
15.9.2021 date-accepted
Ophtalmology Oftalmologija discipline
Professional article Strokovni članek article-type
Artificial intelligence for retinal disease man-
agement
Umetna inteligenca v obravnavi bolnika z bolezni-
jo mrežnice article-title
Artificial intelligence for retinal disease man-
agement
Umetna inteligenca v obravnavi bolnika z bolezni-
jo mrežnice alt-title
artificial intelligence, machine learning, age 
related macular degeneration, diabetic reti-
nopathy, optical coherence tomography
umetna inteligenca, strojno učenje, starostna 
degeneracija makule, diabetična retinopatija, 
optična koherentna tomografija
kwd-group
The authors declare that there are no conflicts 
of interest present.
Avtorji so izjavili, da ne obstajajo nobeni 
konkurenčni interesi. conflict
year volume first month last month first page last page
2022 91 11 12 516 5241
name surname aff email
Maša Koce 1 masa.koce@gmail.com
name surname aff
Polona Jaki Mekjavič 2,3
Mojca Urbančič 2
Manca Tekavčič Pompe 2,3
Mojca Globočnik Petrovič 2,3
eng slo aff-id
Division of Surgery, University 
Medical Centre Ljubljana, 
Ljubljana, Slovenija
Kirurška klinika, Univerzitetni 
klinični center Ljubljana, 
Ljubljana, Slovenija
1
Department of Ophthalmology, 
University Clinical Centre 
Ljubljana, Slovenia
Očesna klinika, Univerzitetni 
klinični center Ljubljana, 
Ljubljana, Slovenija
2
Department of Ophthalmology, 
Faculty of Medicine, University 
of Ljubljana, Ljubljana, Slovenia
Katedra za oftalmologijo, 
Medicinska fakulteta, Univerza v 
Ljubljani, Ljubljana, Slovenija
3
Slovenian Medical Journallovenian Medical Journal
517
PROFESSIONAL ARTICLE
Artificial intelligence for retinal disease management
1 Introduction
Artificial intelligence (AI) means the ability of a com-
puter system to mimic the human mind while acting 
independently of a human. This type of intelligence en-
compasses the ability to connect and process data, learn, 
and infer based on algorithms (1,2).
Machine learning is a branch of artificial intelligence 
that is based on the principles of deep learning or convo-
lutional neural networks (3,4). It is a concept that allows 
the creation of mathematical models based on input da-
ta, and the use of these algorithms to identify and search 
for complex patterns in the large amounts of data (5-7). 
For the system to reliably analyze a specific problem, a 
large enough data set is needed on the basis of which 
the system practices and learns. In the medical profes-
sion, AI is most often used in diagnostic imaging, which, 
due to its prevalence and standardized use, enables the 
analysis and interpretation of imaging data samples 
(8,9). Modern software, based on the principles of AI, 
can already independently identify deviations and irreg-
ularities in the images, compare them with each other, 
and even give a clinical interpretation of the identified 
changes (3,5,7,10).
Ophthalmology is one of the branches of medicine 
Legend: DR – diabetic retinopathy; fundus image – retinal image; OCT – optical coherence tomography; AUC – area under the 
curve; ROP – retinopathy of prematurity; AMD – age-related macular degeneration; AI – artificial intelligence; / – no data.
Study Software Imaging exams Usability AUC
Abramoff (11) IDx-DR fundus image identification of moderate DR 0.98
Chakravarthy (12) Notal OCT 
Analyser
OCT home monitoring for neovascular AMD 0.87
Schmidt-Erfurth 
(13)
/ OCT prediction of the probability of AMD progression 0.68
Russakoff (14) AMDnet OCT prediction of the probability of AMD progression 0.78
Rogers (15) Pegasus OCT, fundus 
image
recognition of pathognomonic changes for glaucoma, 
AMD, DR
0.89–0.94
De Fauw (16) DeepMind 
Health
OCT identification of 50 different retinal pathologies 0.99
Brown (4) I-ROP DL fundus image detection of PLUS disease in preterm infants 0.98
Redd (17) I-ROP DL fundus image identification of individual ROP stages 0.96
Yildiz (18) I-ROP Assist fundus image detection of PLUS disease in preterm infants 0.99
Table 1: Summary and review of AI systems for retinal disease recognition.
in which imaging examinations are part of the basic 
medical treatment of the patient. As we encounter a 
huge amount (several terabytes) of imaging data every 
day in ophthalmology, this field of medicine is unique 
for introducing artificial intelligence systems into clini-
cal practice. Based on retinal images and optical coher-
ence tomography (OCT) imaging of individual layers of 
the retina or optic papilla, deep learning algorithms can 
recognize diabetic retinopathy (DR), age-related mac-
ular degeneration (AMD), retinopathy of prematurity 
(ROP), and glaucoma (Table 1) (4,11-18).
2 Early detection of diabetic retinopathy
Diabetic retinopathy (DR) is the most common 
chronic complication of diabetes and a cause of blindness 
in the working age population (19). About one-third of 
diabetics have DR, and in 10% of patients, the progres-
sive disease poses a serious threat to vision (20,21). The 
disease may not show any symptoms for a long time; 
therefore, timely or early detection of changes is essen-
tial to prevent complications, including permanent vi-
sion loss. The introduction of a screening program for 
518
OPHTALMOLOGY
Zdrav Vestn | November – December 2022 | Volume 91 | https://doi.org/10.6016/ZdravVestn.3259
Figure 1: Moderate non-proliferative diabetic retinopathy (DR), retinal image.
Schematic representation of possible automated recognition of optic papilla and macula (dashed), bleeding and/or aneurysm 
(yellow), soft exudates (blue), as provided by IDx-DR. The image is the source owned by the Department of Ophthalmology, 
University Medical Centre Ljubljana.
early detection of DR has significantly reduced the in-
cidence of blindness among diabetic patients. However, 
the implementation of the program in its current form 
additionally burdens the health care system. The num-
ber of patients with diabetes in the world, as well as in 
our country, is growing rapidly. According to statistics 
from the National Institute of Public Health (NIJZ), the 
incidence increases by 3% every year (22). Due to the 
growing trend of morbidity and the limited availability 
of qualified specialists, the medical staff is overburdened, 
and access to appropriate ophthalmology treatment is 
therefore already difficult in some places (23).
Due to the established standard international classifi-
cation of individual stages of the disease, retinal involve-
ment in DR is highly suitable for the use of AI in image 
analysis. AI systems can easily identify patterns of retinal 
involvement and thus help in screening and diagnosing 
DR (11,16,29). AI algorithms that detect clinically-sig-
nificant macular oedema and advanced stages of DR are 
particularly effective (11,23).
In April 2018, the US Federal Food and Drug Ad-
ministration (FDA) approved an artificial intelligence 
system that is capable of self-diagnosing DR without the 
results having to be interpreted by an ophthalmologist 
(24). Developed by Abramoff et al., IDx-DR software 
(Digital Diagnostics, Coralville, Iowa, USA, former-
ly known as IDx) is based on machine learning and a 
computer algorithm that acts as a screening program 
for DR detection in patients with pre-existing diabetes 
based on standardized retinal images (11). The program 
detects microaneurysms, hard exudates, retinal haem-
orrhages, venous malformations, and soft exudates (mi-
cro infarctions of the nerve fibre layer, the cotton-wool 
spots) on retina images taken with a non-mydriatic cam-
era (Topcon TRC NW 400) (Figure 1). The images at the 
Department of Ophthalmology of the University Medi-
cal Centre contain a schematic representation (Figures 
1, 2 and 3) with program analysis simulation prepared 
using the MacOS Preview software.
If the program detects moderate non-proliferative 
DR (stage of DR > 35 according to the ETDRS [Ear-
ly Treatment for Diabetic Retinopathy Study] severi-
ty scale) or oedema > 300 μm in the wider area of the 
macula on at least one of the four retinal images, fur-
ther treatment of the patient at the ophthalmologist is 
recommended (23,25). The test result is given in binary 
form, 0 or 1, where the value 0 indicates the absence of 
significant DR and the value 1 indicates the presence of 
DR that needs further treatment (11). In a clinical study 
in 900 volunteers, IDx-DR achieved a sensitivity of 87% 
and a specificity of 91% (23).
The program is not suitable for the treatment of pa-
tients with confirmed DR, pre-existing retinal vascular 
pathology such as radiation retinopathy or retinal vein 
519
PROFESSIONAL ARTICLE
Artificial intelligence for retinal disease management
Figure 2: Schematic representation of the analysis of the 
optical coherence tomography (OCT) image of the macula.
Identification of the inner limiting membrane and the layer 
of pigment epithelium (green), the location of the subretinal 
fluid (red), as provided by the proprietary algorithm NOA. 
The image is the source owned by the Department of 
Ophthalmology, University Medical Centre Ljubljana.
occlusion, and patients who have already received intra-
ocular injection or have undergone eye surgery. The pro-
gram is also not intended for the blind, cataract patients, 
children, and pregnant women with rapidly progressive 
DR (11,24).
The equipment is intended exclusively for use at the 
primary health care level, as it only identifies patients 
who need further treatment by an ophthalmologist due 
to advanced DR. However, the system does not provide 
information on the general eye condition of the individ-
ual or the absence of DR (23,25). The main drawback 
of such a system is that among patients with an earlier 
stage of DR, it does not recognize those with a higher 
risk of disease progression. Due to hormonal imbalance, 
diabetes is one of the known risk factors for cataract de-
velopment, age-related macular degeneration, and glau-
coma optic neuropathy (26-28). As the system does not 
recognise other pathologies that need to be examined 
by an ophthalmologist and may need treatment, it may 
give patients a false sense of safety due to a negative test 
result.
The IDx-DR program has been commercially avail-
able in the United States since 2018 (24). Although the 
program was CE certified in 2013, it is not yet available 
for marketing in most European countries, as the inde-
pendent research to demonstrate the effectiveness and 
comparability of software with clinical ophthalmology 
treatment is still ongoing (29).
3 Monitoring age-related macular 
degeneration
Age-related macular degeneration (AMD) is the 
leading cause of functional blindness after the age of 55 
in the developed world (30,25). It is a progressive disease 
that affects the central vision needed for daily tasks such 
as reading, driving, and face recognition. The disease is 
initially asymptomatic; only with its progression does it 
manifest itself with deterioration of vision, metamorph-
opsia, and later with a central scotoma. In approximately 
10–15% of patients with AMD, the disease progresses 
into a neovascular form, characterized by the growth of 
new vessels in the macula (macular neovascularisation; 
MNV) and rapid deterioration of central visual acuity 
(31). Today, the deterioration of vision and the develop-
ment of functional blindness due to neovascular AMD 
can be successfully prevented with biological drugs 
against vascular endothelial growth factor (VEGF), the 
anti-VEGF drugs. This treatment is effective only in the 
initial stages when the newly formed pathological ves-
sels are not yet scarred. Identification of at-risk groups 
of patients in whom disease progression and early detec-
tion of neovascular AMD can be expected is crucial for 
successful treatment.
In addition to regular ophthalmic examinations, 
self-monitoring of vision with the Amsler grid is recom-
mended in the treatment of patients with AMD, through 
which patients themselves detect visual impairment. 
Despite adequate patient participation, the subjective 
perception of changes in visual acuity is often a late in-
dicator of disease progression (32). If the treatment is 
started with poorer visual acuity, the functional effect of 
the treatment or the visual acuity after treatment is also 
poor and vice versa: the better the visual acuity at the be-
ginning of treatment, the better it will be after treatment 
(33,34). Early detection of the disease can significantly 
improve the prognosis for vision.
For early detection of the progression of dry AMD 
into neovascular form, Notal Vision (Notal Vision, 
Manassas, Virginia, USA) has developed an OCT for 
home use. The Home OCT combines the research tech-
nology of optical coherence tomography and the princi-
ples of machine learning and neural networks (12). The 
device records a macular tomography, which is auto-
matically analyzed by NOA™ (Notal OCT Analyzer). If 
in the central 10 degrees of the retina, in the macula, it 
detects fluid in or under the retina, it immediately noti-
fies the chosen ophthalmologist (Figure 2). If it does not 
detect fluid, it transmits monthly reports to the cloud, 
which can be accessed by the chosen physician (12). The 
device thus enables continuous monitoring of macular 
520
OPHTALMOLOGY
Zdrav Vestn | November – December 2022 | Volume 91 | https://doi.org/10.6016/ZdravVestn.3259
Figure 3: Neovascular age-related macular degeneration (AMD), optical coherence tomography (OCT) image (right), and 
possible schematic representation of automated image analysis with DeepMind software (left).
Anatomical segmentation of the retina (green), the layer of pigment epithelium (orange) and the choroid (grey). Identification 
of intraretinal fluid (blue), hyperreflective foci (purple), and neovascularization (red). The image is the source owned by the 
Department of Ophthalmology, University Medical Centre Ljubljana.
disease in high-risk patients, contributing to the rapid 
identification of disease progression, thus shortening the 
time from disease exacerbation to the start of treatment 
(12). It is known that the results of treating AMD with 
anti-VEGF drugs in clinical practice are not as good as 
in clinical trials (12,35). The reason is the high burden 
of treatment, which requires regular monitoring and 
treatment at appropriate intervals. Therefore, monitor-
ing of patients with macular changes that pose a high 
risk of progression into neovascular disease is not reg-
ular. The Home OCT device detects pathological events 
in the macula even before the changes significantly af-
fect vision, when the initial visual acuity is still good. If 
treatment is timely, intervention is more effective and 
treatment outcomes are better. The Home OCT device 
is intended for self-testing at home between regularly 
scheduled clinical examinations and is not a substitute 
for regular examinations in the retinal disease clinic, as it 
is insufficiently accurate compared to professional OCT 
devices routinely used in specialist clinics (12). The ar-
rival of Home OCT on the market has already been an-
nounced for this year. The device would be mainly avail-
able to patients at high risk for disease progression (12).
Recently, an AI model has been developed at the Uni-
versity of Vienna that can evaluate the likelihood of pro-
gression into neovascular AMD based on demographics, 
genetic polymorphisms, and image data analysis (13). 
Using automated segmentation in individual retinal 
layers, the system detects drusen, pseudodrusen, hyper-
reflective foci, and other abnormalities in the pigment 
epithelium that indirectly indicate a high risk of devel-
oping neovascular AMD (13). A similar pre-processing 
algorithm for the analysis of OCT images was also used 
by Russakoff et al. as the basis of their own AI model 
(AMDnet). The basic analysis of the retinal layers is 
followed by automated recognition of pathognomonic 
retinal abnormalities according to structural parameters 
such as thickness and volume (14).
4 Software for simultaneous detection of 
multiple retinal diseases
The British company Visulytix (London, United 
Kingdom) has developed a series of advanced algorithms 
that simultaneously identify various retinal pathologies. 
Their Pegasus software interprets and analyses OCT im-
ages in a detailed way: retinal images and stereoscopic 
images of the optic papilla. It provides immediate sup-
port in deciding on the further treatment of patients 
with DR, AMD and glaucoma (15,36,37). In OCT imag-
es, Pegasus detects drusen, atrophic areas of the retina, 
and intraretinal fluid (IRF) or accumulation of fluid un-
der the retina and thus assesses the presence of macular 
oedema, and dry, atrophic, and neovascular AMD (36). 
It detects microaneurysms, retinal haemorrhages, and 
exudates in retinal images, also recognizes optic papilla 
anomalies, and accurately measures the vertical cup to 
disk ratio (CDR) (15,37,38). The company completed its 
sales in 2019.
Pegasus was designed to unburden ophthalmologists 
in the tertiary health care, who interpret vast amounts 
of imaging data taken daily in outpatient clinics where 
basic screening tests for eye diseases are performed. The 
program automatically analyses image data, which is 
521
PROFESSIONAL ARTICLE
Artificial intelligence for retinal disease management
either accessed directly from a device connected to the 
OCT or with a fundus camera, or captures information 
from the cloud, where data can also be uploaded by ex-
ternal imaging providers (15,37). In specialist clinics, the 
program can be used to assist in triage and treatment 
of patients, and outside clinical centres, Pegasus allows 
for earlier detection of the disease and thus ensures that 
emergencies are treated in a timely manner (15,37). In 
clinical studies, the Pegasus software has achieved the 
accuracy of a specialist in the detection of glaucoma 
neuropathy and retinal pathology (40).
In 2016, Google’s DeepMind (London, UK), in col-
laboration with the Moorfields Eye Hospital in London, 
developed an AI system that successfully detects more 
than 50 pathological changes on OCT images (16). The 
system is one of the most advanced in this field, com-
bining the functions of two neural networks and allow-
ing two-stage analysis of OCT images. In the first stage 
it detects changes in the retina (drusen, geographical 
atrophy, epiretinal membrane, MNV, macular oedema, 
macular foramen, vitreomacular traction, and central 
serous retinopathy), which it then interprets in the sec-
ond stage and concludes the diagnosis (Figure 3) (16). 
Depending on the severity of retinal pathology, it also 
proposes a recommendation for further treatment of 
the patient, so this software is primarily intended for 
technological support in specialized ophthalmological 
clinics.
5 The use of artificial intelligence to guide 
and determine the effect of treatment
Today, treatment with anti-VEGF drugs (aflibercept, 
bevacizumab, brolucizumab, ranibizumab) is consid-
ered the gold standard in the treatment of patients with 
neovascular AMD, proliferative DR, and diabetic mac-
ular oedema (DME), as by binding to VEGF receptors 
these drugs prevent endothelial cell proliferation and 
neovascularization, reduce increased vascular permea-
bility and thus prevent the accumulation of intraretinal 
fluid (IRF) and subretinal fluid (SRF). Today, measure-
ments of central retinal thickness (CRT) on the OCT 
image help us to quantify the fluid in the macula.
Rasti et al. have developed the CADNet (Convolu-
tional Attention-to-DME Network) software to help 
predict the response to anti-VEGF treatment in DME 
patients (40). Assuming the effectiveness of the drug 
used and based on the measured CRT on OCT images 
obtained before the start of intravitreal treatment, the 
algorithm automatically estimates the predicted reduc-
tion in retinal thickness, or the effect of treatment after 
three months (40). Compared to similar algorithms 
(Extra Trees, Inception V3, ResNet50, Visual Geometry 
Group 16, Xception), which require the input of longi-
tudinal OCT images to assess the effect of treatment, the 
CADNet algorithm has proven to be a more accurate 
tool for predicting disease outcome (39,40).
In recent years, there is growing evidence that retinal 
thickness is by no means an ideal method for captur-
ing morphological changes in the retina. CRT should 
certainly not be the only parameter for deciding on 
further treatment of the patient (42-45). In addition to 
retinal fluid, the effect of anti-VEGF treatment is also 
influenced by structural changes in the retina, such as 
pigment epithelial atrophy, changes in the photorecep-
tor layer and vitreoretinal interface, the IRF pattern 
distribution, cystic degeneration, hyperreflective foci, 
disorganization of retinal inner layers (DRIL), and vas-
cular thickness (46). The presence of IRF is the most 
important risk factor for visual acuity deterioration in 
patients with neovascular AMD due to retinotoxicity 
(47,48). Since IRF volume and area are directly related 
to visual function, the predicted treatment outcome can 
be determined by measuring the amount and distribu-
tion of IRF (48). Based on these findings, the University 
of Vienna has developed a complex algorithm that auto-
matically detects the presence of IRF, SRF, and pigment 
epithelial detachment on a three-dimensional OCT 
image (49). The algorithm also reliably detects and spa-
tially evaluates diffusely distributed IRF and IRF with 
multilocular cysts, which is otherwise clinically difficult 
to assess (49). In the study, the algorithm proved to be a 
promising tool for deciding on the intravitreal therapy 
regimen of patients with neovascular AMD (49).
6 Automated screening for retinopathy of 
prematurity
Over the last decade, survival of highly immature 
preterm infants has increased significantly due to ad-
vances in perinatal care which means that in neona-
tology we are facing the problem of higher morbidity 
of surviving preterm infants. Due to retinal immaturi-
ty, preterm infants with very low gestational age are at 
particular risk of developing retinopathy of prematurity 
(ROP). ROP is a relatively rare vasoproliferative retinal 
disease in our geographical area, but it is nevertheless 
one of the most common causes of irreversible vision 
loss in premature infants (50). In Ljubljana, on average, 
20% of preterm infants meeting the criteria for ROP 
screening developed the disease in the last 5 years, and 
15–35% of preterm infants with ROP required treatment 
522
OPHTALMOLOGY
Zdrav Vestn | November – December 2022 | Volume 91 | https://doi.org/10.6016/ZdravVestn.3259
(51). With a timely diagnosis, most of the sight threat-
ening ROPs can be successfully treated, so in Slovenia, 
all preterm infants born before the 31st week of gesta-
tion and/or with a birth weight of less than 1500 g are 
included in the ROP screening program (51).
Screening of preterm infants for ROP involves eye 
examination with an indirect ophthalmoscope at one- 
to two-week intervals, mostly between the ages of 30 
and 37 weeks of gestational age. The number of exam-
inations, the age at the start of the screening, and the 
completion of screening depend on the stage of ROP 
and the associated systemic conditions of the preterm 
infant. The degree of ROP is determined according to 
the size of the avascular area, the appearance of the 
border (ridge) between the vascularized and avascu-
lar region of the retina, and according to the dilatation 
and tortuosity of the vessels of the already vascularized 
retina (52). To more closely monitor the state of retinal 
vascular development, most institutions use wide-angle 
cameras to document the appearance of the retina at 
each screening examination. Comparing a large num-
ber of images showing a normal appearance of the ret-
ina at different gestational ages of preterm infants with 
those showing different forms of ROP is the foundation 
of all software programs that incorporate AI elements. 
Most existing computer aided ROP detection systems, 
e.g. Vessel finder (53), RISA (54), RIVERS (55), CAIAR 
(56), are semi-automated and do not yet achieve the 
ability of a conventional ophthalmic examination.
The analytical procedures of more advanced auto-
mated algorithms for determining the degree of ROP 
are mostly based on the evaluation of retinal vessel dil-
atation and tortuosity, the PLUS disease. Such AI sys-
tems provide greater objectivity and repeatability, mak-
ing the accuracy of PLUS disease detection comparable 
to the conventional clinical examination. Brown et al. 
devised a fully automated screening program system 
for ROP based on deep learning and neural networks. 
It recognizes the PLUS disease in retina images with a 
specificity of 94% and a sensitivity of 93% (4). The au-
tomated computer system has even surpassed the expe-
rienced paediatric ophthalmologists involved in clinical 
research in identifying the disease (4). By upgrading the 
same AI system, Redd et al. (17) enabled differentiation 
and quantification of stages of ROP while maintaining 
the efficiency and accuracy of the program.
E. Ataer-Cansizoglu and V. M. Yildiz developed a 
similar model of artificial intelligence. It automatical-
ly detects clinically significant changes in the retinal 
vessels in retinal images of preterm infants at risk of 
developing ROP, thus providing a specific recognition 
of individual stages of ROP (18,57).
Software programs supported by AI, which are al-
ready installed in some wide-angle cameras for docu-
menting the retina of preterm infants, can already be of 
great help today to a paediatric ophthalmologist during 
the examination. For the time being, they distinguish 
only between present and absent PLUS disease, and 
perform less well at assessing the degree of (pre-)PLUS 
disease and at other parameters for assessing ROP, such 
as extension of vasculature, presence and height of the 
ridge between avascular and vascular retina, in recog-
nizing individual vessels passing over the ridge (and 
representing a good predictor of continued vasculari-
sation), etc. The development of image processing tech-
nology, growing databases, more accurate images, and 
added information provided by Doppler ultrasonogra-
phy of the retinal vessels will undoubtedly significantly 
facilitate screening work in many institutions around 
the world in the future.
7 Conclusion
With the advancement of information and cognitive 
science, AI in the medical profession has experienced 
a rise. Software based on the principles of deep learn-
ing and convolutional neural networks has proven to 
be an effective tool in ophthalmology due to its high 
performance in image data processing and analysis, 
especially in the treatment of patients with retinal dis-
eases. The most common diseases of the retina, diabet-
ic retinopathy and age-related macular degeneration, 
are the leading causes of permanent vision loss, and at 
the same time their early detection allows for effective 
treatment and thus significantly increases the chances 
of maintaining functional vision. The prevalence of DR 
and AMD in the developed world is rising sharply, and 
the economic burden of these diseases is increasing at 
the expense of permanent vision loss. With early de-
tection and treatment of retinal diseases, blindness can 
be largely prevented and functional vision in patients 
maintained for a longer period of time. The first AI al-
gorithms in ophthalmology were therefore developed 
with the aim of early recognition of preclinical signs. 
Due to the accuracy in recognizing retinal pathology, 
they can also be used to support decision-making on 
further treatment of ophthalmic patients.
Conflict of interest
None declared.
523
PROFESSIONAL ARTICLE
Artificial intelligence for retinal disease management
References
1. Grewal PS, Oloumi F, Rubin U, Tennant MT. Deep learning in 
ophthalmology: a review. Can J Ophthalmol. 2018;53(4):309-13. DOI: 
10.1016/j.jcjo.2018.04.019 PMID: 30119782
2. Ahuja AS, Halperin LS. Understanding the advent of artificial intelligence 
in ophthalmology. J Curr Ophthalmol. 2019;31(2):115-7. DOI: 10.1016/j.
joco.2019.05.001 PMID: 31317087
3. Lu W, Tong Y, Yu Y, Xing Y, Chen C, Shen Y. Applications of artificial 
intelligence in ophthalmology: general overview. J Ophthalmol. 
2018;2018:5278196. DOI: 10.1155/2018/5278196 PMID: 30581604
4. Brown JM, Campbell JP, Beers A, Chang K, Ostmo S, Chan RV, et 
al.; Imaging and Informatics in Retinopathy of Prematurity (i-ROP) 
Research Consortium. Imaging and Informatics in Retinopathy of 
Prematurity (i-ROP) Research Consortium.Automated diagnosis of plus 
disease in retinopathy of prematurity using deep convolutionalneural 
networks. JAMA Ophthalmol. 2018;136(7):803-10. DOI: 10.1001/
jamaophthalmol.2018.1934 PMID: 29801159
5. Kapoor R, Whigham BT, Al-Aswad LA. Artificial intelligence and optical 
coherence tomography imaging. Asia Pac J Ophthalmol (Phila). 
2019;8(2):187-94. PMID: 30997756
6. Kapoor R, Walters SP, Al-Aswad LA. The current state of artificial 
intelligence in ophthalmology. Surv Ophthalmol. 2019;64(2):233-40. DOI: 
10.1016/j.survophthal.2018.09.002 PMID: 30248307
7. Keskinbora K, Güven F. Artificial intelligence and ophthalmology. Turk 
J Ophthalmol. 2020;50(1):37-43. DOI: 10.4274/tjo.galenos.2020.78989 
PMID: 32167262
8. Xu J, Xue K, Zhang K. Current status and future trends of clinical diagnoses 
via image-based deep learning. Theranostics. 2019;9(25):7556-65. DOI: 
10.7150/thno.38065 PMID: 31695786
9. Yu KH, Beam AL, Kohane IS. Artificial intelligence in healthcare. Nat 
Biomed Eng. 2018;2(10):719-31. DOI: 10.1038/s41551-018-0305-z PMID: 
31015651
10. Ting DS, Pasquale LR, Peng L, Campbell JP, Lee AY, Raman R, et al. Artificial 
intelligence and deep learning in ophthalmology. Br J Ophthalmol. 
2019;103(2):167-75. DOI: 10.1136/bjophthalmol-2018-313173 PMID: 
30361278
11. Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal trial 
of an autonomous AI-based diagnostic system for detection of 
diabeticretinopathy in primary care offices. NPJ Digit Med. 2018;1:39. 
DOI: 10.1038/s41746-018-0040-6 PMID: 31304320
12. Chakravarthy U, Goldenberg D, Young G, Havilio M, Rafaeli O, Benyamini 
G, et al. Automated Identification of Lesion Activity in Neovascular Age-
Related Macular Degeneration. Ophthalmology. 2016;123(8):1731-6. DOI: 
10.1016/j.ophtha.2016.04.005 PMID: 27206840
13. Schmidt-Erfurth U, Waldstein SM, Klimscha S, Sadeghipour A, Hu 
X, Gerendas BS, et al. Prediction of individual disease conversion 
in early AMD using artificial intelligence. Invest Ophthalmol Vis Sci. 
2018;59(8):3199-208. DOI: 10.1167/iovs.18-24106 PMID: 29971444
14. Russakoff DB, Lamin A, Oakley JD, Dubis AM, Sivaprasad S. Deep learning 
for prediction of AMD progression: A pilot study. Invest Ophthalmol Vis 
Sci. 2019;60(2):712-22. DOI: 10.1167/iovs.18-25325 PMID: 30786275
15. Rogers TW, Jaccard N, Carbonaro F, Lemij HG, Vermeer KA, Reus NJ, et 
al. Evaluation of an AI system for the automated detection of glaucoma 
from stereoscopicoptic disc photographs: the European Optic Disc 
Assessment Study. Eye (Lond). 2019;33(11):1791-7. DOI: 10.1038/s41433-
019-0510-3 PMID: 31267086
16. De Fauw J, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N, 
Blackwell S, et al. Clinically applicable deep learning for diagnosis and 
referral in retinal disease. Nat Med. 2018;24(9):1342-50. DOI: 10.1038/
s41591-018-0107-6 PMID: 30104768
17. Redd TK, Campbell JP, Brown JM, Kim SJ, Ostmo S, Chan RV, et al.; 
Imaging and Informatics in Retinopathy of Prematurity (i-ROP) Research 
Consortium. Evaluation of a deep learning image assessment system 
for detecting severe retinopathyof prematurity. Br J Ophthalmol. 
2018;bjophthalmol:2018-313156. PMID: 30470715
18. Yildiz VM, Tian P, Yildiz I, Brown JM, Kalpathy-Cramer J. Plus Disease 
in Retinopathy of Prematurity: Convolutional Neural Network 
PerformanceUsing a Combined Neural Network and Feature Extraction 
Approach. Transl Vis Sci Technoll. 2020;9(2):10. DOI: 10.1167/tvst.9.2.10 
PMID: 32704416
19. Bourne RR, Stevens GA, White RA, Smith JL, Flaxman SR, Price H, et al.; 
Vision Loss Expert Group. Causes of vision loss worldwide, 1990-2010: a 
systematic analysis. Lancet Glob Health. 2013;1(6):e339-49. DOI: 10.1016/
S2214-109X(13)70113-X PMID: 25104599
20. Yau JW, Rogers SL, Kawasaki R, Lamoureux EL, Kowalski JW, Bek T, et 
al.; Meta-Analysis for Eye Disease (META-EYE) Study Group. Global 
prevalence and major risk factors of diabetic retinopathy. Diabetes Care. 
2012;35(3):556-64. DOI: 10.2337/dc11-1909 PMID: 22301125
21. Lee R, Wong TY, Sabanayagam C. Epidemiology of diabetic retinopathy, 
diabetic macular edema and related vision loss. Eye Vis (Lond). 
2015;2(1):17. DOI: 10.1186/s40662-015-0026-2 PMID: 26605370
22. Nacionalni inštitut za javno zdravje. Sladkorna bolezen v Sloveniji: Kje 
smo in kam gremo? Ljubljana: NIJZ; 2018 [cited 2020 Apr 25]. Available 
from: https://www.nijz.si/sl/sladkorna-bolezen-v-sloveniji-kje-smo-in-
kam-gremo.
23. van der Heijden AA, Abramoff MD, Verbraak F, van Hecke MV, Liem A, 
Nijpels G. Validation of automated screening for referable diabetic 
retinopathy with the IDx-DRdevice in the Hoorn Diabetes Care System. 
Acta Ophthalmol. 2018;96(1):63-8. DOI: 10.1111/aos.13613 PMID: 
29178249
24. US Food and Drug Administration. FDA permits marketing of artificial 
intelligence-based device to detect certain diabetes-relatedeye 
problems. Maryland: Food and drug administration; 2018 [cited 2020 
Apr 25]. Available from: https://www.fda.gov/news-events/press-
announcements/fda-permits-marketing-artificial-intelligence-based-
device-detect-certain-diabetes-related-eye.
25. Li JQ, Welchowski T, Schmid M, Mauschitz MM, Holz FG, Finger RP. 
Prevalence and incidence of age-related macular degeneration in 
Europe: a systematicreview and meta-analysis. Br J Ophthalmol. 
2020;104(8):1077-84. DOI: 10.1136/bjophthalmol-2019-314422 PMID: 
31712255
26. Chen X, Rong SS, Xu Q, Tang FY, Liu Y, Gu H, et al. Diabetes mellitus 
and risk of age-related macular degeneration: a systematic reviewand 
meta-analysis. PLoS One. 2014;9(9):e108196. DOI: 10.1371/journal.
pone.0108196 PMID: 25238063
27. Khan A, Petropoulos IN, Ponirakis G, Malik RA. Visual complications in 
diabetes mellitus: beyond retinopathy. Diabet Med. 2017;34(4):478-84. 
DOI: 10.1111/dme.13296 PMID: 27917530
28. Mitchell P, Smith W, Chey T, Healey PR. Open-angle glaucoma and 
diabetes: the Blue Mountains eye study, Australia. Ophthalmology. 
1997;104(4):712-8. DOI: 10.1016/S0161-6420(97)30247-4 PMID: 9111268
29. Nakhuda H. IDx-DR: Automated Screening for Diabetic Retinopathy/
Detection of Cognitive Impairmentand Dementia With Artificial 
Intelligence. Health Technology Update. 2018(22):4.
30. Jonas JB, Cheung CM, Panda-Jonas S. Updates on the epidemiology 
of age-related macular degeneration. Asia Pac J Ophthalmol (Phila). 
2017;6(6):493-7. PMID: 28906084
31. Bowling B, Kanski JJ. Kanski’s clinical ophthalmology: A systematic 
approach. 68th ed. Edinburgh: Elsevier; 2016.
524
OPHTALMOLOGY
Zdrav Vestn | November – December 2022 | Volume 91 | https://doi.org/10.6016/ZdravVestn.3259
32. Ying GS, Huang J, Maguire MG, Jaffe GJ, Grunwald JE, Toth C, et al.; 
Comparison of Age-related Macular Degeneration Treatments Trials 
Research Group. Baseline predictors for one-year visual outcomes 
with ranibizumab or bevacizumab forneovascular age-related macular 
degeneration. Ophthalmology. 2013;120(1):122-9. DOI: 10.1016/j.
ophtha.2012.07.042 PMID: 23047002
33. Arias L, Armadá F, Donate J, García-Arumí J, Giralt J, Pazos B, et al. Delay 
in treating age-related macular degeneration in Spain is associated 
with progressivevision loss. Eye (Lond). 2009;23(2):326-33. DOI: 10.1038/
sj.eye.6703053 PMID: 18202712
34. Muether PS, Hermann MM, Koch K, Fauser S. Delay between 
medical indication to anti-VEGF treatment in age-related macular 
degenerationcan result in a loss of visual acuity. Graefes Arch Clin Exp 
Ophthalmol. 2011;249(5):633-7. DOI: 10.1007/s00417-010-1520-9 PMID: 
20865421
35. Joachim N, Colijn JM, Kifley A, Lee KE, Buitendijk GH, Klein BE, et al. 
Five-year progression of unilateral age-related macular degeneration 
to bilateralinvolvement: the Three Continent AMD Consortium 
report. Br J Ophthalmol. 2017;101(9):1185-92. DOI: 10.1136/
bjophthalmol-2016-309729 PMID: 28108569
36. Notal Vision. Notal Vision Announces FDA Grants Breakthrough Device 
Designation for Pioneering Patient-Operated Home Optical Coherence 
Tomography (OCT) System. Tel Aviv: Notal Vision; 2018 [cited 2020 Apr 
25]. Available from: https://nv.steadfa.st//assets/press-releases/Notal-
Vision-Announces-FDA-Grants-Breakthrough-Device-Designation-for-
Pioneering-Patient-Operated-Home-Optical-Coherence-Tomography-
OCT-System.pdf.
37. Cohen SY, Mimoun G, Oubraham H, Zourdani A, Malbrel C, Queré S, et 
al.; LUMIERE Study Group. Changes in visual acuity in patients with wet 
age-related macular degeneration treatedwith intravitreal ranibizumab 
in daily clinical practice: the LUMIERE study. Retina. 2013;33(3):474-81. 
DOI: 10.1097/IAE.0b013e31827b6324 PMID: 23266880
38. Bhatia KK, Graham MS, Terry L, Wood A, Tranos P, Trikha S, et al. Disease 
classification of macular optical coherence tomography scans using 
deep learningsoftware: validation on independent, multicenter data. 
Retina. 2020;40(8):1549-57. DOI: 10.1097/IAE.0000000000002640 PMID: 
31584557
39. Rogers TW, Gonzalez-Bueno J, Garcia Franco R, Lopez Star E, Méndez 
Marín D, Vassallo J, et al. Evaluation of an AI system for the detection 
of diabetic retinopathy from images capturedwith a handheld portable 
fundus camera: the MAILOR AI study. Eye (Lond). 2021;35(2):632-8. DOI: 
10.1038/s41433-020-0927-8 PMID: 32382145
40. Al-Aswad LA, Kapoor R, Chu CK, Walters S, Gong D, Garg A, et al. 
Evaluation of a deep learning system for identifying glaucomatous 
optic neuropathybased on color fundus photographs. J Glaucoma. 
2019;28(12):1029-34. DOI: 10.1097/IJG.0000000000001319 PMID: 
31233461
41. Rasti R, Allingham MJ, Mettu PS, Kavusi S, Govind K, Cousins SW, et al. 
Deep learning-based single-shot prediction of differential effects of anti-
VEGF treatmentin patients with diabetic macular edema. Biomed Opt 
Express. 2020;11(2):1139-52. DOI: 10.1364/BOE.379150 PMID: 32133239
42. Liu B, Zhang B, Hu Y, Cao D, Yang D, Wu Q, et al. Automatic prediction 
of treatment outcomes in patients with diabetic macular edemausing 
ensemble machine learning. Ann Transl Med. 2021;9(1):43. DOI: 10.21037/
atm-20-1431 PMID: 33553336
43. Keane PA, Liakopoulos S, Chang KT, Wang M, Dustin L, Walsh AC, 
et al. Relationship between optical coherence tomography retinal 
parameters and visual acuityin neovascular age-related macular 
degeneration. Ophthalmology. 2008;115(12):2206-14. DOI: 10.1016/j.
ophtha.2008.08.016 PMID: 18930551
44. Moutray T, Alarbi M, Mahon G, Stevenson M, Chakravarthy U. 
Relationships between clinical measures of visual function, fluorescein 
angiographicand optical coherence tomography features in patients 
with subfoveal choroidal neovascularisation. Br J Ophthalmol. 
2008;92(3):361-4. DOI: 10.1136/bjo.2007.123976 PMID: 18303157
45. Sadda SR, Wu Z, Walsh AC, Richine L, Dougall J, Cortez R, et al. Errors 
in retinal thickness measurements obtained by optical coherence 
tomography. Ophthalmology. 2006;113(2):285-93. DOI: 10.1016/j.
ophtha.2005.10.005 PMID: 16406542
46. Simader C, Ritter M, Bolz M, Deák GG, Mayr-Sponer U, Golbaz I, et al. 
Morphologic parameters relevant for visual outcome during anti-
angiogenic therapyof neovascular age-related macular degeneration. 
Ophthalmology. 2014;121(6):1237-45. DOI: 10.1016/j.ophtha.2013.12.029 
PMID: 24684838
47. Schmidt-Erfurth U, Waldstein SM. A paradigm shift in imaging biomarkers 
in neovascular age-related macular degeneration. Prog Retin Eye Res. 
2016;50:1-24. DOI: 10.1016/j.preteyeres.2015.07.007 PMID: 26307399
48. Schmidt-Erfurth U, Vogl WD, Jampol LM, Bogunović H. Application 
of automated quantification of fluid volumes to anti-VEGF therapy 
of neovascularage-related macular degeneration. Ophthalmology. 
2020;127(9):1211-9. DOI: 10.1016/j.ophtha.2020.03.010 PMID: 32327254
49. Kodjikian L, Decullier E, Souied EH, Roux A, Aulagner G, Huot L; GEFAL 
Study Group. Predictors of one-year visual outcomes after anti-
vascular endothelial growth factortreatment for neovascular age-
related macular degeneration. Retina. 2018;38(8):1492-9. DOI: 10.1097/
IAE.0000000000001736 PMID: 28654629
50. Waldstein SM, Philip AM, Leitner R, Simader C, Langs G, Gerendas 
BS, et al. Correlation of 3-dimensionally quantified intraretinal and 
subretinal fluid with visualacuity in neovascular age-related macular 
degeneration. JAMA Ophthalmol. 2016;134(2):182-90. DOI: 10.1001/
jamaophthalmol.2015.4948 PMID: 26661463
51. Bashinsky AL. Retinopathy of Prematurity. N C Med J. 2017;78(2):124-8. 
DOI: 10.18043/ncm.78.2.124 PMID: 28420777
52. Tekavčič Pompe M, Markelj Š, Kornhauser Cerar L.. Retinopathy of 
prematurity in the period 2015-2019 in a tertiary referral centre inSlovenia. 
Zdrav Vestn. 2021;90(9-10):488-95. DOI: 10.6016/ZdravVestn.3085
53. International Committee for the Classification of Retinopathy 
of PrematurityThe international classification of retinopathy of 
prematurity revisited. Arch Ophthalmol. 2005;123(7):991-9. DOI: 10.1001/
archopht.123.7.991 PMID: 16009843
54. Heneghan C, Flynn J, O’Keefe M, Cahill M. Characterization of changes 
in blood vessel width and tortuosity in retinopathy ofprematurity using 
image analysis. Med Image Anal. 2002;6(4):407-29. DOI: 10.1016/S1361-
8415(02)00058-0 PMID: 12426111
55. Gelman R, Martinez-Perez ME, Vanderveen DK, Moskowitz A, Fulton AB. 
Diagnosis of plus disease in retinopathy of prematurity using Retinal 
Image multiScaleAnalysis. Invest Ophthalmol Vis Sci. 2005;46(12):4734-8. 
DOI: 10.1167/iovs.05-0646 PMID: 16303973
56. Tsai CL, Madore B, Leotta MJ, Sofka M, Yang G, Majerovics A, et al. 
Automated retinal image analysis over the internet. IEEE Trans Inf 
Technol Biomed. 2008;12(4):480-7. DOI: 10.1109/TITB.2007.908790 PMID: 
18632328
57. Ataer-Cansizoglu E, Bolon-Canedo V, Campbell JP, Bozkurt A, Erdogmus 
D, Kalpathy-Cramer J, et al. i-ROP Research Consortium. Computer-based 
image analysis for plus disease diagnosisin retinopathy of prematurity: 
performance of the “i-ROP” system and image featuresassociated with 
expert diagnosis. Transl Vis Sci Technol. 2015;4(6):5. DOI: 10.1167/
tvst.4.6.5