Abstract
Artificial Intelligence
(AI) has become a transformative force in changing the healthcare system in
many ways. Machine learning (ML) and deep learning (DL) are used to analyze the
medical images like CT scans, MRIs, ultrasounds, ad X-rays. Using DL algorithms
systems can analyze and detect anomalies such as tumors, lesions and other
abnormalities at a high accuracy and assists the medical professionals in a
better way. AI can assist in automating the diagnosis process by analyzing the
medical images and identify potential issues for review. This paper examines
the role of AI in diagnosing retinal disorders and eye diseases. AI
technologies, particularly DL, have demonstrated remarkable accuracy and
efficiency in detecting conditions like diabetic retinopathy (DR), age-related
macular degeneration (AMD), retinopathy of prematurity (ROP), retinal vein
occlusion (RVO), hypertensive retinopathy (HR). This paper explores current AI
applications, methodologies, challenges and prospects, offering insights into
how AI can revolutionize ophthalmic care.
Keywords: Artificial Intelligence, Machine Learning,
Deep Learning, Retinal disorders, eye disease, retinal imaging, retinopathy
1. Introduction
Artificial
intelligence (AI) has emerged as a game-changer in healthcare, with the
potential to overcome many of the challenges. AI technologies become
increasingly embedded in healthcare, the adoption of AI is not just about the
technological integration but also ensuring these systems are deployed and responsibly1,2. There is several AI technologies deployed in
the healthcare, neural network-based techniques2 such as DL and ML, enhance diagnostic
accuracy, improve efficiency. These AI techniques analyze retinal images,
detects anomalies and assists clinicians in making informed decisions3.
Conditions like
diabetic retinopathy (DR), age-related macular degeneration (AMD), retinopathy
of prematurity (ROP), retinal vein occlusion (RVO), hypertensive retinopathy
(HR) can often go undiagnosed in their early stages due to limited access to
specialized care and resource-intensive diagnostic methods. Early detection is
crucial, as timely intervention can prevent irreversible vision loss.
This paper
explores the current state of AI in diagnosing retinal disorders, highlighting
its applications, manual and AI driven workflows, technologies and benefits. It
also examines the challenges associated with implementing AI in clinical
practice and discusses future opportunities for this transformative technology.
2. The Need for AI in Retinal Diagnostics
The need for AI
in retinal diagnostics is its potential to enhance the accuracy, turnaround
time for diagnosis and better access to eye care.
2.1. Current challenges in ophthalmology
·Global Burden of Retinal Diseases: Retinal disorders affect millions worldwide,
with over 1 billion have moderate to severe vision impairment from preventable
eye disease4. According to Lama et al, existing evidence
suggests that vision impairment is associated with lower quality of life, like
physical emotional and social well-being.
·Limited Access to Specialized Care: Rural and low-income regions often lack
ophthalmologists and diagnostic tools, leading to delayed detection and
irreversible damage. The primary reasons are availability, affordability and
accessibility of eye care services globally especially in developing countries5. A study conducted in Ethiopia by Bekele et
al, shows that the utilization of the eye care services depended on the factors
like, education, awareness, income levels, availability of insurance,
affordability. It was identified that old age people, people who can afford,
people with insurance, people who already diagnosed with eye problems tend to
utilize the eye care services. According to the study conducted by Gilbert et
all in India, the reason for the eye care visit was primarily due to the impact
of diabetic patients visit the eye care facility. It is evident that the study
shows that the awareness and availability of trained resources and
affordability of care are the reasons for less utilization of the care in India6. The rural people must travel miles to reach
eye care facilities and they are visiting as need based not as regular
screening.
·Resource-Intensive Diagnostics: Traditional methods require highly skilled
professionals and expensive imaging equipment, limiting scalability7. The examination of fundoscopic images by
ophthalmologist requires expensive equipment, trained personal for image
acquisition and interpretation. Similar changes exist with different imaging
techniques. Some of the challenges with these traditional screening types can
be accessibility, patient compliance, cost, not being able to detect in an
early stage.
2.2. How AI fills the gap
·Early detection: DL models have demonstrated the ability to
detect early signs of retinal disease even before the disease is noticed by the
patient or by an ophthalmologist. The demand for retinal diagnostic imaging
exams has increased, but the but the number of eye physicians are too little to
cover the need in the world. Early detection of eye disease with technology
support providing doctors with new insights than the traditional approach8,9. Early detection helps prevention of eye
diseases in a better way. Another advancement experiment conducted by Shu et al
by enabling AI scanning the photographs taken by the mobile to detect eye
disease in pediatric children10, the DL model for early detection of myopia, strabismus and
ptosis, the model demonstrated high accuracy in detecting the pediatric eye
disease.
·Automation of diagnostics: AI can automate the analysis of retinal
images, such as optical coherence tomography (OCT), fundus images, scans which
are manually time consuming, automating these with AI can help improve the
screening process and can help reduce the burden on ophthalmologists11. AI algorithms analyze retinal images with
high precision, reducing dependency on specialists.
·Improved accuracy: AI system can provide high level of accuracy
in identifying retinal pathologies, research has shown that DL models can
exceed the human expertise in diagnosing the retinal diseases. Consistent
improvement can help reduce the misdiagnosis and enhanced patient results. DL
algorithms were trained with larger volumes of images with expected outcomes11. This process improved the accuracy of the
algorithms to scan, understand and detect the disease in an efficient way. The
databases trained with this data helps provide predictions on the patient’s
potential to get a type of disease with a time period like number year the
patient can be affected with retinal disease based on other medical conditions
of the patient like diabetes.
·Scalability and Accessibility: AI tools enable mass screenings and extend
diagnostic capabilities to underserved areas. A study conducted by Lama et al
to examine the benefits of mobile, real-time, community based teleophthalmology
program required eye exam of large volume of the community who came forward to
have the eye exam done with the mobile unit in New York suburb area12. They were able to scan or conduct eye exam
for almost 957 participants. However, with the utilization of AI and technology
innovations these mobile units can scale to larger volumes as the manual
reading of images by an ophthalmologist can slow down the process. These mobile
units can help more suburbs and more underserved areas with low income and no
insurance people. AI reading images and interpreting results takes the eye care
to a next level. With more case studies conducted by the teleophthalmology
screening they concluded that there is an emerging need for larger volumes of
screening and identifying the disease at an early stage.
Table 1: A table comparing the manual and AI-assisted diagnostic
workflows.
|
Step |
Manual workflow |
AI Assisted workflow |
|
Patient
Consultation |
Patient visit
to the healthcare provider; health history is symptoms are discussed. |
Similar to
manual process, patient history is gathered. |
|
Visual Acuity
Test |
Assessing the
patient ability to read numbers and symbols on an eye chart is performed. |
This step
remains same as the manual workflow. |
|
Retinal
Imaging |
Fundus
photography, optical Tomography (OCT) or Fluorescein Angiography (FA) are
performed. |
The same
imaging techniques are used like in manual workflow. |
|
Image Review |
A specialist
manually reviews and analyzes the retinal images to check for signs of
diabetic retinopathy (DR), age-related macular degeneration (AMD) etc. |
Retinal images
are processed by AI algorithms using machine learning models trained on large
datasets to identify potential signs of eye disease. ML algorithms flags the
areas of concern |
|
Diagnosis |
Diagnosis is
made based on the image interpretation, clinical exam, history, diagnosis is
recommended. |
Technician
reviews the AI generated results, clinician focus more closely on the issues
identified or suggested by AI. Base don both AI and clinician review final
diagnosis is made. |
|
Treatment plan |
Treatment
options such as injections, surgery, laser therapy is discussed. |
Treatment is
determined similar to the workflow, but AI can assist in predicting the
disease progressions, helping to make more precise plans of treatment. |
Qingyu et al
conducted a study using the manual and AI assisted workflow of diagnosing the
retinal disease [13], with a set of clinicians participating in the manual
workflow and reviewing the AI workflow outcomes. The datasets used for training
ML algorithms improving with more training data and DL algorithms are most
highly utilized in retinal disease identification.
3. AI Technologies in Retinal Diagnostics
Artificial
Intelligence (AI) is growing in many medical specialties15, using different algorithms, applications,
devices. AI is defined as a branch of computer science to design the machines
capable of learning like human beings to solve complex problems. ML is a subset
of AI which is designed to process large volumes of data and can learn and
recognize patterns in data.
Neural networks
is a technique of ML that mimics the human brains with neurons. It consists of
input, hidden and output layers15. These neural networks can be used a variety of problem
solving like pattern recognition, classification, clustering of data,
dimensionality reduction, natural language processing (NLP), regression and
predictions etc.
Deep neural
networks (DNN) is technique of neural network as the name implied the layer of
this neural network is deeper with many layers of learning. The depth of DNN is
greater than 4, a multiplayer perceptron with more than one hidden layer is a
DNN framework16.
Deep learning
algorithms are built using the artificial neural networks which follow the
concept of deep neural networks with more than one hidden layer.
3.1. Deep learning algorithms
Deep learning
(DL) has already made a huge impact in medical diagnosis, self-driving cars,
speech recognition, predicting and forecasting15. There are many deep learning algorithms, few
of the are listed below, the detailed study on those network algorithms can be
studied in detail with the research by Shrestha and Mahmood15.
·Convolutional Neural Networks (CNNs)
· Long Short-Term Memory Networks (LSTMs)
· Generative Adversarial Networks (GANs)
·Restricted Boltzmann Machines (RBMs)
· Radial Basis Function Networks (RBFNs)
·Multilayer Perceptron (MLPs)
·Recurrent Neural Networks (RNNs)
·Deep Belief Networks (DBNs)
·Self-Organizing Maps (SOMs)
Each network
algorithm has its own advantages and disadvantages. CNNs being a type of neural
network specially designed for processing grid like data such as images with
grids of pixels and it consists of layers like
·Convolutional layer
·Rectified Linear Unit layer
·Pooling layer
·Fully connected layer
The
convolutional layer applies filters to the input data to extract features,
making the CNNs very effective for image recognition. The input data for
convolutional layer can be retinal images to render patterns and anomalies
across multiple layers.
CNN can vary
based on number of layers, pooling operations, region-based and attention-based
mechanism16,17, Below is the list of CNNs based on a layer
modification17.
·LeNet
·AlexNet
·GoogLeNet
· VGGNet
· ResNet
·DenseNet
CNN algorithms
like ResNet and DenseNet has become more popular due to their expert-level
performance in classifying retinal diseases18-21. DL based technologies have proven to achieve
equivalent diagnostic performance compared to clinicians in identifying the
retinal diseases.
3.2. Integration with imaging modalities
There are
investments and continuous improvements made towards integrating AI into
retinal imaging technologies. The IDx-DR is the first FDA approved AI system
designed to use in primary care setting for screening diabetic retinopathy22. This system is designed based on the deep
learning algorithm23, Grzybowski et al conducted a study to evaluate the results
based on IDx-DR and Medios AI systems developed using the deep learning
algorithms and the results demonstrated high accuracy and sensitivity.
24According to Dolar-Szczasny, J., Drab, A.,
& Rejdak, another product that was approved by FDC to help patients
self-administer OCT scans was “Notal Vision Home Optical Coherence Tomography
(OCT) System.” This Ai enabled home tool can help patients to self-screen and
provide the results their healthcare provider. AI enhances traditional imaging techniques such
as fundus photography, OCT and fluorescein angiography by automating image
analysis and improving diagnostic accuracy.
3.3. Case studies and tools
Promotion of
AI-enabled solutions for healthcare may improve the efficiency and patientcare soon.
It is also important to ensure the integrity of the AI systems to have positive
impact on eye care. In the recent year many studies have applied AI algorithms
to retinal imaging to improve the retinal disease diagnosis outcomes. Many
institutions and technology companies engage in AI research and continuing to
make new research and innovations.
Below is the
list of AI enabled systems that are approved by FDA to be used for diagnostic
screening of retinal diseases. The list is taken from FDA website25 of Artificial Intelligence and Machine
Learning-Enabled Medical Devices.
Table 2: List of all approved AI and ML enabled medical
devices for Ophthalmology by FDA. Source [25] FDA website of Artificial
Intelligence and Machine Learning-Enabled Medical Devices
|
FDA Approval Date |
Device |
Category |
Description |
Company |
|
5/15/2024 |
Notal Vision Home Optical Coherence Tomography (OCT) System |
Home Monitoring Ophthalmic Optical Coherence Tomography (Oct)
Imaging Device |
The Notal Vision Home Optical Coherence Tomography (OCT) System
is an Artificial |
Notal Vision Inc |
|
4/23/2024 |
AEYE-DS |
Diabetic Retinopathy Detection Device |
Software as a medical device designed to analyze digital fundus
images taken with an ophthalmic camera. Using artificial intelligence
algorithms, the device is able to determine whether a patient has referable
retinopathy. |
AEYE Health Inc. |
|
6/16/2023 |
EyeArt v2.2.0 |
Diabetic Retinopathy Detection Device |
Software as a medical device designed to analyze digital fundus
images taken with an ophthalmic camera. Using artificial intelligence
algorithms, the device is able to determine whether a patient has referable
retinopathy. |
Eyenuk, Inc. |
|
6/17/2022 |
IDx-DR v2.3 |
Diabetic Retinopathy Detection Device |
Software as a medical device designed to analyze digital fundus
images taken with an ophthalmic camera. Using artificial intelligence
algorithms, the device can determine whether a patient has referable
retinopathy. |
Digital Diagnostics Inc. |
4. Applications AI in Specific Retinal Diseases
4.1. Diabetic retinopathy
According to
International Diabetes Federation, there are 537 million adults living with
diabetes. The projections show that there are 783 million live with diabetes by
2045. There is 1 person out of 10 in the world population with diabetes. Diabetic
retinopathy (DR) is caused by diabetes in a person26. Diabetes can cause damage to the blood
vessels and high sugar levels effect the blood vessel that goes to retina, it
can damage the retinal soft tissue and cause diabetic retinopathy27. A common procedure of identifying the
diabetic retinopathy is the physician examines the eye retina with fundus
photography and captures high resolution images of retina to identify blood
circulation and anomalous vessels.
The initial
version of deep learning embedded system approved by FDA, the IDx-DR has been
tested and being used with upgraded versions. The IDx-DR has been successful in
identifying and classifying the fundus images to detect diabetic retinopathy
with more accuracy and efficiency22.
Pandey et al
developed and trained an ensemble of five CNNs to classify retinal fundus
photographs to detect and identify diabetic retinopathy (DR), age-related
macular degermation (AMD), glaucoma and normal retina [28]. As per the
literature, the deep convolutional ensemble (DCE) was trained with 43,055
fundus images and 100 new images for testing. These images were screened by the
board-certified ophthalmologists. The ophthalmologist’s results were 72.7%
accurate vs the 79.2% DCE algorithm-based results.
Chai et al
conducted a similar testing on the deep learning models on CNNs, their results
showed improved sensitivity, improved specificity compared to the retina
specialists for DR detection29.
AI detects
microaneurysms, hemorrhages and other signs of DR, enabling timely
intervention. Remote screenings powered by AI have improved accessibility in
underserved areas.
4.2. Age-related macular degeneration
Age-related
macular degermation is a leading cause of vision loss in may parts of the
world. Many of the cases left undiagnosed until they become an irreversible
issue. AMD us projected to reach 288 million by 204030,31. To address this issue, the AAO recommends
regular screening for people aged over 65. This explains the need for the AI
systems to assist in screening and identifying the disease with accurately and
with scalability to be able to screen larger volumes.
The most common imaging
modalities that are being explored in the AI fields for AMD is optical
coherence tomography (OCT). OCT angiography is also used in DL algorithms to
classify and detect AMD. There are several DL algorithms that were trained and
used to effectively classify AMD. Bhuiyan et al has applied six deep CNN models
to train and segregate images to identify different AMD patterns like dry, wet,
Pre and Late AMD predictions32. AI differentiates between wet and dry AMD and monitors
disease progression using OCT imaging, helping clinicians optimize treatment
plans.
According to the
Crincoli et al, there are efforts to study and apply DL algorithms to evaluate,
incipient AMD on a genetic point of view, predicting the progressions of AMD to
late and advanced using bio markers, the risk of AMD conversion to neovascular
ae evaluated by DL algorithms33.
4.3. Glaucoma
Glaucoma is an
eye condition that damages optic nerve on eyes retina. This damage can lead to
vision loss or blindness. Optic nerve is vital for the vision as it send signal
from eye to the brain, it is important to have healthy optic nerve34. A fluid called aqueous humor flows freely
through anterior chamber in the eye and exits through the drainage system. This
is called trabecular meshwork. If this drain system is blocked or does not
function well, the pressure in eye builds up and damages the optic nerve. In
most of the cases this leads to gradual vision loss35. Glaucoma usually happens at the age of 60
years but can happen at any age.
According to
mayo clinic, there are varieties of glaucoma with different stages. Open-angle
glaucoma, acute angle-closure glaucoma, normal tension glaucoma, pigmentary
glaucoma and glaucoma in children. The most common glaucoma is open-angle
glaucoma35.
AI algorithms
that detect and predict glaucoma has increased exponentially. Many parameters
like intraocular pressure, optic disc evaluation, retinal nerve fiber layer
measurement and gonioscopy have been evaluated by DL algorithms36.
OCT has become a
dominant imaging modality for glaucoma assessment37, according to Yousefi, a recent study with
five conventional ML classifications named as linear discriminant analysis,
SVM, recursive partitioning and regression tree, generalized linear model,
generalized additive model testing with normal and glaucoma’s eye images
achieved 95% specificity and 92.5% sensitivity37. While there are two FDA approved AL enabled
systems to screen DR and macular edema, there needs to be more algorithms to
monitor, diagnosis and prognosis of glaucoma.
According to
Yousefi, there are glaucoma progression algorithms based on linear regression
assume that glaucoma progresses linearly, while there is evidence that glaucoma
progress non-linearly, concluding that unsupervised AI algorithms can provide
unbiased results of the progression of glaucoma.
4.4. Retinopathy of prematurity (ROP)
Retinopathy of
prematurity (ROP) is an ocular disease that grows in premature babies or
newborns less than 3 pounds of weight. The retinal blood vessels of the infant
suffering from ROP develops more blood vessels and can affect the vision and
may even lead to blindness. Screening for ROP can help identify and prepare for
the treatment plan for the children diagnosed with this condition.
Wang et al has
developed large scale ROP databases with adequate clinical fundus images with labels
and developed an automated ROP detection system called DeepROP38, using deep neural networks (DNNs). They
created two network called identification network Id-Net and grading network
called Gr-Net. With the test set they were able to achieve 96.62% sensitivity
and 99.32% sensitivity. For grading they were able to achieve 88.6% sensitivity
and 92.31% sensitivity. They compared the model testing with human evaluations
the % of error is reduced from 6.7% in the first test to 2.1% in the fourth
test. They confirmed that DNNs can learn ROP features from a large set of test
data and provide efficient results on real cases.
Integration of
deep learning algorithms in detecting the disease, progression and monitoring
is giving satisfied results and more AI devices needs to be developed and
approved to help address the ROP in children.
5. Challenges and Limitations
·Bias in training data: In all the example see the model training was
done using the public available data sets. This can be a bias when there are
similar set of images used by several DL algorithms. It is important to train
the models with more participation from the real-world ophthalmology clinics. Models
trained on non-diverse datasets may underperform in certain populations.
·Regulatory hurdles: It is important to have stringent approval
process so that no product can be easily approved without supported evidence of
proven results. However, approval processes delay clinical implementation, it
is important to have more attention towards improving the approval process for
the clinical devices developed using AI.
·Ethical concerns: The AI has the “black-box problem”, which it
is not easy to explain what happens within the black box, the layers and the
weights and logics of deep learning algorithms. Issues like data privacy,
informed consent when there are hundred and thousand of fundus images and OCT
images of the patients, there needs to be a proper patient privacy and informed
consent put in place.
·Integration into clinical practice: Irrespective of technology advancements if practitioners
do not trust and use AI systems it remains a challenge for these AI systems. AI
demonstrates a powerful solution for the growing need of retinal disease in the
world.
6. Future Directions
·Explainable AI (XAI): To have better explanation of the black box of
AI, it is helpful to have enhanced transparency to build trust among
clinicians. The innovation and research towards explainable AI can help gain
this confidence.
·Wearable technology: AI-powered devices for continuous retinal
health monitoring.
·Open-source collaboration: There are several datasets that are available
for public use, but they are limited resources for the variety that is required
to train AI models. When healthcare providers share datasets the AI development
can improve global diagnostic standards.
·Integration with genomics: So far, the discovery and AI model’s
implementation is on training datasets using the images, combining AI with
genetic data for deeper insights into hereditary disorders can make more
meaningful predictions.
·Predictive analytics: The repository of millions of patient’s data
for a variety of eye diseases can provide a lot of valuable information to have
AI-driven predictions and to provide proactive disease management. With
diversified data predictive analytics can provide valuable insight for a
variety of demographics.
7. Conclusion
AI-driven
diagnostics are revolutionizing ophthalmology by enhancing early detection and
management of retinal diseases. By addressing the limitations of traditional
diagnostic methods, AI has the potential to reduce global vision loss
significantly. However, its widespread adoption depends on overcoming
challenges like data bias, ethical concerns and regulatory barriers. With
continued advancements and collaboration, AI is poised to reshape the future of
retinal care.
8. References