Hindawi

Journal of Healthcare Engineering

Volume 2023, Article ID 2728719, 18 pages
https://doi.org/10.1155/2023/2728719

Review Article
Deep Learning and Medical Image Processing Techniques for
Diabetic Retinopathy: A Survey of Applications, Challenges, and
Future Trends

Posham Uppamma and Sweta Bhattacharya

School of Information Technology and Engineering, Vellore Institute of Technology University, Vellore 632014, India

Correspondence should be addressed to Sweta Bhattacharya; [email protected]

Received 12 July 2022; Revised 28 October 2022; Accepted 25 November 2022; Published 2 February 2023

Academic Editor: Oana Geman

Copyright © 2023 Posham Uppamma and Sweta Bhattacharya. Tis is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.

Diabetic retinopathy (DR) is a common eye retinal disease that is widely spread all over the world. It leads to the complete loss of
vision based on the level of severity. It damages both retinal blood vessels and the eye’s microscopic interior layers. To avoid such
issues, early detection of DR is essential in association with routine screening methods to discover mild causes in manual
initiation. But these diagnostic procedures are extremely difcult and expensive. Te unique contributions of the study include the
following: frst, providing detailed background of the DR disease and the traditional detection techniques. Second, the various
imaging techniques and deep learning applications in DR are presented. Tird, the diferent use cases and real-life scenarios are
explored relevant to DR detection wherein deep learning techniques have been implemented. Te study fnally highlights the
potential research opportunities for researchers to explore and deliver efective performance results in diabetic retinopathy
detection.

1. Introduction the secretion of insulin hormones and hinders the ability of

patients to maintain proper glucose levels in the body,
DR happens to a person afected by diabetes, and conse- leading to multiple health issues. In a Type 2 diabetes sit-
quently, the eyes get afected. In diabetic patients, irregular uation, the insulin hormone does not function in converting
blood sugar levels are observed. Generally, glucose in the glucose to energy. On the contrary, the human body needs
human body needs to be transformed into some form of an adequate amount of energy to sustain a healthy and
energy while performing daily activities. When blood sugar normal life and hence becomes dependent on medication to
levels exceed their normal control limit, organs such as the support the defciency in the insulin level. In both cases of
retina, heart, nerves, and kidneys begin to sufer. In the Type 1 or Type 2, regardless of the blood sugar levels, the
presence of high blood sugar levels in the human body, there patient may be afected by diabetic retinopathy, and fur-
is a serious risk of eye blood vessels being afected to the thermore, it may lead to complete loss of eyesight [1].
point where they become completely obstructed, causing Te initial symptoms of DR include blurred vision and
blood leakage, the retinal vessels swell, and even new tiny eye foaters, and at an advanced stage, patients can lose
blood vessels are formed on the eye retina. Tis is known as complete vision if the disease is chronically sustained for
hyperglycemia; diabetes is categorized mainly into two types: a longer period. Tis disease is indicated by symptoms of
Type 1 and Type 2 diabetes. Type 1 diabetes is an incurable diferent types of eye retinal lesions. Tese are comprised of
disease that results from abnormalities in the insulin hor- hemorrhages (HM), microaneurysms (MA), hard exudates,
mone level, which is responsible for controlling blood and soft exudates. Patients who are aficted with the illness
glucose levels. Te evidence is that Type 1 diabetes reduces often fail to avail themselves of early diagnosis by identifying
2 Journal of Healthcare Engineering

initial symptoms. Hence, it is of utmost importance for

a person to ensure an eye examination is conducted at least Artificial
Intelligence
twice a year to eliminate or proactively avoid disease
complications. Considering the disease’s severity, diabetic
retinopathy is categorized, namely, as one that has a low
severity and is recognized at an early stage. On the other Machine
Learning
hand, PDR is of a higher severity level, being detected at an
advanced stage [2].
Various techniques are implemented for the diagnosis of
DR at an early stage, and machine learning has pre- Deep
dominantly contributed to the same. A subset of artifcial Learning
intelligence is a machine learning technique implemented to
improve the learning experience without human in-
tervention. Deep learning is also a technique that is pre-
dominantly used in various types of disease prediction. It is
a part of artifcial intelligence and machine learning, which
contribute to numerous developments in areas such as Figure 1: Relationship between the AI, ML, and DL.
computer vision techniques, image coloring, image cap-
tioning, medical image analysis, drug discovery, neural
networks, and even fraud detection systems. In deep 2.1. Review of CNN-Based Techniques. CNN is one of the
learning, large datasets are used on which feature extraction deep learning models which is mainly concentrated on
techniques are implemented to identify the most signifcant image classifcation and object identifcation techniques. Te
attributes, and then various data mining algorithms are study [6] highlighted that deep learning approaches have
implemented to perform data analysis and predictions. a signifcant impact on the automated detection of diabetic
Tese techniques work towards achieving optimized, ac- retinopathy in comparison to traditional detection methods
curate predictive outcomes without human intervention, as it yielded better accuracy, specifcity, and sensitivity. Te
and hence, disease predictions become trivial at early stages study examined PDR and NPDR in terms of low, medium,
saving the lives of millions of patients. Te following Figure 1 and high levels using the Messidor-2 data source consti-
shows the relationship between deep learning, machine tuting fundus photographs. Te IDX-DR X2.1 software was
learning, and artifcial intelligence. used for implementing CNN (convolutional neural net-
work)-based detection techniques so that microvascular
diseases could be easily identifed. Te authors [7] explained
1.1. Global Statistics on Diabetic Retinopathy. Te American the various severity levels of DR emphasizing DME (diabetic
Diabetes Association Position (ADAP) has observed sig- macular edema). A DCNN (deep convolutional neural
nifcant improvement in diabetic retinopathy detection and network) technique was implemented with a clinical scaling
treatment due to the prevalent use of optical coherence system to classify the diferent stages in disease identifcation
tomography and intraretinal pathology. Tese results enable methods when applied to the Digi fundus image dataset
the evaluation of the thickness of retinal fundus images and from Finland. Te grading aspects of retinal images were
generate results about microvascular lesions with enhanced explored and a measure of the disease’s seriousness was
accuracy [3]. Te statistical report also validates the fact that identifed with sensitivity, specifcity, and AUC (area under
diabetic retinopathy is a major public health concern in the curve). Te study [8] implemented a hybrid-deep
today’s society. Te prevalence of diabetes mellitus (DM) learning method for easy recognition of DR with fundus
since 1980 men’s rates have risen by 110 percent, while pictures using the EyePACS dataset from the Kaggle re-
women’s rates have risen by 58 percent, which will further pository. Also, linear-SVM techniques based on convolu-
increase by 9 and 7.9 percent globally in 2014. It is estimated tional neural networks were used to classify the images,
that there will be an increase of almost 422 million patients which led to enhanced performance.
sufering from diabetes worldwide in the next couple of Te authors in [9] presented a deep learning framework
years, which might further increase to almost 629 million by that implemented a CNN model consisting of three diferent
2045 [4]. Similarly, a computer-aided diagnosis system can layers to identify the diferent retinal layers that helped to
be used to reduce the burden on ophthalmologists for predict diseases such as DR, drusen, CNV (choroidal neo-
performing regular screening procedures. Also, various vascularization), and DME. Various preprocessing tech-
feature extraction methods, classifcation techniques, and niques were applied to get enhanced image quality, and
diferent algorithms are implemented for the early diagnosis optimization techniques were further implemented to
of DR [5]. remove unnecessary noise from OCT (optical coherence
tomography) images, ensuring that accurate results are
generated. Te authors in [10] mentioned diabetic reti-
2. Related Works
nopathy as being one of the most challenging issues. To
Tere exists some of the signifcant studies which are detect the level of severity, deep visual features are extracted
reviewed in the following sections. from the images that are D-color-SIFT (color dense in scale-
Journal of Healthcare Engineering 3

invariant), and the GLOH (gradient location-orientation a modifed deer hunting optimization algorithm is employed
histogram technique) was implemented. Te selected fea- for extracting optimal features for enhanced classifcation
tures from the three image datasets, namely, DIRETDB1 and accuracy [17].
MESSIDOR-2, are fed into the semisupervised multilayer
method. Te results are validated using the AUC, sensitivity,
and specifcity metrics. Similarly, the study [11] presented 2.3. Review of Deep Learning-Based Diabetic Retinopathy
ultrawide-feld fundus photographs that provided detailed Screening in Various Areas. Te initial survey is conducted
imaging of the surface of the retina. Te study helped to highlighting that most deep learning applications use image
automatically detect diabetic retinopathy using deep processing techniques for DR recognition. Te number of
learning algorithms such as the ResNet-34 architecture, individuals afected by the disease is expected to grow
consisting of 34 layers that enabled classifcation. Also, the rapidly in the next 22 years, especially in developing
image segmentation method ETDRS-7SF (early treatment of countries. Te WHAGAP (World Health Assembly Global
diabetic retinopathy study-7 standard feld) was proposed to Action Plan) has been intended to diminish this ever-
achieve enhanced results when validating the output using increasing number since its inception in 1990 and is still
various metrics. Te authors of [12] proposed a pretrained ongoing. Te objective is to identify and report all vision-
image classifcation technique using CNN models, namely, threatening problems and fnd ways to eliminate them. Te
AlexNet, VGG-16, and SqueezeNet, for classifying diabetic only way to reduce the occurrence of the disease is to initiate
retinopathy images. Te customized 5-layer CNN model was early and frequent screening of the disease at an early stage
a fully connected neural network classifer, which helped in among diabetic patients. Te traditional techniques fail to
the easy training of the image datasets, and the ReLU ac- serve the purpose, and ophthalmologists remain unable to
tivation function was used to achieve better performance. detect the severity level. Te use of machine learning detects
Te existing contributions to various diabetic retinopathy the disease at varying severity levels, reduces the com-
datasets with deep learning approaches are shown in Table 1. plexities involved in traditional detection systems, and thus
helps to save the lives of potential patients having a higher
chance of getting afected by the disease [18, 19]. Te
2.2. Review of DCNN-Based Techniques. Te study in [13] comparison between the normal eye and the diabetic
highlighted that a convolutional neural network is a major retinopathy-afected eye is specifed in Figure 2.
contributor to image prediction. Te study proposed Te present study highlights various information per-
a multistage transfer learning approach for various labeling tinent to diferent deep-learning approaches utilized for the
when applied to relevant datasets. Moreover, an automatic screening of diabetic retinopathy on both publicly available
identifcation of diabetic retinopathy using fundus photo- and real-time datasets. Te unique oferings of studies are
graphs generated Chen’s quadratic kappa score, sensitivity, the following:
and specifcity. Te study in [14] focused on the analysis of (i) Te basic information was relevant to the diabetic
microaneurysms at the early stages of DR and DME (diabetic retinopathy disease, its types, the cause, and its
macular edema) identifcation. A DCNN (deep convolu- presence globally which kindles the need to review
tional neural network) model with a lesion detection al- the diferent techniques involved in it
gorithm was applied to fundus images. A semantic
segmentation method categorized the pictures where oph- (ii) Te various imaging techniques and implementa-
thalmologists could easily detect the severity levels of pro- tions of deep learning procedures for diabetic ret-
liferative and nonproliferative diabetic retinopathy, ensuring inopathy screening are explored
enhanced accuracy and efciency. Te authors in [15] (iii) Recent studies and real-life examples of how image
aligned with the similar facts that diabetic retinopathy could processing can be used to help people with diabetic
be identifed using deep learning approaches and they can retinopathy are discussed in detail
predict various risk factors in patients. As part of the study, (iv) Various datasets are identifed, contemporary DL
the authors proposed a one-feld and three-feldinception- techniques are reviewed, and associated challenges
V3 architecture wherein validation was performed consid- are highlighted acting as scope for future study in
ering data from one eye selected randomly from patient data this area
of two datasets. Finally, an internal validation dataset and an
external validation dataset of 3678 and 2345 eyes, re-
spectively, were generated. Te model’s results were checked 2.4. Organization of the Study. Te information can be or-
with AUC, which produced better results than traditional ganized for this study as follows: Section 2 represents
systems. In [16], the authors presented an entropy-based a detailed review of studies conducted relevant to deep
enhancement technique for refning the features of the learning implementations on diabetic retinopathy. Section 3
images on the imbalanced dataset. Tis method resulted represents the background study on diabetic retinopathy
from an enhanced classifcation using a hybrid neural emphasizing on deep learning procedures for medical im-
network and low computational cost. Te deep visual fea- aging techniques. Section 4 provides an exhaustive review of
tures of DR images are enhanced using graph convolutional the medical imaging procedures utilizing deep learning
networks (GCNs) in association with relation- approaches for the screening of diabetic retinopathy. Section
awarechannel-spatial attention (RACSA). In addition, 5 highlights DR case studies conducted in diversifed
 4
Table 1: Various diabetic retinopathy datasets with deep learning implementations.
Outcomes and Pros and
Ref. Dataset Used method Research challenges
metrics cons
Pros:
MESSIDOR-2 fundus image Accuracy, AUC, sensitivity, (i) Efcient screening of DR Diversifed image datasets are not
[6] CNN for deep learning
dataset and specifcity Cons: considered
(i) Not cost efective
Pros:
(i) Deep learning CNN based on (i) Ability to increase size of
fve-stage diabetic retinopathy image without modifying Inclusion of high-resolution image
Diabetes image dataset + digi AUC, sensitivity, and
[7] screening structure data but exclusion of low-resolution
fundus Ltd. from Finland specifcity
(ii) Clinical grading system on macular Cons: image data
edema (i) Biased results in image feature
grading
Pros:
(i) Easier identifcation of DR
Severity level of the disease is not
[8] EyePACS dataset Linear-SVM in VGG-Net Sensitivity and specifcity Cons:
included
(i) Results lack required
satisfaction level
Pros:
(i) Efcient ocular structure
Accuracy, specifcity, (i) Use of OCT images alone in
OCT (optical coherence identifcation
[9] Tree-layered CNN sensitivity, precession, recall, implementation
tomography) images Cons:
F1 score, and kappa score (ii) Exclusion of fundus images
(i) Exclusion of variable eye
structure images in the dataset
Pros:
(i) Deep visual features (i) Deep visual features are used
(DVF) + D-color-SIFT GLOH for reducing time and image
DIRETDB-1, MESSIDOR-2, and ROC, sensitivity, specifcity, Inability to classify multiple images
[10] technique uncertainty problems
FAZ images dataset and training errors at a time
(ii) Semisupervised multilayer deep Cons:
learning algorithm (i) Inability to perform multiple
image classifcations at a time
Pros:
(i) Ease in training of huge retinal
Ultrawide feld fundus Segmentation of the huge retinal
Accuracy, AUC, sensitivity, surface image dataset
[11] photographs from St. Mary’s ResNet architecture surface fails to cover mid and far
and specifcity Cons:
hospital in South Korea areas of the retina
(i) Model dependent on single
data source for images collection
Pros:
(i) Enhanced classifcation
accuracy by the pretrained model Implementation confned to
RGB images from public CNN models-AlexNet and VGGNet-16 Sensitivity, specifcity, and
[12] Cons: specifc problem, generalized
MESSIDOR dataset SqueezeNet accuracy
(i) Methodology is much conclusion not appropriate
generalized to handle diversifed
problems
Journal of Healthcare Engineering
 Table 1: Continued.
Journal of Healthcare Engineering

Outcomes and Pros and

Ref. Dataset Used method Research challenges
metrics cons
Pros:
Shapley additive explanation (SHAP) (i) Achievement of decreased
IDRID (Indian association of Quadratic weighted Cohen’s Efective hyperparameter
multistage transfer learning approach, variation and improved
[13] diabetic retinopathy dataset) and kappa score, sensitivity, and optimization method are not
(EfcientNet-B4, B5, and generalization
MESSIDOR dataset specifcity included
SE-ResNeXt50) Cons:
(i) Lack of stability
Pros:
(i) Minimizes darkness of the
Deep-CNN model with lesion detection Implementations on larger datasets
[14] IEEE dataset (IDRID) ROC, accuracy, and efciency retinal lesions
algorithm are not performed
Cons:
(i) Limitations in the data source
Pros:
(i) Ability to identify several
Deep learning model lacks
Two datasets, Eye-PACS and 13 Inception-V3 architecture with hazards relevant to DR prognosis
[15] AUC including risk factors efciency, risk analysis, and is not
medical centers in Tailand one-feld and three-feld Cons:
cost efcient
(i) Does not include features
relevant to DR risks
5
6 Journal of Healthcare Engineering

Figure 2: Existence of diabetic retinopathy. Source: https://www.kaggle.com/datasets/sovitrath/diabetic-retinopathy.

locations. Section 6 consolidates various issues and chal- Adaptive Learning Big Data Biological Classification
lenges associated with deep learning implementations for
DR. Finally, the conclusions and future directions of re-
search are discussed in Section 7. Document Analysis Deep Vision System

3. Background Study Remote Sensing

Deep Learning
Applications
Data Flow Chart

3.1. Diabetic Retinopathy and Related DL/ML Techniques. Text Document

Summarization Recognition
DR is an eye retinal disease that is one of the most common
chronic diseases all over the world. It usually appears when Image
Healthcare
the person has had diabetes for a long period, which the Segmentation

disease may cause. Te major issue is that the disease does Image RADAR
Smart Parking
System
not reveal diabetic retinopathy’s early signs and symptoms.
Te disease was classifed as low, medium, or rigorous to Stock Market Object Detection
determine its severity. Numerous researchers have worked
Re-Identification
to prevent this type of disease complication using various Speech Recognition
System
machine learning and deep learning techniques. Deep
learning applications incorporating medical imaging tech- Figure 3: Deep learning applications in various sectors.
niques have been spotlighted due to their ability to generate
efective results. Tese techniques help in the classifcation produce better results. Te fre-fy algorithm and standard
and prediction of diseases without much efort. Also, the scalar strategies were used to normalize the data in the study.
CNN architecture enables achieving optimized and en- Te study in [23] proposed a Siamese-like structure with
hanced quality outcomes in comparison to traditional dis- a binocular network on CNN for taking fundus images of the
ease prediction techniques. According to the authors of [20], left and right eyes. Tose features adapt the transfer learning
DR is one of the most complicated diseases that results in method for classifcation and prediction. Te study com-
vision problems all over the world. Also, this study con- puted the quadratic kappa score as one of the metrics to fnd
centrated on image quality improvement using a contrast- accuracy and similarity between histogram matrix pre-
constrained adaptive histogram equalization model for diction and actual images. Te study in [24] proposed
image segmentation. Te image was optimized using a convolutional neural network with novel AlexNet, VggNet,
a Bayesian optimization technique, and the hyperparameter GoogleNet, and ResNet models for image classifcation to
tuning inception-v4 model improved the evaluation results. automatically recognize a set of fundus images. To enhance
Some of the applications of deep learning in various dis- the performance of image datasets, hyperparameters and
ciplines are specifed in Figure 3. transfer learning techniques are considered. Data stan-
Te authors [21] study explained an image classifcation dardization and data argumentation methods were imple-
approach using CLAHE (contrast limited adaptive histo- mented to improve the quality of images.
gram equalization) and HE (Histogram Equalization) Te advancement in CNN architecture addressed the
techniques. Te study implemented the image screening CNN pooling layer’s low accuracy issue when specifying
method using CNN on the MESSIDOR dataset. In [22], the image view characteristics. Matrix multiplication, dynamic
study employed the easy identifcation of retinal disease routing, and squashing functions are shown in CapsNet
using deep learning methods. Te image preprocessing architecture features on the MESSIDOR dataset for en-
procedures reduced the image’s dimensions using principal hanced results [25]. In [26], the study proposed three dif-
component analysis (PCA) and also eliminated noisy data to ferent neural network models and preprocessing techniques
Journal of Healthcare Engineering 7

to describe how to classify diabetic retinopathy fundus already been trained. Tis approach restricts a smaller
images. Te Fuzzy C-means algorithm trains fundus images training dataset due to the variance and amount of data
without sacrifcing image quality and the labels cannot be distribution. Te study in [38] presented a preprocessing
modifed. Te authors in [27] focused on retinal image le- approach and a contrast-limited histogram method to refne
sions to identify the severity level of the illness. Te study the images on a dataset and resulting in high-quality images.
explained a lesion localization model and a patch-based A multiclass classifer transfer learning method trained with
technique on a deep neural network for various layers of GoogleNet architecture was used to extract the image features
patches applied, wherein the localization of the red lesions to generate efcient results. Te study in [39] proposed au-
was conducted. Te results revealed improvement in AUC tomatic diagnosis approaches implemented on retinal images
level, sensitivity, and specifcity metrics. Te authors in [28] to identify the severity of the disease. In this proposed ma-
employed a convolutional neural network-based system that chine learning algorithm, the inception network architecture
works on an entropy image of a green component for is trained and tested on the EYEPACS dataset for classifying
classifcation. Along with that, the grayscale unsharp- the fundus images. Tese approaches yielded enhanced ac-
masking (UM) technique of image extraction used grey- curacy results in comparison to the traditional approaches.
scale images to improve the performance. In [29], the au- Tese deep learning and machine learning models also help to
thors proposed a bag of words model for detecting the level detect and prevent various health-related issues, for example,
of severity in diabetic retinopathy images on the basic ap- mental disabilities in humans [40]. Te CNN model in [41] is
plications of support vector machine and random forest. used to identify the patient’s condition based on image
Various preprocessing methods were included as multiclass emotion exposure. Moreover, deep learning methods were
classifers were used to classify various severity levels of the used to process biomedical signal information. Te prognosis
disease. Te authors in [30] developed a novel automated- of dermatological diseases and related CNN implementations
feature approach that worked on a large collection of fundus are discussed in [42]. A summary of the various studies
image datasets. Te proposed approach helped in the easy reviewed and discussed in this section is presented in Table 2.
identifcation of diabetic retinopathy using a computer-
aided model integrated with a visual heat map. Te study
in [31] proposed a new dataset named DDR that was used for 3.2. Performance Evaluation of the Reviewed DL/ML
fundus image collection. Tose images for lesion recognition Implementations. Te various DL/ML techniques used in
and segmentation at each level recognized the four stages of DR diagnosis are critically reviewed. Each of the imple-
lesion detection annotations. Te study in [32] presented mentations were evaluated using commonly used metrics
a deep learning interpretable classifer to organize retinal such as accuracy, specifcity, sensitivity, precision, recall, and
images into various levels concerning the condition. In this ROC. Tis section presents a comparative analysis of the
approach, a distribution score was obtained in the last layer performance of these techniques as published in the
of image pixel classifcation, and the pixelwise score prop- reviewed study. Te section also identifes the limitations of
agation model identifed the image visual maps. each of the implementations which enable researchers to
Te DNN (deep neural network) concentrates on a lay- understand the associated challenges thereby providing
ered approach for feature extraction methods to compute ideas for model optimization. Te consolidated parametric
implicit results. Similarly [33], the authors surveyed diferent evaluation for performance analysis is presented in Table 3.
computer vision-related approaches for classifcation, pre-
processing, and feature extraction methods recognizing the 4. Significance of Deep Learning Applications
irregularities in retinal blood vessels. Te study in [34] Using Medical Imaging Techniques
presented the training and test datasets, and DCNN worked
to classify fundus photographs. Tese outcomes indicated DR Applications of deep learning have had a signifcant impact
and the presence of DME (diabetic macular edema). Te [43] on medical image analysis over the past few years. Tese
authors in [35] proposed to fnd the severity level of the applications have helped ophthalmologists easily detect
disease using fundus image classifcation techniques and an various eye diseases and identify the severity levels of patient
image normalization approach to detect the size and shape of conditions relevant to the disease. Te signifcant applica-
the images. On the basis of the processed images, an tions of deep learning for medical image processing are
inception-v4 algorithm was used to enhance the quality of an specifed in Figure 4. Te study in [44] stated that multiple
image. Te study in [36] proposed machine learning algo- people were identifed as sufering from any eye disease at its
rithms such as SVM, KNN (K-nearest neighbor), and bagged initial stages by the ophthalmologists using various types of
trees applied to diabetic patients’ physical health records to classifcation rules and prediction techniques to generate the
easily identify diabetic retinopathy. Te model generated best results. Te study in [45], explained that the layered
a superior level of accuracy using the bagged trees prediction approach of CNN helped to improve the accuracy of pre-
method. Te authors in [37] proposed a collective intelligence dictions concerned with fundus photographs.
approach (human + eye) for enhanced results in comparison Te authors in [46], addressed a convolutional neural
to the traditional artifcial intelligence techniques to fnd the network model, which was implemented as an efective way
various categories of diabetic retinopathy. Te proposed to classify and predict the images in diferent layers leading
method used EfcientNetB3, EfcientNetB4, and Ef- to optimized performance. It provided real-time classifca-
cientNetB5 to improve the accuracy of images that had tion techniques to employ deep network features to achieve
 8
Table 2: Various signifcant machine and deep learning methods in medical imaging analysis.
Ref. Dataset Used method Outcomes and metrics Research challenges
HPTI (hyperparameter tuning inception)-V4 Sensitivity, specifcity, accuracy, and Implementation of diferent classifcation
[20] MESSIDOR database
model precision factor are models not included
Histogram equalization and limited adaptive Sensitivity, specifcity, accuracy, Alternative medical databases are not
[21] MESSIDOR database with 400 images
histogram equalization methods precision, recall, F-score, and G-means considered for performance evaluation
Principal component analysis (PCA) with frefy Accuracy, precision, recall, sensitivity, Not suitable for high-dimensional data in
[22] MESSIDOR-2 dataset
algorithm and specifcity various domains
Use of limited fundus image data having
CNN-based binocular network with Sensitivity, specifcity, and quadratic
[23] EyePACS dataset missing values are collected from same
siamese-like structure kappa score
patient
Dataset with 35126 fundus images from Feature transfer learning and hyperparameter Variable image classifcation methods are
[24] Accuracy, sensitivity, and specifcity
kaggle tuning method not considered to improve model accuracy
Complete classes of image datasets are not
[25] MESSIDOR dataset Capsule network architecture Accuracy
trained in CapsNet
System is not implemented in GPU
[26] Dataset has 2000 images from the Kaggle Fuzzy C-means algorithm Accuracy environment, considered limited data
sources
Standard diabetic retinopathy database CNN with entropy images grayscale
[27] Accuracy, sensitivity, and specifcity Noninclusion of larger datasets
(DIARETDB1) unsharp-masking (UM) method
Considering distinctive labelling model is
[28] Kaggle dataset of 21,123 images Patch-based DNN Accuracy, sensitivity, and specifcity
costly, use of limited image data
Alternative image classifer methods are not
[29] Dataset has 240 images taken from Kaggle SVM and random Forest techniques Accuracy
implemented
Data-driven algorithm for deep feature Time-consuming and increased sensitivity
[30] MESSIDOR2 and E-ophtha databases Sensitivity, specifcity, and AUC
extraction for dataset with multiple variances
Detection and segmentation of fundus
Average precision (AP) and IoU
Deep neural network algorithms-VGGNet-16, image lesions from various perspectives are
[31] DDR dataset (intersection over union) metrics for
ResNet-18, GoogleNet, and DenseNet-121. critical, use of larger image datasets are not
each type of lesion identifcation
considered
Sensitivity and specifcity performance Pixel score identifcation process is not
[32] Kaggle repository dataset Score propagation deep learning model
for lesion recognition included
Feature extraction + image segmentation Sensitivity and specifcity performance Hybrid robust DL methods are not included
[33] STARE & DRIVE image dataset
method for lesion recognition for performance enhancements
Excludes the possibility of model
[34] EyePACS-1 and Messidor-1 dataset Deep-CNN-based model High sensitivity and specifcity
implementation in clinical environment
STARE, DIARETDB1, MESSIDOR,
[35] DRIVE, STARE, REVIEW, and E-ophtha Deep CNN model with inception-V4 algorithm Accuracy, precision, and recall Data from other domains are not included
datasets
Data collected from patients in central PCA (principal component analysis) and linear
[36] Accuracy Larger datasets are not considered
India regression
Exclusion of balanced dataset leading to
[37] EyePACS dataset Gaussian flters and EfcientNet Kappa score and accuracy
reduced efciency
CLAHE method and CNN + transfer learning
[38] Messidor-1 Accuracy Exclusion of minor diseases
approach
Sensitivity, specifcity, accuracy, and Automated image prognosis system is not
[39] EyePACS dataset Deep-CNN + inception network
Journal of Healthcare Engineering

precision included
 Table 3: Te performance analysis report of various DR detection techniques.
Ref. Methods Metrics (%) Performance analysis
Accuracy � 99
Specifcity � 98
[20] Hyperparameter tuning inception V4 model Need employment of versatile classifcation methods
Sensitivity � 99
Journal of Healthcare Engineering

Precision � 97
Accuracy � 97
Specifcity � 98
Limited adaptive histogram equalization Sensitivity � 94
[21] Image quality needs to be improvised
methods Precision � 97
F1 score � 94
G-Means � 94
Specifcity � 82
[23] CNN-based binocular network Sensitivity � 7 Availability of limited data leading to biased results
Kappa score � 0.824
[25] Capsule network Accuracy � 95 Training of limited image features
CNN-based approaches
[26] Fuzzy C-means Accuracy � 98 Inclusion of limited datasets
Accuracy, precision, and recall evaluated DR
[33] Feature extraction + segmentation methods Need for robust DL techniques to improve results
severity
Sensitivity � 91
[34] Deep CNN model System excluded possibilities implementation in clinical environment
Specifcity � 90
Accuracy � 88
Information from other domains with plausible impact on results are
[35] Deep CNN with inception-V4 model Precision � 96
not included
Recall � 94
AUC � 0.68
[37] EfectiveNet method Biased results are generated from unbalanced dataset
Kappa score � 0.36
Accuracy � 92
Specifcity � 94
[39] Deep CNN with inception method Automated image prognosis method is not implemented
Sensitivity � 81
Precision � 93
Accuracy � 97
[24] Transfer learning approach Specifcity � 95 Biased classifcation results
Transfer learning approaches
Sensitivity � 92
[38] CNN with transfer learning Accuracy � 74 Inability to detect minor diseases
9
 10

Table 3: Continued.
Ref. Methods Metrics (%) Performance analysis
Accuracy � 97
Specifcity � 95
[22] PCA with frefy algorithm using DNN Sensitivity � 92 Low-dimensional data are not considered
Precision � 96
Recall � 96
Accuracy � 97
[27] DNN patch-based approach Specifcity � 95 Use of limited data leading to enhanced cost
Deep neural network-based
Sensitivity � 92
approaches
Sensitivity � 94
[30] Deep feature extraction method Specifcity � 98 High computational cost
AUC curve � 0.97
Average precision (AP) � 0.88
[31] DNN algorithms Lack of larger datasets
IoU (intersection over union) � 0.17
Sensitivity � 91
[32] Score propagation method Image pixel score identifcation model is not included
Specifcity � 90
Machine learning-based [29] SVM and random forest techniques Accuracy � 90 Enhanced image classifer methods are not implemented
approaches [36] PCA with linear regression Accuracy � 92 Larger datasets are not considered
Journal of Healthcare Engineering
Journal of Healthcare Engineering 11

Prostate Image Fetal Localization Tyroid Diagnosis

Analysis

Brain Tumor Deep Learning Lung Nodule

Detection Classifcation

Covid-19 Diabetes Identifcation

Cancer Detection Cardiac Functioning
Detection
Figure 4: Applications of deep learning for various medical image processing.

accurate results. Te study in [47] discussed medical imaging the level of severity [51]. Te study in [52] surveyed various
processes that were used as an efective way to grade the traditional works that were performed using several
severity levels using a neural network classifer for fundus screening methods for early detection of DR. Tese
images. A multistage classifcation deep neural network was screening methods are used for image cropping, rotating,
used in association with a shallow dense layer, being applied and resizing aspects to fnd mild variations in retinal vessels
to the diabetic retinopathy dataset that generated grading of that generate optimized outcomes.
the disease with enhanced accuracy. Te study also imple- Additional approaches for evaluating DR levels at the
mented a graded equalization method to improve the ef- initial stage include using the visual graphic (VG) method
fectiveness and transparency of the pictures for for the classifcation of images and then the performance is
microaneurysm detection. calculated using error-correcting output codes (ECOC)
classifer [53]. Te study in [54, 55], explained the use of
pretrained CNN models with a transfer learning approach
4.1. Classifcation Techniques. In image processing, the for the refnement of images. Tese models classify images
fundus images are classifed using a CAD (computer-aided using a support vector machine (SVM) considering publicly
diagnosis) system by implementing various imaging tech- available databases. DR detections and lesions identifcation
niques to achieve early diagnosis and clarity regarding pa- problems can also be resolved using deep MIL (multiple
tient health conditions. Te previous studies explain several instance learning) wherein handling of the improper DR
preprocessing procedures for DR-related images and classify lesions is performed using a well-trained framework to
the presence of DR in a patient. Studies have implemented achieve enhanced quality Te earliest sign of retina disease
various classifers incorporating CNN layers to predict the include the identifcation of exudates, namely, soft and hard
disease. Also, multicast lesion detection procedures and exudates. Tese exudates were identifed in terms of yellow
modifed-CNN applications are implemented on fundus color cotton wool spots using morphological methods in-
images to classify normal PDR and severe NPDR, thereby volving multiple numbers used for the fltering of the retina
generating a PDR histogram, threshold edge values, accu- lesions. Te authors in [56, 57] proposed a training method
racy, and specifcity [48, 49]. Te authors in [50] used an that efciently identifed the existence of exudates using
automatic screening tool for reducing manual interruption CNN-based framework. Various studies show how nor-
and ensuring the diagnosis of the disease in a shorter time malizing pixel values for training and testing images can be
frame. Tese screening methods are implemented in dif- performed with the help of image-enhancing techniques. In
ferent datasets to classify and detect the DR grade levels [58], the authors proposed a multilayer deep CNN and SVM
thereby implementing diferent levels of image-enhancing classifer method that performed DR segmentation and
rules such as normalization and extension of the quality of detection. Tese retinal images were preprocessed using the
lesions. Tese lesions can be classifed using image-level gaussian fltering method. Tis method also explained about
descriptors to monitor the patient’s condition and grade ways to fnd the age-related retinal abnormalities along with
12 Journal of Healthcare Engineering

DR classifcation. Te authors in [59] presented an optical in [68] explained retinal images based on color character-
coherence tomography (OCT) approach using 3D-Markov istics extraction with an ensemble machine learning ap-
Gibbs random feld (MGRF) for extracting higher-order proach. In this approach, the image analysis, image
image features. Tese image features are applied to each cropping, resizing, and removal of irrelevant image data
layer of the artifcial neural network (ANN) classifer to were performed using the data source available from the
detect DR disease more accurately. Kaggle repository. Finally, the improved classifcation
metrics were achieved under the ROC curve. Te study in
[69] presented a novel transfer learning approach for
4.2. Detection Techniques. Detection means identifying an
detecting the exudates in retinal images using ResNet-50,
object or specifc part of important data in images. Tis
inception-v3, and VGG- 19 techniques. Tese were trained
mechanism mainly worked on the medical imaging process,
using a CNN model that had already been trained, and then
identifying the defects, and security of the surveillance
they were applied to one-Ophtma and DIRETDB1 data
system. Tese approaches are bound with computer vision
sources to get accurate classifcation results.
techniques that computed the human vision system com-
plications. In this scenario, artifcial intelligence techniques
computed enhanced results in image recognition. When it 4.3. Segmentation Techniques. Segmentation is a technique
comes to multiimage recognition methods, machine for partitioning images into subgroups or segments to re-
learning and deep learning mechanisms are efectively duce complexities in subsequent processing and analysis.
performed. Te study in [60] proposed the detection of Te purpose of segmentation is always to change the ori-
medical images using deep learning applications. Tis study entation of an image so as to evaluate it more easily. Also, it
implemented blood vasculature method to detect the retina could be used to partition the image into sets of pixels, and
blood vessels’ position in the afected area on images. Tese each pixel gives a clear contrast. Tis technique yields a high
vascular results are used by ophthalmologists to give degree of accuracy and enhanced results. In [70], the study
treatment to patients at early stages. Te study in [61], proposed a segmentation method that worked on several
explained the SVM method to detect NPDR. Te SVM tissues in the human brain, facial parts, ears, and eyes easily
classifer estimated the disease’s severity using around 400 segmented on the visual Chinese human head (VCH) model.
test images from the CHASE dataset. Apart from that, Deep learning-based techniques were used on fve separate
a confusion matrix was used to evaluate the results. MRI imaging datasets to identify the disease’s consequences.
Another study in [62] developed a hybrid machine learning It results achieved enhanced accuracy in comparison to the
method as a detection tool to identify if a patient has reti- manual segmentation processes. Tis technique had limi-
nopathy or not using retinal images. Tis technique used tations and is not suited to very big MRI datasets. In [71],
multiple instances of learning strategies and feature extraction Te study described how to predict leaf disease and segment
techniques for the results generated. Te study in [63], sur- it for easy identifcation, as well as how to use color
veyed the results of numerous studies on diabetic retinopathy transforms to fnd the afected leaf. Tis study used a k-
disease detection and classifcation issues. It also reviewed means clustering algorithm to categorize disease symptoms
microaneurysms, exudates, hemorrhages, and optic disc dis- and separate them as clusters at various stages. Te study in
orders and their prognosis details. Tese were used to create [72], proposed a new wavelet-based image segmentation
awareness on potential future research studies. Te authors in method that involves transforming input images into dif-
[64] developed a DTL (deep transfer learning) approach to ferent orientations while maintaining the standard pixels
detect the various layers of CNN architecture using the needed to detect image alterations. To state a clear repre-
inception-v3 algorithm. Tese methods divided the network sentation of the specifc item, we used morphological ap-
into depths and widths to reduce noisy data for enhanced proaches to retain specifc image segment frontiers.
computation results such as sensitivity and specifcity. Te authors in [73] proposed a deep-learning archi-
Te study in [65], proposed an image processing method tecture with 2D and 3D images for cardiac function seg-
instead of using lesion segmentation and implemented an mentation. In this work, 2D modifed U-Net architecture
image classifcation method. Tese approaches helped to was implemented in CNN models segmented on short-axis
identify the retinal lesion patches and classifed viable retinal MR images, resulting in enhanced performance. Te authors
vascular along with encoding the texture of the image with in [74] stated that segmentation applications play an im-
local binary patterns. Te study in [66] demonstrated an portant role in medical images and Francis’ investigation of
image recognition system for the neurovascular retinal iris, biometric, and vein segmentation for real-world
detection method that used a matching flter and a fuzzy problems. Deep learning models used encoder-decoder
entropy-based technique. Tis technique employs a gaussian architecture and pretrained classifers such as VGG-Net,
flter to precisely recognize small retinal blood vessels in DenseNet, and ResNet to get information that could be used
datasets such as MESSIDOR and DRIVE with precise on a large scale. Te study in [75] explained MRI brain
measurements. Te study in [67] proposed toboggan seg- tumor images and observed a typical tissue to remove noise.
mentation and the multiagent process that worked on the Te proposed method concentrated on decreasing the noise
two-sided segmentation of damaged retinal vessels. Tese from a tumor cell and used image segmentation models such
were assessed to enhance the quality of the image using as FSSN (Fernandez-Steel Skew Normal), GMM (Gaussian
flters such as the gaussian and modifed kirsch. Te authors Mixture Model), and Fuzzy C-means.
Journal of Healthcare Engineering 13

4.4. Registration Techniques. Image registration is an im- immediate demand for AI-based DL methods combined
portant aspect of the imaging system because it aids in the with radiographic imaging technology in Mexico to ensure
completion of certain tasks in medical image processing. In the rapid expansion of artifcial intelligence (AI) methods in
feature and intensity-based classifcations, image attributes diagnosis, prognosis, and the provision of quality medical
such as image alignment and diferent measurements are services. ARIA (automated retinal image analysis) is a basic
required to acquire results from the target image. Te au- premise of a web-based resource platform consisting of an
thors in [76] proposed various image registration frame- image analysis module for referable and nonreferable DR
works for enhanced outcomes based on picture longitudinal classifcation. Te hospital administration provided high-
values. Tis study focused on removing noisy data from quality services to eye technicians to check the patient’s
retinal vessels for enhanced outcomes using quadratic, lower condition and determine the severity of the illness. Both the
order, and elastic transformation framework models that ophthalmologists’ and the ARIA system’s results were used
eliminate errors. Te study in [77] proposed fuorescein to determine the patient’s condition. It is an extreme en-
angiography (FA) and optical coherence tomography (OCT) vironment to recognize chronic eye diseases such as glau-
methods that were identifed as scanning laser ophthal- coma and the DR diagnosis. It also required more
moscope images for future extraction purposes. Te re- ophthalmologists to meet the extremely high standards.
sultant extracted features were highly successful in
generating global registration results of the complex blood
vessel networks pertinent to retinal images. Also, deformable 5.2. Use Case 2: Andalusian Reginal Government (Spain) to
and intensity-based transformations were implemented to Develop an Automatic Diagnosis Diabetic Retinopathy Pre-
improve the motion magnitude of FA and OCT images. vention Strategy for the Early Stages. Te study in [83]
Te study in [78] proposed a supervised multispectral addressed the Andalusian Reginal Government of the health
fundus image registration (MSI-R-NET) method by ana- ministry of Spain and established an automatic diagnosis
lyzing various levels of blood vessel structures in the retina. preventive system at the beginning stage of diabetic retinopathy.
It also observed the eye movement of retinal vessels for Te study used supervised classifcation algorithms on 1058
a short period and worked on image irregularity positions. images of 529 diabetic retinopathy patients identifed by three
Tese multispectral images were forecasted without classi- eye specialists. Tese experts were from Huelva’s Juan Ramon
fcation and improved outcomes in the training and testing Jimenez Hospital (JRJHH), Cordoba’s NASMC (North Area of
parts. Te retinal vessel complications lead to complete Sanitary Management), and Seville’s VMUHS (Virgen de
vision loss in adults. Te authors in [79] presented a novel Macarena University Hospital) who worked on ground truth
volume of interest (VOI) for various levels of OT image image assessments to identify the earliest pace of the disease.
registration, which ofered superior results in B-spline Te complication of diabetic retinopathy caused blindness in
transformations, utilizing a stochastic gradient descent 4–11 percent of Spaniards. To ensure the safety of patients, the
optimizer. Also, it used the Jacobian determinant trans- researchers used an automatic prescreening method to detect
formations expansion on retinal vessels and monitored hemorrhages and microaneurysms. Tis system thus needs to
precise values. Te study in [80] mentioned the fact that concentrate on all eye-related diseases for diabetes patients.
complete image rotations, intensity, and scaling factors are
all key challenges in image registrations. To achieve ap- 5.3. Use Case 3: Monitoring of Diabetic Retinopathy at the
propriate image feature augmentation, the feature descriptor Initial Stage in Chinese Patients-A Case Study. Te authors in
statistical properties (FiSP) models were employed to ensure [84] developed the Triglyceride-Glucose Index (TyG) to an-
minimal scalability. Te resized parameters were regulated alyze the diabetes results in Chinese patients. Diabetic patients
efciently, and the resultant outcomes were superior in were registered in 2018 and evaluated by the endocrinology
comparison to traditional approaches. Te authors in [81] department of a diabetic eye disease center at China Medical
focused on enhancing the accuracy of retinal images using University. Te two ocular clinical experts on the team
the SIFT (scale-invariant feature transform) registration classifed the patient’s condition as mild, moderate, or severe.
method. Tis method helped in depicting an image of Tis study concentrated on the results of the TyG index, BMI
a higher pixel value without using noise data. Te automatic (body mass index), and HbA1 (hemoglobin A1c) to predict
image scalability and image alterations were applied to the the level of diabetic retinopathy in people who have had
FIRE dataset to obtain robust values to enable refned scaling diabetes for a long time. Tese advancements will be used in
and brightness of images, ensuring enhanced performance. the future to enhance diabetic retinopathy prognosis and
detect other eye diseases in diabetic-assisted patients.
5. Deep Learning Algorithms for Diabetic
Retinopathy Diagnosis: Case Studies
5.4. Use Case 4: Diabetic Retinopathy Afected Regions in
5.1. Use Case 1: Remote Image Analysis Was Used to Diagnose America, China, and India. Te authors in [85], highlighted
Diabetic Retinopathy in Mexican Patients. Te problem of diabetic retinopathy to be the primary reason for blindness
diabetic retinopathy occurs in young- and middle-aged in America and 99 percent of cases in India. India and
people. Te study in [82] suggested the development of China have over 90 million patients sufering from diabetis
an assistive index framework for identifying diseases that are which could result in unprecendted number of individuals
unlikely to be controlled. Tis study responded to the becoming blind in the next few years unless detected early
14 Journal of Healthcare Engineering

enough. Aravind Eyecare hospitals provide high-quality It is still an ongoing process to be found in deep learning
ophthalmic care. In this study, WiLDNet (Wif-based previous work problems, as shown below:
Long Distance Networks) technique was used to com-
(i) Most of the cases in traditional works lacked the high
municate with patients in rural locations over the online
resolution of quality images in larger datasets. As
platform. Tis system ran a mobile health organization for
a result, creating a proper dataset that collects images
the people who lived in rural areas, which increased the
of the required quality is necessary. Tose images
number of patients with favorable prognoses. In this study,
came from several sources, including public and
various layers of retinal vascular architecture and the
private, and are combined to yield positive results.
system’s overftting of specifc dimensions methods
implementations cannot be precise. To address such issues, (ii) Improved image enhancement and contrast are
the organization intended to use a large fundus image needed to detect damaged retinal vessels. Because
dataset to improve its performance. retinal images are small and noisy, it is difcult to
identify the disease in the early stage.
(iii) Te lack of standardization in data collecting is one
5.5. Use Case 5: Computer-Assisted Screening and Detection of the most signifcant issues in medical image
System in Europe Countries: A Case Study. In this study in analysis. It is vital to remember that as the amount
[86], according to the World Health Organization (WHO), of data available grows, so does the necessity for big
approximately 422 million people in low- and middle- datasets to ensure that deep learning models pro-
income countries sufer from diabetic retinopathy. Many duce realistic results.
people in Europe struggle with vision impairment which can
lead to total blindness as well as eye screening programs for
diabetic patients are inefective. During the COVID-19 6.2. Issues and Challenges of Diabetic Retinopathy. Te fol-
pandemic, many patients were diagnosed with telemedicine lowing are the challenges and concerns specifc to medical
and internet connectivity to consult doctors via online media image analysis for deep learning applications in the early
sources. To accomplish these methods one should be in screening of diabetic retinopathy:
online mode to save time which is necessary to raise the
number of persons registered to a large extent. In [87], the 6.2.1. Limited Real-Time Data for Training. Training a large
authors discussed fractional-order flters for edge identif- number of images in deep-learning applications is one of the
cation in biomedical imaging technologies. Te study most challenging tasks. Te activity of an image dataset with
employed left-sided and right-sidedfractional-order mask limited images may be tough to integrate with real-time data
flters to fne-tune the image with whole edges detected from other sources, such as hospitals. Also, it obtained
without noisy data and the performance evaluation of enhanced classifcation and detection accuracy when
fundus images on the STARE dataset. Tis process provides working with larger datassets.
complete noise-free images for disease diagnosis.
In the study in [88], the authors proposed a case study to 6.2.2. For Medical Images Collective DL Architecture. In
prevent risk factors in diabetic retinopathy in diabetic patients medical image processing, transfer learning approaches were
at Tikur Anbessa hospital. A person sufering from diabetes used to identify object detection and classifcation to achieve
and related complications fail to maintain the required glu- enhanced accuracy. But there are scenarios where these
cose level in the body. On the contrary, such individuals tend approaches failed to deliver optimal classifcation accuracy.
to eat healthy diet which have contrareactions of increased In such cases, it becomes necessary to retrain the model to
blood pressure level and consequential neovascularization. achieve enhanced accuracy.
Tis study suggested frequent health check-ups and glucose
levels were used to identify diabetic retinopathy problems in
the early years. Te study in [89], described the signifcant 6.2.3. Te Process of Incorporating DL Applications and
factors for the occurrences of diabetic retinopathy in those Telemedicine. In rural areas, neither medical nor hospitality
who had diabetes recognized in Debre Markos Referral resources are available to diagnose health issues, especially
hospital in Northwest Ethiopia. Various patients from this eye diseases, for which there are no routine screenings. Te
hospital maintained retinal vascular images examined using integration of telemedicine with cloud data sources enables
binary and multiple logistic regression approaches. Te pri- the use of AI techniques to upload images of the retina.
mary aim of this study is to prevent the occurrences of this
disease and conduct a systematic review of screening methods 6.2.4. Pool-Based Data Sampling. In this article, traditional
to maintain a healthy glycaemic level. image datasets are usually classifed as either deep learning
or machine learning models. Te main concern is that it uses
6. Various Issues, Challenges, and a large volume of records for DR classifcation and uses both
a large amount of processing capacity and memory.
Lessons Learned
6.1. Lessons. Te processing of medical images in a deep- 6.2.5. Low Computational Power and Network Size. Two
learning environment ensures accurate results. In this study, major factors contributed to the deep learning model’s
various methods stated the performance level in 2D images. overftting needs: computational power and network size.
Journal of Healthcare Engineering 15

An object-based screening model is identifed as more lesions easily [92]. To develop a system with minimal
comfortable in comparison to the image-based screening confguration, overftting and computational cost need to be
method popular in medical image processing that examines reduced.
hemorrhages and other ocular lesions. To overcome the data argumentation issues mentioned
in [93], use of more features on the large dataset need to be
analyzed to improve accuracy. Both data preprocessing
. Conclusion and Future Directions
methods and feature enhancement techniques are combined
7.1. Conclusion. Deep learning is regarded as an efective to design a robust hybrid deep learning model for detecting
technique for ofering technical solutions in disease pre- the DR classifcation and detection. Tese methods were
diction and classifcation. Tis study discusses various used to evaluate the information about diferent patients. It
machine and deep learning techniques for the early di- also provides the best outcomes for DR and other eye-related
agnosis of diabetic retinopathy. Tese studies were con- diseases in further proceedings.
ducted in various geographical locations wherein various DL
applications for medical image processing were imple- Data Availability
mented in the last century. DL has been used to achieve DR
identifcation solutions employing medical imaging tech- Te datasets and material used to support the fndings of the
niques such as classifcation, detection, segmentation, and study can be obtained from the corresponding author upon
registration. But, all these various applications of DL in- request.
corporating medical image processing methods require
enhancement in accuracy, performance, and reduction in Conflicts of Interest
computational cost. Tis study presents an exhaustive review
of the various deep learning and image processing-based Te authors declare that they have no conficts of interest.
techniques and relevant applications of the same in DR
detection. It also enlists the challenges identifed in the References
earlier studies, namely, relevant to the availability of real-
time data, computational power, network size, and various [1] E. J. Duh, J. K. Sun, and A. W. Stitt, “Diabetic retinopathy:
others. Finally, the study mentions certain recommenda- current understanding, mechanisms, and treatment strate-
tions for future scope in research that would enable diabetic gies,” JCI insight, vol. 2, no. 14, Article ID 93751, 2017.
retinopathy detection at an early stage. Tese research de- [2] T. R. Gadekallu, N. Khare, S. Bhattacharya, S. Singh,
tections involved facilitating efective hybridization tech- P. K. R. Maddikunta, and G. Srivastava, “Deep neural net-
works to predict diabetic retinopathy,” Journal of Ambient
niques and the implementation of advanced hyperparameter
Intelligence and Humanized Computing, pp. 1–14, 2020.
tuning methods to overcome some of the prominent chal- [3] S. D. Solomon, E. Chew, E. J. Duh et al., “Diabetic retinopathy:
lenges identifed in DR detection using DL techniques. a position statement by the American diabetes association,”
Diabetes Care, vol. 40, no. 3, pp. 412–418, 2017.
[4] R. Cheloni, S. A. Gandolf, C. Signorelli, and A. Odone,
7.2. Future Directions. Deep learning algorithms have “Global prevalence of diabetic retinopathy: protocol for
gained immense momentum in improving diagnosis in a systematic review and meta-analysis,” BMJ Open, vol. 9,
medical imaging systems. Te traditional studies are con- no. 3, Article ID 22188, 2019.
centrated on CNN models and deep-layered architectures to [5] M. R. K. Mookiah, U. R. Acharya, C. K. Chua, C. M. Lim,
detect diabetic retinopathy. Furthermore, it specifes the use E. Ng, and A. Laude, “Computer-aided diagnosis of diabetic
of the massive amount of data presented on image datasets, retinopathy: a review,” Computers in Biology and Medicine,
which was combined with retinal lesions to design a robust vol. 43, no. 12, pp. 2136–2155, 2013.
[6] M. D. Abramof, Y. Lou, A. Erginay et al., “Improved auto-
model for efective implementation. Due to the limited
mated detection of diabetic retinopathy on a publicly available
number of publicly available datasets, DR diagnosis remains dataset through integration of deep learning,” Investigative
a challenge. Te recent advancements in DL have produced Opthalmology & Visual Science, vol. 57, no. 13, pp. 5200–5206,
promising classifcation results despite having difculty 2016.
identifying the damaged lesions. For example, the study in [7] J. Sahlsten, J. Jaskari, J. Kivinen et al., “Deep learning fundus
[90, 91] identifed segmented intraretinal variations. Images image analysis for diabetic retinopathy and macular edema
were classifed according to their severity level using the grading,” Scientifc Reports, vol. 9, no. 1, Article ID 10750,
multiclass classifcation approach. Tis system performed 2019.
well on various fundus image datasets and identifed the [8] S. Seth and B. Agarwal, “A hybrid deep learning model for
patient’s condition. Tese techniques have promising po- detecting diabetic retinopathy,” Journal of Statistics &
tential in the detection of diferent eye disorders, namely, Management Systems, vol. 21, no. 4, pp. 569–574, 2018.
[9] A. Tayal, J. Gupta, A. Solanki, K. Bisht, A. Nayyar, and
glaucoma and other intraretinal abnormalities.
M. Masud, “Dl cnn-based approach with image processing
Another future enhancement is identifed DR at very techniques for the diagnosis of retinal diseases,” Multimedia
large image datasets such as Messidor and EyePACS. To Systems, pp. 1–22, 2021.
classify the severity level, this system completely relies on the [10] Q. Abbas, I. Fondon, A. Sarmiento, S. Jime´nez, and
quality of the images that are presented easily. Tose images P. Alemany, “Automatic recognition of severity level for
are retrained using the transfer learning method to detect the diagnosis of diabetic retinopathy using deep visual features,”
16 Journal of Healthcare Engineering

Medical, & Biological Engineering & Computing, vol. 55, capsule networks,” Com- plex & Intelligent Systems, vol. 7,
no. 11, pp. 1959–1974, 2017. pp. 1–14, 2021.
[11] K. Oh, H. M. Kang, D. Leem, H. Lee, K. Y. Seo, and S. Yoon, [26] S. Dutta, B. C. Manideep, S. M. Basha, R. D. Caytiles, and
“Early detection of diabetic retinopathy based on deep N. C. S. N. Iyengar, “Classifcation of diabetic retinopathy
learning and ultra-widefeld fundus images,” Scientifc Re- images by using deep learning models,” International Journal
ports, vol. 11, no. 1, pp. 1897–1899, 2021. of Grid and Distributed Computing, vol. 11, no. 1, pp. 99–106,
[12] M. Rehman, S. H. Khan, Z. Abbas, and S. D. Rizvi, “Classi- 2018.
fcation of diabetic retinopathy images based on customised [27] G. T. Zago, R. V. Andrea􏽥o, B. Dorizzi, and E. O. Teatini Salles,
cnn architecture,” in Proceedings of the 2019 Amity In- “Diabetic retinopathy detection using red lesion localization
ternational Conference on Artifcial Intelligence (AICAI), and convolutional neural networks,” Computers in Biology
pp. 244–248, IEEE, Dubai, UAE, February 2019. and Medicine, vol. 116, Article ID 103537, 2020.
[13] B. Tymchenko, P. Marchenko, and D. Spodarets, “Deep [28] S.-I. Pao, H.-Z. Lin, K.-H. Chien, M.-C. Tai, J.-T. Chen, and
learning approach to diabetic retinopathy detection,” 2020, G.-M. Lin, “Detection of diabetic retinopathy using bichannel
https://arxiv.org/abs/2003.02261. convolutional neural network,” Journal of Ophthalmology,
[14] L. Qiao, Y. Zhu, and H. Zhou, “Diabetic retinopathy detection vol. 2020, Article ID 9139713, 7 pages, 2020.
using prognosis of microaneurysm and early diagnosis system [29] S. Honnungar, S. Mehra, and S. Joseph, “Diabetic retinopathy
for non- proliferative diabetic retinopathy based on deep identifcation and severity classifcation,” Fall, vol. 2016, 2016.
learning algorithms,” IEEE Access, vol. 8, Article ID 104292, [30] R. Gargeya and T. Leng, “Automated identifcation of diabetic
2020. retinopathy using deep learning,” Ophthalmology, vol. 124,
[15] A. Bora, S. Balasubramanian, B. Babenko et al., “Predicting the no. 7, pp. 962–969, 2017.
risk of developing diabetic retinopathy using deep learning,” [31] T. Li, Y. Gao, K. Wang, S. Guo, H. Liu, and H. Kang, “Di-
Te Lancet Digital Health, vol. 3, no. 1, pp. e10–e19, 2021. agnostic assessment of deep learning algorithms for diabetic
[16] M. Imran, M. Imran, A. Ullah, M. Arif, and R. Noor, “A retinopathy screening,” Information Sciences, vol. 501,
unifed technique for entropy enhancement based diabetic pp. 511–522, 2019.
retinopathy detection using hybrid neural network,” Com- [32] J. de La Torre, A. Valls, and D. Puig, “A deep learning in-
puters in Biology and Medicine, vol. 145, Article ID 105424, terpretable classifer for diabetic retinopathy disease grading,”
2022. Neurocomputing, vol. 396, pp. 465–476, 2020.
[17] V. Purna Chandra Reddy and K. K. Gurrala, “OHGCNet: [33] T. A. Soomro, J. Gao, T. Khan, A. F. M. Hani, M. A. U. Khan,
optimal feature selection-based hybrid graph convolutional and M. Paul, “Computerised approaches for the detection of
network model for joint DR-DME classifcation,” Biomedical diabetic retinopathy using retinal fundus images: a survey,”
Signal Processing and Control, vol. 78, Article ID 103952, 2022. Pattern Analysis & Applications, vol. 20, no. 4, pp. 927–961,
[18] L. Guariguata, D. R. Whiting, I. Hambleton, J. Beagley, 2017.
U. Linnenkamp, and J. E. Shaw, “Global estimates of diabetes [34] V. Gulshan, L. Peng, M. Coram et al., “Development and
prevalence for 2013 and projections for 2035,” Diabetes Re- validation of a deep learning algorithm for detection of di-
search and Clinical Practice, vol. 103, no. 2, pp. 137–149, 2014. abetic retinopathy in retinal fundus photographs,” JAMA,
[19] Gbd 2019 Blindness and Vision Impairment Collaborators, vol. 316, no. 22, pp. 2402–2410, 2016.
Vision Loss Expert Group of the Global Burden of Disease [35] Z. Gao, J. Li, J. Guo, Y. Chen, Z. Yi, and J. Zhong, “Diagnosis
Study, P. S. Briant et al., “Causes of blindness and vision of diabetic retinopathy using deep neural networks,” IEEE
impairment in 2020 and trends over 30 years, and prevalence Access, vol. 7, pp. 3360–3370, 2019.
of avoidable blindness in relation to vision 2020: the right to [36] M. P. Rathi and D. S. Padmavati Shrivastava, “Prediction of
sight: an analysis for the global burden of disease study,” diabetic retinopathy using classifcation techniques,” Solid
Lancet Global Health, vol. 9, no. 2, pp. e144–e160, 2021. State Technology, vol. 63, pp. 9479–9494, 2020.
[20] K. Shankar, Y. Zhang, Y. Liu, L. Wu, and C.-H. Chen, [37] P. Bhatter, E. Frisch, E. Duhaime, A. Jain, and C. Fischetti,
“Hyperparameter tuning deep learning for diabetic retinop- “Diabetic retinopathy detection using collective intelligence,”
athy fundus image classifcation,” IEEE Access, vol. 8, Article Journal of Scientifc Innovation in Medicine, vol. 3, no. 1, 2020.
ID 118164, 2020. [38] D. Lam, M. G. Yi, and T. Lindsey, “Automated detection of
[21] D. J. Hemanth, O. Deperlioglu, and U. Kose, “An enhanced diabetic retinopathy using deep learning,” AMIA summits on
diabetic retinopathy detection and classifcation approach translational science proceedings, vol. 2018, p. 147, 2018.
using deep convolutional neural network,” Neural Computing [39] A. J. Reddy, A. Dang, A. A. Dao, G. Arakji, J. Cherian, and
& Applications, vol. 32, no. 3, pp. 707–721, 2020. H. Brahmbhatt, “A substantive narrative review on the usage
[22] T. R. Gadekallu, N. Khare, S. Bhattacharya et al., “Early de- of lidocaine in cataract surgery,” Cureus, vol. 13, no. 10, Article
tection of diabetic retinopathy using pca-frefy based deep ID 19138, 2021.
learning model,” Electronics, vol. 9, no. 2, p. 274, 2020. [40] P. R. Kshirsagar, H. Manoharan, S. Selvarajan, H. A. Alterazi,
[23] X. Zeng, H. Chen, Y. Luo, and W. Ye, “Automated diabetic D. Singh, and H. N. Lee, “Perception exploration on ro-
retinopathy detection based on binocular siamese-like con- bustness syndromes with pre-processing entities using ma-
volutional neural network,” IEEE Access, vol. 7, Article ID chine learning algorithm,” Frontiers in Public Health, vol. 10,
30744, 2019. Article ID 893989, 2022.
[24] S. Wan, Y. Liang, and Y. Zhang, “Deep convolutional neural [41] H. Manoharan, S. Selvarajan, A. Yafoz et al., “Deep conviction
networks for diabetic retinopathy detection by image classi- systems for biomedical applications using intuiting pro-
fcation,” Computers & Electrical Engineering, vol. 72, cedures with cross point approach,” Frontiers in Public
pp. 274–282, 2018. Health, vol. 10, Article ID 909628, 2022.
[25] G. Kalyani, B. Janakiramaiah, A. Karuna, and L. V. N. Prasad, [42] P. R. Kshirsagar, H. Manoharan, S. Shitharth, A. M. Alshareef,
“Diabetic retinopathy detection and classifcation using N. Albishry, and P. K. Balachandran, “Deep learning
Journal of Healthcare Engineering 17

approaches for prognosis of automated skin disease,” Te Life, [58] A. Mohanarathinam, C. S. Manikandababu, N. B. Prakash,
vol. 12, no. 3, p. 426, 2022. G. R. Hemalakshmi, and K. Subramaniam, “Diabetic reti-
[43] Y. Niu, L. Gu, Y. Zhao, and F. Lu, “Explainable diabetic nopathy detection and classifcation using hybrid multiclass
retinopathy detection and retinal image generation,” IEEE SVM classifer and deep learning techniques,” Mathematical
Journal of Biomedical and Health Informatics, vol. 26, no. 1, Statistician and Engineering Applications, vol. 71, no. 3,
pp. 44–55, 2022. pp. 891–903, 2022.
[44] A. Rakhlin, “Diabetic retinopathy detection through in- [59] M. Elsharkawy, A. Sharafeldeen, A. Soliman et al., “A novel
tegration of deep learning classifcation framework,” 2018, computer-aided diagnostic system for early detection of di-
https://www.biorxiv.org/content/10.1101/225508v2. abetic retinopathy using 3D-OCT higher-order spatial ap-
[45] H. Pratt, F. Coenen, D. M. Broadbent, S. P. Harding, and pearance model,” Diagnostics, vol. 12, no. 2, p. 461, 2022.
Y. Zheng, “Convolutional neural networks for diabetic reti- [60] J. Kaur and D. Mittal, “A generalized method for the detection
nopathy,” Procedia Computer Science, vol. 90, pp. 200–205, of vascular structure in pathological retinal images,” Bio-
2016. cybernetics and Biomedical Engineering, vol. 37, no. 1,
[46] V. Deepa, C. S. Kumar, and T. Cherian, “Ensemble of multi- pp. 184–200, 2017.
stage deep convolutional neural networks for automated [61] V. Carrera, A. Gonza´lez, and R. Carrera, “Automated de-
grading of diabetic retinopathy using image patches,” Journal tection of diabetic retinopathy using svm,” in Proceedings of
of King Saud University- Computer and Information Sciences, the 2017 IEEE XXIV international conference on electronics,
vol. 34, no. 8, pp. 6255–6265, 2022. electrical engineering and computing (IN- TERCON), pp. 1–4,
[47] R. Sayres, A. Taly, E. Rahimy et al., “Using a deep learning IEEE, Cusco, Peru, August 2017.
algorithm and integrated gradients explanation to assist [62] M. H. Mahmoud, S. Alamery, H. Fouad, A. Altinawi, and
grading for diabetic retinopathy,” Ophthalmology, vol. 126, A. E. Youssef, “An automatic detection system of diabetic
no. 4, pp. 552–564, 2019. retinopathy using a hybrid inductive machine learning al-
[48] I. Qureshi, J. Ma, and Q. Abbas, “Recent development on gorithm,” Personal and Ubiquitous Computing, vol. 25,
detection methods for the diagnosis of diabetic retinopathy,” pp. 1–15, 2021.
Symmetry, vol. 11, no. 6, p. 749, 2019. [63] M. Mateen, J. Wen, M. Hassan, N. Nasrullah, S. Sun, and
[49] C. Mahiba and A. Jayachandran, “Severity analysis of diabetic S. Hayat, “Automatic detection of diabetic retinopathy: a re-
retinopathy in retinal images using hybrid structure de- view on datasets, methods and evaluation metrics,” IEEE
scriptor and modifed cnns,” Measurement, vol. 135, Access, vol. 8, Article ID 48784, 2020.
pp. 762–767, 2019. [64] F. Li, Z. Liu, H. Chen, M. Jiang, X. Zhang, and Z. Wu,
[50] C. Valverde, M. Garcia, R. Hornero, and M. Lopez-Galvez, “Automatic detection of diabetic retinopathy in retinal fundus
“Automated detection of diabetic retinopathy in retinal im- photographs based on deep learning algorithm,” Translational
ages,” Indian Journal of Ophthalmology, vol. 64, no. 1, p. 26, vision science & technology, vol. 8, no. 6, p. 4, 2019.
2016. [65] A. Colomer, J. Igual, and V. Naranjo, “Detection of early signs
[51] M. Bhaskaranand, C. Ramachandra, S. Bhat et al., “Automated of diabetic retinopathy based on textural and morphological
diabetic retinopathy screening and monitoring using retinal information in fundus images,” Sensors, vol. 20, no. 4, p. 1005,
fundus image analysis,” Journal of diabetes science and 2020.
technology, vol. 10, no. 2, pp. 254–261, 2016. [66] S. Sil Kar and S. P. Maity, “Gradation of diabetic retinopathy
[52] R. Sarki, K. Ahmed, H. Wang, and Y. Zhang, “Automatic on reconstructed image using compressed sensing,” IET
detection of diabetic eye disease through deep learning using Image Processing, vol. 12, no. 11, pp. 1956–1963, 2018.
fundus images: a survey,” IEEE Access, vol. 8, Article ID [67] S. Karkuzhali and D. Manimegalai, “Distinguising proof of
151133, 2020. diabetic retinopathy detection by hybrid approaches in two-
[53] Z. Mohammadpoory, M. Nasrolahzadeh, N. Mahmoodian, dimensional retinal fundus images,” Journal of Medical Sys-
and J. Had- dadnia, “Automatic identifcation of diabetic tems, vol. 43, no. 6, pp. 173–212, 2019.
retinopathy stages by using fundus images and visibility graph [68] N. Sikder, M. S. Chowdhury, A. S. M. Arif, and A.-A. Nahid,
method,” Measurement, vol. 140, pp. 133–141, 2019. “Early blindness detection based on retinal images using
[54] X. Li, T. Pang, B. Xiong, W. Liu, P. Liang, and T. Wang, ensemble learning,” in Proceedings of the 2019 22nd In-
“Convolutional neural networks-based transfer learning for ternational Conference on Computer and Information Tech-
diabetic retinopathy fundus image classifcation,” in Pro- nology (ICCIT), pp. 1–6, IEEE, Dhaka, Bangladesh, December
ceedings of the 2017 10th international congress on image and 2019.
signal processing, biomedical engineering and informatics [69] M. Mateen, J. Wen, N. Nasrullah, S. Sun, and S. Hayat,
(CISP- BMEI), pp. 1–11, IEEE, Shanghai, China, October “Exudate detection for diabetic retinopathy using pretrained
2017. convolutional neural networks,” Complexity, vol. 2020, Article
[55] L. Zhou, Y. Zhao, J. Yang, Q. Yu, and X. Xu, “Deep multiple ID 5801870, 11 pages, 2020.
instance learning for automatic detection of diabetic reti- [70] Z. Kong, T. Li, J. Luo, and S. Xu, “Automatic tissue image
nopathy in retinal im- ages,” IET Image Processing, vol. 12, segmentation based on image processing and deep learning,”
no. 4, pp. 563–571, 2018. Journal of healthcare engineering, vol. 2019, Article ID
[56] Darvish and M. Ezoji, “Morphological exudate detection in 2912458, 10 pages, 2019.
retinal images using pca-based optic disc removal,” Journal of [71] K. S Archana and A. Sahayadhas, “Automatic rice leaf disease
AI and Data Mining, vol. 7, no. 4, pp. 487–493, 2019. segmentation using image processing techniques,” In-
[57] G. Li, S. Zheng, and X. Li, “Exudate detection in fundus ternational Journal of Engineering & Technology, vol. 7, no. 27,
images via convolutional neural network,” in International pp. 182–185, 2018.
Forum on Digital TV and Wireless Multimedia Communi- [72] R. Sahoo and C. Sekhar, “Detection of diabetic retinopathy
cations, pp. 193–202, Springer, New York, NY, USA, 2017. from retinal fundus image using wavelet-based image
18 Journal of Healthcare Engineering

segmentation,” International Journal of Computer Applica- europe—narrative review,” Healthcare, vol. 9, no. 6, p. 745,
tion, vol. 182, no. 47, pp. 46–50, 2019. 2021.
[73] C. F. Baumgartner, L. M. Koch, M. Pollefeys, and [87] S. M. Ismail, L. A. Said, A. H. Madian, and A. G. Radwan,
E. Konukoglu, “An exploration of 2d and 3d deep learning “Fractional- order edge detection masks for diabetic reti-
techniques for cardiac mr image segmentation,” in In- nopathy diagnosis as a case study,” Computers, vol. 10, no. 3,
ternational Workshop on Statistical Atlases and Computa- p. 30, 2021.
tional Models of the Heart, pp. 111–119, Springer, New York, [88] K. Seid, T. Tesfaye, A. Belay, and H. Mohammed, “De-
NY, USA, 2017. terminants of diabetic retinopathy in tikur anbessa hospital,
[74] S. Ghosh, N. Das, I. Das, and U. Maulik, “Understanding deep Ethiopia: a case–control study,” Clinical Diabetes and En-
learning techniques for image segmentation,” ACM Com- docrinology, vol. 7, no. 1, pp. 12–19, 2021.
puting Surveys, vol. 52, no. 4, pp. 1–35, 2019. [89] M. Tilahun, T. Gobena, D. Dereje, M. Welde, and G. Yideg,
[75] N. Iriawan, A. A. Pravitasari, K. Fithriasari, S. W. Purnami, “Prevalence of diabetic retinopathy and its associated factors
and W. Ferriastuti, “Comparative study of brain tumor among diabetic patients at Debre Markos referral hospital,
segmentation using diferent segmentation techniques in Northwest Ethiopia, 2019: hospital-basedcross-sectional
handling noise,” in Proceedings of the 2018 Interna- tional study,” Diabetes, Metabolic Syndrome and Obesity: Targets
Conference on Computer Engineering, Network and Intelligent and Terapy, vol. 13, pp. 2179–2187, 2020.
Multimedia (CENIM), pp. 289–293, IEEE, Surabaya, Indo- [90] A. Sungheetha and R. Sharma R, “Design an early detection
nesia, November 2018. and classifca- tion for diabetic retinopathy by deep feature
[76] S. K. Saha, D. Xiao, A. Bhuiyan, T. Y. Wong, and extraction-based convolution neural network,” Journal of
Y. Kanagasingam, “Color fundus image registration tech- Trends in Computer Science and Smart technology, vol. 3, no. 2,
niques and applications for automated analysis of diabetic pp. 81–94, 2021.
retinopathy progression: a review,” Biomedical Signal Pro- [91] S. Majumder and N. Kehtarnavaz, “Multitasking deep
cessing and Control, vol. 47, pp. 288–302, 2019. learning model for detection of fve stages of diabetic reti-
[77] E. Golkar, H. Rabbani, and A. Dehghani, “Hybrid registration nopathy,” IEEE Access, vol. 9, Article ID 123220, 2021.
of retinal fuorescein angiography and optical coherence to- [92] A. Erciyas and N. Barıs¸c¸ı, “An efective method for detecting
mography images of patients with diabetic retinopathy,” and classifying diabetic retinopathy lesions based on deep
Biomedical Optics Express, vol. 12, no. 3, pp. 1707–1724, 2021. learning,” Computational and Mathematical Methods in
[78] X. Sui, Y. Zheng, Y. Jiang, W. Jiao, and Y. Ding, “Deep Medicine, vol. 2021, Article ID 9928899, 13 pages, 2021.
multispectral image registration network,” Computerized [93] G. U. Nneji, J. Cai, J. Deng, H. N. Monday, M. A. Hossin, and
Medical Imaging and Graphics, vol. 87, Article ID 101815, S. Nahar, “Identifcation of diabetic retinopathy using
2021. weighted fusion deep learning based on dual-channel fundus
[79] M. M. Khansari, J. Zhang, Y. Qiao et al., “Automated scans,” Diagnostics, vol. 12, no. 2, p. 540, 2022.
deformation-based analysis of 3d op- tical coherence to-
mography in diabetic retinopathy,” IEEE Transactions on
Medical Imaging, vol. 39, no. 1, pp. 236–245, 2020.
[80] R. Ramli, M. Y. I. Idris, K. Hasikin et al., “Local descriptor for
retinal fundus image registration,” IET Computer Vision,
vol. 14, no. 4, pp. 144–153, 2020.
[81] C. Hernandez-Matas, X. Zabulis, and A. A. Argyros, “An
experimental evaluation of the accuracy of key points-based
retinal image registration,” in Proceedings of the 2017 39th
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), pp. 377–381, IEEE, Jeju
Island, Republic of Korea, July 2017.
[82] A. Noriega, D. Meizner, D. Camacho et al., “Screening di-
abetic retinopathy using an automated retinal image analysis
system in Mexico: independent and assistive use cases mexico:
randomized controlled trial,” JMIR Formative Research, vol. 5,
2020.
[83] M. E. Gegundez-Arias, D. Marin, B. Ponte et al., “A tool for
automated diabetic retinopathy pre-screening based on ret-
inal image computer analysis,” Computers in Biology and
Medicine, vol. 88, pp. 100–109, 2017.
[84] L. Yao, X. Wang, Y. Zhong et al., “Te triglyceride–glucose
index is associated with diabetic retinopathy in Chinese pa-
tients with type 2 diabetes: a hospital- based, nested, case-
–control study,” Diabetes, Metabolic Syndrome and Obesity:
Targets and Terapy, vol. 14, pp. 1547–1555, 2021.
[85] N. Silberman, K. Ahrlich, R. Fergus, and L. Subramanian,
“Case for automated detection of diabetic retinopathy,” in
Proceedings of the 2010 AAAI spring symposium series, Palo
Alto, CA, USA, March 2010.
[86] E. Hristova, D. Koseva, Z. Zlatarova, and K. Dokova, “Di-
abetic retinopathy screening and registration in