Content uploaded by Gbenga Gbotoso
Author content
All content in this area was uploaded by Gbenga Gbotoso on Oct 13, 2023
Content may be subject to copyright.
Healthcare Analytics 4 (2023) 100261
Available online 22 September 2023
2772-4425/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A systematic review of retinal fundus image segmentation and classication
methods using convolutional neural networks
Ademola E. Ilesanmi
a
,
*
, Taiwo Ilesanmi
b
, Gbenga A. Gbotoso
c
a
University of Pennsylvania, 3710 Hamilton Philadelphia, PA, 19104, United States
b
National Population Commission, Abuja, Nigeria
c
Lagos State University of Science and Technology, Ikorodu, Nigeria
ARTICLE INFO
Handling Editor: Madijd Tavana
Keywords:
Convolutional neural network
Retinal fundus
Image segmentation and classication
Eye-related disorders
Retinal disease
Computer-aided image detection
ABSTRACT
Retinal fundus images play a crucial role in the early detection of eye problems, aiding in timely diagnosis and
treatment to prevent vision loss or blindness. With advancements in technology, Convolutional Neural Network
(CNN) algorithms have emerged as effective tools for recognition, delineation, and classication tasks. This study
proposes a comprehensive review of CNN algorithms used for retinal fundus image segmentation and classi-
cation. Our review follows a systematic approach, exploring diverse repositories to identify studies employing
CNN to segment and classify retinal fundus images. Utilizing CNNs in the segmentation and classication of
retinal fundus images can enhance the precision of segmentation outcomes and alleviate the sole dependence on
human experts. This approach enables more accurate segmentation results, reducing the burden on human ex-
perts. A total of sixty-two studies are included in our review, analyzing aspects such as database usage and the
advantages and disadvantages of the methods employed. The review provides valuable insights, limitations,
observations, and future directions in the eld. Despite certain limitations, the ndings indicate that CNN al-
gorithms consistently achieve high accuracies. The comprehensive examination of the included studies sheds
light on the potential of CNN in retinal fundus image analysis.
1. Introduction
The prevalence of eye-related diseases has witnessed a signicant
global increase, leading to a rise in the number of individuals experi-
encing acute conditions. According to the World Health Organization
(WHO), approximately 2.2 billion people worldwide suffer from visual
impairment. Alarmingly, around 1 billion cases could have been pre-
vented if timely detection had occurred. In the United States alone, an
estimated 40 million people are affected by varying degrees of severe
eye-related diseases, with a primary focus on conditions related to the
retina, such as glaucoma, among others [1,2]. In Africa, visual impair-
ment affects approximately 26.3 million individuals, with 20.4 million
experiencing low vision and 5.9 million being blind, contributing to
15.3% of the global blind population. The leading causes of blindness
and visual impairment in this region are uncorrected refractive errors
and cataracts. While individuals over the age of 50 are more susceptible
to visual impairment and blindness, it is important to note that these
conditions can affect people of all age groups. The International Clas-
sication of Diseases 11 (2018) categorizes visual impairment into two
main groups: 1) Distance vision impairment, ranging from mild to
blindness, and 2) Near vision impairment, characterized by acuity worse
than N6 or M.08 at 40 cm.
The degree of visual impairment is inuenced by several factors,
including the availability of preventive measures and treatment options,
access to vision rehabilitation services, and the availability of basic
amenities such as inclusive buildings, transportation, and information.
The leading causes of vision impairment encompass a range of condi-
tions, including: Cataracts, Glaucoma, Diabetic retinopathy, Corneal
opacity, Trachoma, Uncorrected refractive error, Age-related macular
degeneration. The causes of visual impairment among children can vary
depending on the country and social status. In low-income countries,
cataracts often emerge as the primary cause, whereas premature reti-
nopathy is a prominent cause in middle-income countries. Uncorrected
refractive error remains a leading cause of visual impairment in children
across all countries. In the adult population, glaucoma emerges as a
signicant concern, affecting nearly 80 million people globally and
ranking among the leading causes of blindness. Glaucoma can be clas-
sied into two main types: open-angle and closed-angle glaucoma.
* Corresponding author.
E-mail address: Ademola.Ilesanmi@Pennmedicine.upenn.edu (A.E. Ilesanmi).
Contents lists available at ScienceDirect
Healthcare Analytics
journal homepage: www.elsevier.com/locate/health
https://doi.org/10.1016/j.health.2023.100261
Received 10 May 2023; Received in revised form 6 August 2023; Accepted 12 September 2023
Healthcare Analytics 4 (2023) 100261
2
Clinicians estimate that approximately 90% of affected individuals suf-
fer from open-angle glaucoma. Diagnosis of glaucoma involves pro-
cedures such as assessing neuroretinal rim loss, conducting visual eld
tests, and evaluating nerve ber characteristics [3].
The optic cup, resembling a white cup-like structure at the center of
the eye, is positioned within the optic disk. It serves as one of the
diagnostic indicators for glaucoma. Diabetic retinopathy (DR) is another
prevalent cause of visual impairment, characterized by damage to the
tissues located in the blood vessels at the back of the eye. The WHO
identies DR as a signicant contributor to blindness, ranking as the
fourth leading cause globally [4] (refer to Figs. 2 and 3). In the early
stages of DR, there are usually no noticeable symptoms or changes in
eyesight. However, if left untreated, it can result in permanent vision
impairment. A study conducted by the Center for Disease Control and
Prevention in the United States revealed that almost one-third of adults
over the age of 40 have DR. Furthermore, over one-third of
African-Americans and Mexican-Americans have diabetes, a risk factor
for developing DR [5].
Retinal fundus images are photographs that provide a direct optical
representation of the eye’s internal processes. They capture various
morphological and pathological components, including blood vessels,
the macula, fovea, optic disk, hemorrhages, arterioles, venules, exu-
dates, and microaneurysms [2]. Fig. 1 illustrates a diagram of retinal
fundus images, highlighting different signs of DR, such as exudates,
microaneurysms, hemorrhages, and neovascularization.
To prevent visual loss and maintain healthy eyesight, it is advisable
to follow the recommendations outlined below: 1) Maintain stable blood
sugar levels to minimize the risk of diabetic complications that can affect
vision. 2) Be aware of your family’s eye health history as certain eye
conditions can have a hereditary component. 3) Adopt a balanced and
nutritious diet that includes eye-healthy foods to support overall eye
health. 4) Maintain a healthy weight, as obesity and excessive weight
can increase the risk of various eye diseases. 5) Wear appropriate pro-
tective eyewear, such as safety goggles or sunglasses, to shield your eyes
from potential injuries and harmful UV rays. 6) Incorporate regular
exercise into your routine, as it promotes overall well-being and good
blood circulation, which is benecial for eye health. 7) Reduce or
eliminate tobacco usage, as smoking has been linked to several eye
diseases and can exacerbate existing visual impairments. 8) Practice
proper eye hygiene, including resting your eyes periodically and
ensuring clean hands and eye lenses to reduce the risk of infections and
other eye-related issues [7–9].
Early detection and diagnosis are crucial in safeguarding against
visual impairment and blindness. For individuals with type 1 diabetes, it
is recommended to undergo screening for diabetic retinopathy (DR)
three to ve years after the onset of diabetes, while those with type 2
diabetes should have their rst screening within one year of diagnosis
[10]. Subsequent examinations should be scheduled every six months or
one year, depending on the severity of DR. In cases where the condition
is more severe, early and more frequent examinations may be advised.
The diagnosis of visual impairment involves a comprehensive exami-
nation, including high-quality retinal photography, and a structured
follow-up process. Various techniques, such as direct and indirect
ophthalmoscopy, stereoscopic retinal fundus photography, and mydri-
atic and nonmydriatic photographs, are utilized to detect and classify
visual impairments.
The gold standard for assessing retinal fundus images is the utiliza-
tion of stereoscopic photographs in seven standard elds, covering 30◦
[11–13]. While this approach is accurate and effective, it requires skilled
professionals to operate, and the process can be laborious and
time-consuming. In many countries, clinicians face a heavy workload
due to low doctor-to-patient ratios, leading to potential errors and delays
in diagnosis and treatment. As a result, numerous medical photographs
taken in hospitals and clinics remain unused and stored within the fa-
cilities. Fortunately, these photographs can be valuable resources for
Computer-Aided Detection systems (CADs) [14]. CAD systems assist
doctors in interpreting medical images by processing digital images and
identifying signicant areas that may indicate potential disease spots.
Convolutional neural networks (CNNs) have been developed to
automatically detect and diagnose ophthalmic issues in retinal fundus
images. A notable advantage of CNNs is their ability to learn complex
features automatically and translate them into meaningful results. By
reducing manual procedures, CNN systems can enhance healthcare
practices and serve as a valuable second interpreter. They have the
potential to automate eye screening, detect abnormalities, and provide
exibility for healthcare practitioners. The objective of this study is to
present the ndings of various CNN algorithms used for automatic
segmentation, classication, and disease detection in retinal fundus
images. This review focuses on recent trends from 2015 onwards,
analyzing the performance of different CNN methods, databases, and
validation metrics. It also explores the roadmap and challenges associ-
ated with CNN detection and diagnosis. Previous research, such as
reference [2], has examined deep learning for retinal fundus images, but
this review aims to investigate the latest trends, including the pathology
and morphologies of retinal fundus images using CNN methods. Addi-
tionally, it will address various validation metrics, limitations, and
future directions for the segmentation and classication of retinal
fundus images. The contributions of this review can be summarized as
follows.
1. Analysis of different CNN methods for the segmentation and classi-
cation of retinal fundus images, along with an examination of
various databases and validation metrics.
2. Discussion of potential challenges and roadmap for CNN utilization
in the segmentation and classication of retinal fundus images.
The remainder of the paper is organized as follows: Section 2 pro-
vides background information and reviews methods for the
Fig. 1. Retinal fundus images showing different signs of the DR [6].
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
3
classication and segmentation of fundus images using CNN. Section 3
discusses different evaluation metrics employed in the classication and
segmentation processes. Databases are examined in Section 4, while
Section 5 presents statistical analyses of different algorithms. Finally,
Section 6 concludes the paper.
2. Background
In the 1860s, doctors began exploring potential solutions for
capturing images of the eye. By the 1880s, a partial solution had been
developed, allowing doctors to take pictures by placing a camera on the
patient’s head. However, they still had to wait approximately 3 min for
the lm to develop. Despite its simplicity, this development marked the
rst time that images of the eye could be captured by anyone. In 1926,
the rst fundus camera was invented, enabling photography of a portion
of the eye. It took another 70 years for the development of a retinal
fundus camera capable of capturing images spanning 130◦. Eventually,
in the 21st century, a non-invasive camera was introduced, capable of
capturing a 200-degree view [15].
In the realm of computers, the rst algorithm executed on a machine
was created by Ada Lovelace in 1843. Since then, numerous algorithms
have been developed to perform specic tasks. The eld of computer
vision emerged in 1966 when computers were employed to identify
objects. Mathematical analysis and quantitative applications were
introduced to computer vision in the subsequent decade, incorporating
techniques such as scale-space representation [16], contour models
[17], and Markov random elds [18]. Marvin Minsky’s research in
mimicking the human brain paved the way for computers to process
Fig. 2. Cases of vision loss by Countries [1].
Fig. 3. Global blindness gure [1].
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
4
information for decision-making. In 1959, Russell Kirsch invented a
digital image scanner capable of transforming images into digital data.
Lawrence Robert processed 3D information about solid objects from 2D
photographs in 1963. Kunishiko Fukushima developed the precursor of
modern CNNs in 1980. In 1999, David Lowe described a visual recog-
nition system utilizing local invariant features. In 2001, the rst
real-time face detection framework was introduced.
The breakthrough moment in computer vision occurred in 2012
when AlexNet won the ImageNet competition. Since then, numerous
researchers have utilized CNN methods for segmenting and classifying
medical images, particularly retinal fundus images. In recent years, the
U-shaped Network has emerged as a prominent and effective approach
for medical and biomedical image segmentation. Notably, the method
introduced by Ronneberger et al. [19] has gained signicant recogni-
tion. Building upon the U-Shaped architecture, several modications
and variants have been proposed, aiming to further enhance its perfor-
mance. A comprehensive discussion and exploration of these different
U-Shaped variants can be found in Ref. [20].
3. Review of methods
Numerous CNN methods have been proposed by researchers for
segmenting and classifying retinal fundus images, employing diverse
network architectures to develop sophisticated AI platforms. For
instance, Rohit Thanki [21] proposed a CNN method specically
designed to detect glaucoma in fundus images. This approach utilizes a
deep neural network for feature extraction and incorporates six machine
learning methods for classication. Promising results were obtained,
with the logistic regression algorithm achieving the highest classier
accuracy of 0.99 (99%). The method was evaluated on a dataset con-
sisting of 15 normal images and 15 images affected by glaucoma. In
another study, Phridviraj et al. [22] proposed a bi-directional Long
Short-Term Memory (LSTM) approach for the detection of DR in fundus
images. This method employs a three-fold approach. Firstly, a pre-
processing technique is applied to enhance the image quality. Secondly,
a deep learning-based efcient network is utilized to extract relevant
features. Finally, a bi-directional long-term memory classier
comprising six LSTM layers is employed to classify the images. The
proposed method achieved an accuracy of 97% when tested on three
different datasets.
In the domain of glaucoma classication in fundus images, Kamesh
Sonti and Ravindra Dhul [23] proposed a CNN-based approach. Their
method consists of 26 layers, including six convolution layers, four
pooling layers, and one fully connected layer. A softmax layer is utilized
to generate a simple mask prediction for the classier. Through
cross-validation and data augmentation techniques, the method ach-
ieved a high-performance accuracy of 96%. In a related study, Raja
Sankari et al. [24] focused on detecting retinopathy in fundus images of
preterm infants using CNN. Their approach involved preprocessing and
segmentation of the images using a modied multiresolution U-Shaped
Network [25]. Features were extracted from both traditional methods
and deep learning techniques. Subsequently, these features were fed into
the random forest algorithm for classication. These research efforts
demonstrate the application of CNN in the accurate identication and
classication of specic eye conditions, such as glaucoma and retinop-
athy. The proposed methods showcase the potential of CNN-based ap-
proaches in improving diagnosis and treatment decisions based on
retinal fundus images.
In this study we emphasize the prowess of CNN methodologies in
tackling distinct ophthalmic conditions. Our ndings underscore the
substantial potential of these techniques for precise diagnosis and
effective classication of retinal fundus images. To provide readers with
an understanding of the current trends, we have organized these CNN
methods into three categories: (1) CNN methods for classifying and
segmenting optic discs, (2) CNN methods for classifying and segmenting
arteries and veins, and (3) CNN methods for classifying and segmenting
retinal blood vessels. Each of these categories comprises a substantial
number of CNN methods that have demonstrated promising results.
Fig. 4 presents a block diagram illustrating the different CNN ap-
proaches, along with a list of prominent CNN methods used within each
category.
3.1. CNN used for segmentation and classication
3.1.1. CNN used for optic disk and optic cup
The automation of optic disk and optic cup segmentation can address
challenges encountered in the manual procedure and those anticipated
in the future. However, this segmentation technique faces various
challenges, including 1) unclear boundaries, 2) signicant variability, 3)
interference from other image components, and 4) mixed pathologies.
To overcome these challenges, researchers have proposed different
convolutional neural network (CNN) methods. Wang et al. [26] intro-
duce an encoder-decoder network that consists of two components
working in tandem. The rst component is the feature detection (FDS)
module, which preserves features by employing two stacked convolu-
tional layers (3 ×3) with batch normalization (BN) and rectied linear
unit (ReLU) activations. The second component is the cross-correction
sub-network (CCS), which reduces the impact of multiple pooling op-
erations. The decoding block, situated in the second layer of the
network, enhances contrast and combines multiple encoding features. In
a related study by Fu et al. [27], the segmentation of the optic disc (OD)
is accomplished using a combination of the U-shaped network (UNet
[19]) and probability bubbles. The images are preprocessed with an
iterative robust homomorphic surface ltering method [28].
A two-stage technique aimed at alleviating the class imbalance
constraint problem was used by Meng et al. [29]. The candidate’s
location was determined through a guided search procedure, and a
weighted neighborhood voting approach was utilized to generate the
localized portable position. Preprocessing of optic disk (OD) and optic
cup (OC) images plays a crucial role in medical image processing. This
signicance is evident in the study conducted by Yuna et al. [30]. In
their research, they employed the contrast-limited adaptive histogram
equalization (CLAHE) technique [31] for enhancing the image quality
before transforming it into polar coordinates [32]. For segmentation
purposes, a CNN known as W-Net, comprising feature extractor and
context extractor modules, was utilized (refer to Fig. 5 for a clearer
visualization).
Due to the intricate connections between structures within the
retinal fundus images, the segmentation of candidate regions becomes
notably complex. Wang et al. [33] and Tan et al. [34] employed both
UNet and a basic CNN for retinal fundus image segmentation. Wang
et al. utilized the UNet approach to segment candidate regions within
the vessel density map, whereas Tan et al. employed the simple CNN to
simultaneously segment and classify three distinct components: vascu-
lature, optic disk (OD), and fovea. The images were normalized and then
inputted into the CNN, which consisted of six layers including two
convolutional, two max pooling, and two fully connected layers. In order
to underscore the signicance of preprocessing in retinal fundus images,
particularly focusing on the optic disk (OD) and optic cup (OC), Veena
et al. [35] employed Gaussian ltering and image normalization tech-
niques to eliminate undesired signals. Global distribution features and
the edge histogram texture descriptor (CLAHE and Sobel edge detector
[36]) were utilized for structure analysis and detection. The Watershed
algorithm [37] was employed to extract contour shapes and localiza-
tions. Finally, the end-to-end decoder-encoder method with 39 layers
was used. In a different study, Imtiaz et al. [38] introduce a semantic
approach that integrates automated augmentation into an
encoder-decoder architecture. The encoder block consists of 18 layers
comprising 13 convolutions and 5 pooling layers, while the decoder
block consists of 20 layers, including 14 convolutions, 5 pooling layers,
and 1 softmax layer. For further details, refer to Fig. 6.
In an independent study centered on the utilization of the UNet
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
5
model, Xie et al. [39] formulated a coarse-to-ne segmentation
approach aimed at establishing an initial segmentation boundary. The
CNN methods, specically a fully convolutional network, were com-
bined with the Viterbi algorithm to segment boundaries. The network
consists of three blocks: the encoder, decoder, and sequence decoder.
The rst two blocks involve traditional encoder and decoder layers,
while the sequence encoder layers comprise a gateway module and
cascaded gate units. The gateway module receives inputs from each
decoder layer and consists of three upsampling and three convolution
layers followed by sigmoid activation, while the gate unit comprises
ReLU and softmax layers. The Viterbi algorithm [40,41,42] was
employed to decode the output of the sequence decoder, modeling the
interaction of prediction and spatial constraints. Sadhukhan et al. [43]
introduced the attention-based fully connected CNN (AFCNN). This
network consists of 19 layers, including three attention blocks, 12
convolution layers, two dropout layers, and one softmax layer. In
Priyanka et al. [44], the zero-phase component analysis was applied to
the OD image, which was then augmented and fed into the CNN. After
being segmented with the CNN, the images were further segmented
using the fuzzy c-means method [45]. Similarly, Raja et al. [46] intro-
duced a nine-layer CNN architecture comprising three convolutional
layers, four ReLU layers, two max-pooling layers, one fully connected
layer, and one softmax layer.
Chowdhury et al. [47] developed a multiscale encoder/deco-
dernguided attention network for multicomponent segmentation of
retinal and optic discs. This network incorporates a self-attention
mechanism and extracts multiscale features. Hervella et al. [48] intro-
duced a pixel-level and image-level multi-task approach for simulta-
neously classifying glaucoma and segmenting the optic disc. Gupta et al.
[49] employ the U-Net architecture for segmentation and utilize the
mayy optimization kernel extreme learning [50] for classication.
Veena et al. [51] developed an enhanced deep learning method for
segmenting the optic disc and optic cup. Xiong et al. [52] introduced the
Bayesian U-Net Hough transform annotation method, which integrates a
Bayesian variant and the copy and crop framework to enhance the
weighting of the U-Net. In a different work, Maiti et al. [53] constructed
an encoder/decoder network comprising seven subnetworks and inte-
grated a long short-term memory framework. Moreover, multiple other
investigations, including Priyanka et al. [44] and Raja et al. [46],
employed CNNs for optic disc (OD) segmentation within retinal fundus
images. In a recent 2023 study by Rajarshi Bhattacharya et al. [54], an
encoder-decoder network was augmented with an auxiliary pyramid
decoder. This network features receptive blocks, a modied attention
block, and a spatial pyramid network. Notably, promising outcomes
were obtained, with dice scores of 94% and 95% achieved for the optic
disc and optic cup segmentation, respectively. The study referenced in
Fig. 4. Categories of CNN methods for retinal fundus images.
Fig. 5. Multi-scale CNN [30].
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
6
[55] employed a single-shot multibox detector comprising three key
components: a resize module, a feature extractor, and additional feature
blocks. A comparison of different methods in this section can be found in
Tables 1 and 2.
3.1.2. CNN used for arteries and vein
In the study by Girarda et al. [59], they put forth an encoder-decoder
CNN model aimed at segmenting arteries and veins. Median ltering was
utilized for preprocessing, followed by the stacking of resultant images
into the encoder-decoder network. The network comprises 32 feature
maps and a nal convolutional layer that reduces the map to three
classes: background, arteries, and veins. Morano et al. [60] introduced a
simultaneous segmentation module inspired by the UNet method. In this
study, fundus images were preprocessed with local intensity normali-
zation and channel-wise global contrast enhancement [60]. Subse-
quently, the UNet was utilized to predict the mask. It is noteworthy that
most of the methods employed for arteries and vein segmentation in
retinal fundus images are traditional methods, as observed in the
available literature. There is a scarcity of CNN-based approaches in this
area. For a detailed comparison of different methods in this subsection,
Fig. 6. Label-based semantic segmentation [38].
Table 1
CNN used for Optic Disk and Optic Cup.
Author Ref Year CNN Name Inspiration for research procedure Accuracy
(Accuracy)
Wan et al. [26] 2021 Asymmetric deep learning network UNet [19]
M-Net [56]
Segmentation 0.937
Fu et al. [27] 2021 Fusing UNet with probability bubbles UNet Segmentation 0.99
Meng et al. [29] 2018 RGV generated CNN model LeNet-5 [57] Segmentation 0.98
Yuan et al. [30] 2021 Multi-scale W-Net M-NET and UNet Segmentation 0.95
Wang et al. [33] 2019 Coarse-to-ne deep learning UNet Segmentation 0.93
Tan et al. [34] 2017 Single convolutional neural network Multiple segmentation Segmentation 0.96
Veena et al. [35] 2021 Deep learning enhanced CNN Encoder-decoder CNN Segmentation 0.98
Imtiaz et al. [38] 2021 Label based encoder and decoder semantic
segmentation
Encoder-decoder CNN Segmentation 0.86
Xie et al. [39] 2020 SU-Net and Viterbi algorithm UNet, dilated CNN [58], Viterbi
algorithm [41,42]
Segmentation –
Sadhukhan et al. [43] 2020 AFCNN FCNN Segmentation –
Priyanka et al. [44] 2017 Patches CNN CNN, Fuzzy C Means Segmentation 0.95
Raja et al. [46] 2020 Traditional CNN CNN Segmentation 0.90
Chwodhung et al. [47] 2022 MGAN Attention Network Segmentation 0.91
Hervella et al. [48] 2022 Multi task method CNN Segmentation 0.95
Gupta et al. [49] 2022 MOKEL UNet Segmentation and
classication
0.89
Xiog et al. [52] 2022 Bayesian UNet UNet Segmentation 0.94
Maiti et al. [53] 2022 Multiple convolutions VGG Segmentation 0.95
Rajarshi Bhattacharya
et al.
[54] 2023 UNet Auxilliary pyramid Network UNet Segmentation 0.95
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
7
please refer to Tables 3 and 4.
3.1.3. CNN used for retinal vessel
Retinal vessel segmentation is a long-standing problem in medical
image analysis [61,62], which is accompanied by several challenges,
including.
•Presence of various abnormalities of different sizes and shapes: The
presence of abnormalities surrounding the vessels in retinal fundus
(RF) images can hinder the effective segmentation of vessels.
•Limited availability of annotated data: The scarcity of annotated data
can lead to overtting issues, making it a major challenge when
segmenting vessels in FR images.
•Vessel structural differences: Retinal vessels exhibit variations in
thickness and structure, making it difcult to nd a single model or
network suitable for segmenting all types of vessels.
•Unstructured prediction: Pixel classication differs from vessel seg-
mentation, posing challenges in predicting the vessel structure
accurately.
To tackle these challenges, several authors have employed CNN
methods for vessel segmentation in RF images. Budak et al. [63] intro-
duced the Densely Connected and Concatenated Multi-Encoder-Decoder
CNN (DCCMED-CNN). DCCMED utilizes a patch-based learning network
and includes both training and testing phases. During training, color
patches extracted from raw retina images were used as inputs, and the
network weights were trained using stochastic gradient descent methods
[64]. The network architecture consists of 2 max-pooling layers, 2 max
unpooling layers, 8 concatenated convolution layers, batch normaliza-
tion, ReLU layers, and a softmax layer. Tang et al. [65] developed the
Multi-Proportion Channel Ensemble Model (MPC-EM) for retinal vessel
segmentation. MPC-EM comprises 5 submodel networks, with each
submodel following an encoder-center-decoder structure. A center ar-
chitecture is employed as a transitional region to adjust the shape of the
feature vector. The subnetworks were optimized using triple convolu-
tional residual blocks to enhance feature extraction and alleviate the
vanishing gradient problem [66]. In another approach, Zhao et al. [67]
employed a region-based CNN consisting of four parts: 1) backbone for
feature extraction, 2) region proposal network, 3) head module for
bounding-box regression, and 4) classication for mask generation. The
ResNet [68,69] was utilized for backbone feature extraction, and a
pyramid structure [70] was adopted to consider multiple scales. The
proposed network combined multi-task CNN with 27 layers, 15 con-
volutional layers, 4 duplicated feature maps, and 2 fully connected
layers.
Czepita and Nska [71] employed a shallow UNet architecture for the
segmentation of retinal fundus images. The method comprised six
stages: 1) Phase image registration: This stage involved aligning the
spatial domain of the fundus image to the moving spatial image domain
[72]. 2) Vessel probability map generator: The vessel probability stage
utilized an encoder-decoder CNN, specically the shallow UNet model,
to extract vessels from the fundus images. 3) Postprocessing: After the
initial segmentation, postprocessing techniques were applied to rene
the results and improve the accuracy of vessel delineation. 4) Second
segmentation: A second round of segmentation was performed to further
enhance the quality of vessel segmentation. 5) Region of interest se-
lection: Relevant regions of interest were selected to focus the analysis
on specic areas of the retinal image. 6) Vessel diameter measurement:
The retinal diameters were measured using a method similar to the one
proposed by Ref. [73], specically the calculation of central venular
equivalent (CRVE). Furthermore, Sun et al. [74] presented the multipath
cascaded UNet (MCU-Net) for retinal vessel segmentation. The MCU-Net
incorporated three types of data as input: raw FFA, small-scale FFA, and
large-scale FFA. The network fused vessel features from these inputs to
generate a vascular probability map as the output. It comprised an
attention gate [75] in conjunction with a residual recurrent unit [76].
The MCU-Net encompassed two specic blocks: the renement block
and the FFA image fusion block. For a more detailed understanding,
please refer to Fig. 7.
As evident from various studies, numerous researchers have
employed the UNet architecture RF segmentation. The UNet was
employed either in its original conguration or with adaptations to the
original structure. An illustration of such a modication can be found in
the work of Zhao et al. [69], who introduced the Nested U-shaped
attention network (NUA-Net) for segmenting and classifying retinal
images. The image was rst enhanced thereafter, the green channel
images were used as the network inputs (the same as the methods in
Ref. [77]). The NUA-Net has an encoder stage and each encoder stage
has a 2 ×2 max-pooling followed by convolution with batch normali-
zation, ReLU, and dropout. Guo et al. [78] proposed the multiscale
deeply supervised network with short connection (BTS-DSN). This
network used short connections to transfer semantic information be-
tween side-output layers. Two approaches were considered: the
bottom-top short connections and the top-bottom short connections. The
key element of this network is the top-bottom, and bottom-top short
connection approaches. A switch of connectivity within layers gives the
BTS-DSN a exible procedure. Guo et al. [79] proposed the multiple
deep convolutional neural network (MDCNN) for a formulated classi-
cation and segmentation task. The MDCNN was constructed by
cascading multiple networks with the same structure.
Noh et al. [60] proposed the scale-space approximation for
multi-scale representation in CNN (SSANet). The SSANet consists of 3
Table 2
Pros and cons of CNN used for Optic Disk and Optic Cup.
Advantages Disadvantages
Directly training a model using an end-to-
end process on both source and target
framework.
Effortlessly transforms data between
different forms.
Might exhibit excessive loss.
In case of inadequate conguration, it
can generate incorrect decoding result.
It is simple to overlook crucial features
originating from the encoder.
Noise detection is readily achievable due
to the superior pixel capture
methodology.
Images containing fewer symmetrical
components are acquired more quickly.
Prone to generating deceptive results.
Results from symmetric components
are notably poor.
Table 3
CNN used for Arteries and Vein.
Author Ref Year CNN Name Inspiration for research procedure Accuracy
Girard et al. [59] 2019 Joint segmentation model UNET Segmentation 0.96
Morano et al. [60] 2021 Simultaneous segmentation UNET Segmentation 0.96
Table 4
Pros and cons of CNN used for Arteries and Vein.
Advantages Disadvantages
Demonstrates efcacy in
preserving edges.
Achieves accurate outcomes
with a reduced image count.
A low noise ratio can disrupt image edges and
generate misleading edge noise, potentially
impacting accuracy.
Errors might arise during the normalization
process.
Due to the network’s shallowness, exceptional
accuracy might not be attainable.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
8
blocks, (feature generation, feature aggregation, and inference) with 33
layers. The feature generation block has 21 layers consisting of 1
convolution layer, 3 upsampling layers, and 17 ResBlock layers. The
aggregation stage perform two key procedure: it moves input before
each upsampling in the generation block. Secondly, it accepts inputs
from the nal block of the generation using 9 layers (5 Convolutional
layers, and 4 upsampling layers). The inference block collects inputs
from the aggregation and transforms these inputs to a mask. Zhuo et al.
[80] combined the size-invariant feature maps [81] with the dense
connectivity [82] (SID
2
Net) for the segmentation and classication of
RF. The SID
2
Net has two bottleneck modules, three dense blocks, two
convolutional layers and a sigmoid layer for prediction. An ablation
experiment was carried out dividing the network into the dense network
(DNet), and DNet with size-invariant feature maps (SIDNet). Hervella
et al. [83] used the multi-instance heating regression to predict RF
image segmentation. This method predicts binary maps with the pixels
corresponding to the location and labeling of the positive class of the
ground truth. The RF images were passed to the UNET framework and
the results were interpolated back into the original RF images.
The skip-chain convolutional network is a model design integrating
skip connections or shortcuts among layers, aiming to amplify infor-
mation ow and improve feature acquisition. For the specic task of
blood vessel segmentation, a vessel-specic skip-chain convolutional
network (VSSC Net) developed by Samuel et al. was employed. The
VSSC Net operates through two stages: preprocessing and segmentation.
The preprocessing stage converts RF images to grayscale then the
adaptive fractional difference approach [84] followed by the CLAHE is
applied to the grayscale image. The segmentation stage (VSSC Net) is an
end-to-end framework that takes input images of arbitrary size pro-
ducing a probability map. VSSC Net has two components: base network
architecture and novel architecture. The base network consists of
different convolutional layers split into 4 pairs. The visual geometry
group (VGG-16) [85] was used as the base network. The attention-based
before-activation residual UNet (BSEResU-Net) [86] exploits the atten-
tion mechanism and the dropblock regularization method to reduce
overtting. The images were preprocessed by transforming RGB images
to grayscale, and the CLAHE algorithm was applied to the grayscale
image. The ResU-Net has 33 convolutional layers with 16 residual op-
erations, 2 transpose convolutional layers, 2 downsampling layers, and 1
output map. Huang et al. [87] proposed the multipath scale network
(MPS-Net) for retinal vessel segmentation. The MPS-Net is an end-to-end
network that uses one high-resolution RF input and produces a proba-
bility map with two low resolutions as output. The network has 13
multi-path scale modules, 3 convolution +ReLU, 3 Normalization +
ReLU, and 1 cropping layer. The multi-path scale module has 3 regional
paths concatenated together and arranged horizontally to produce the
output. The range entropy [88] denition was introduced to describe
vessel information of the feature maps.
Tian et al. [89] proposed the multipath CNN for RF segmentation.
This network converts the original image to low-frequency and
high-frequency images with the low-pass Gaussian lter and the
high-pass Gaussian lter. The CNN consists of a convolution down-
sampling and convolution upsampling and has 32 convolutional layers
with four blocks of 64, 128, 64, and 32. Ultimately, the outcomes of the
initial and secondary CNN were merged (fused) together to generate the
conclusive segmentation result. To nd out if there is a difference be-
tween the preprocessed image and the images without preprocessing,
Atli and Gedik [90] used the Sin-Net for the segmentation of vessels in
RF images. To preprocess, the CLAHE and the multi-scale top hat
transform (MTHT) [91] were used to enhance image contrast. The
Sin-Net architecture consists of 17 layers comprising 11 convolution
operations, 2 upsampling layers, 2 down-sampling layers, 1 output and
input layer each.
The usage of reinforcement learning in RF images is gaining promi-
nence. Guo et al. [92] used CNN with reinforcement learning to segment
vessels in RF images. The images are divided into smaller patches and
sent to CNN for training. The CNN has ve components: convolution,
pooling, dropout, fully connected, and loss function. Deep CNN [93] has
received tremendous recognition in medical image processing. As an
example, Wu et al. [94] introduced the Network Followed Network
(NFN+), which encompasses four distinct modules: 1) encoder and
decoder of the initial network, 2) encoder and decoder of the subsequent
network, 3) front group of intra-network skip connections, and 4) sec-
ond group of inter-network skip connections. In summary, the NFN +
architecture incorporates two connections involving the front and fol-
lowed networks, encompassing a total of 10 integrated components,
including convolutions, batch normalization, and dropout layers. Fully
convolutional networks (FCN) have gained relevance in tasks related to
nonmedical imaging. However, such tides are changing, Hemelings et al.
[95] used the FCN for segmenting retinal vessels in RF images. They
used the method adopted by Ref. [96] to pad the region of interest to
avoid excessive contrast enhancement at the border of the image. The
CLAHE, and noise removal methods were used to preprocess the RF and
passed to network architecture for segmentation. Boudegga et al. [97]
proposed RV-Net for vessel segmentation. This method preprocesses the
RF images by replacing the black area with an average color (see
Ref. [98]), then the image was converted to LAB. The CLAHE algorithm
is applied to the image and the channels are merged and converted back
to RGB. The preprocessed image was augmented by performing image
transformation, cropping, and patch extraction [99]. The images were
fed into the network for segmentation by the RV-Net. Wang et al. [100]
proposed the hybrid CNN and ensemble random forest (RFs) [101]
method. CNN was used for segmentation while the RF was the trainable
traditional method used for classication. The CNN has 5 layers con-
sisting of the convolution, subsampling, and fully connected layer. Hu
et al. [102] conducted a study introducing a multiscale CNN with a
cross-entropy loss function for retinal vessel segmentation. The original
RF image underwent augmentation before being fed into the network.
Fig. 7. MCU-net [74].
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
9
The multiscale CNN framework comprised four distinct stages. The rst
and second stages comprised four convolutional layers and one
max-pooling layer each. At the beginning of the network, there were a
total of 20 convolutional layers and three max-pooling layers. For a
more comprehensive understanding, please refer to Fig. 8.
Zhou et al. [103] proposed the symmetric equilibrium generative
adversarial network (SEGAN) for vessel segmentation. The SEGAN is an
end-to-end synthetic neural network that utilizes the adversarial prin-
ciple. Three fundamental principles were employed in this study:
SEGAN, multiscale feature rene block (MSFRB), and attention mech-
anism (AM) [104]. Overall, there were 13 traditional layers, 5 MSFRB,
and 5 a.m. layers. The research by Ref. [105] proposed a hybrid mul-
titask deep learning for segmenting vessels. The original image was
annotated before being fed into the deep learning algorithm. The
network has two modules: 1) multitask segmentation, 2) fusion network
module. For both modules, the improved UNet framework was adopted
for segmentation and fusion. The network is an encoder-decoder seg-
mentation consisting of 20 layers. The output of the segmentation was
passed to the fusion layer for the nal output. The deformable U-Net
(DUNet) proposed by Jin et al. [106] is a U-shaped architecture with an
encoder and decoder framework. Some of the convolutional layers in the
traditional UNet were replaced with deformable convolutional blocks.
The DUNet integrates the low-level features with high-level features.
The design was constructed with 4 convolution layers, 4 batch nor-
malizations, and 4 ReLU layers. Soomro et al. [107] proposed the strided
fully connected CNN for the segmentation of vessels in RF images. The
images were preprocessed with morphological tactics and the principal
component analysis (PCA) [108]. The network has 5 fully consecutive
convolutional blocks with sizes ranging from 16, 32, 64,128, and 256
features. There is no ablation experiment in this research.
Chala et al. [109] proposed the end-to-end improved CNN for vessel
segmentation. This network used the multi-encoder-decoder principle
and a new progressive reduction model that was integrated into the
network. The network has 4 interconnected components (multi-encoder
and parallel components, RGB-encoder and green channel encoder,
decoder component, and progressive reduction components). Data
augmentation was performed to generate more data for the network.
Sun et al. [110] proposed the contextual information enhanced UNet
(CIEU-Net) with dilated convolutional module for vessel segmentation.
The cascaded dilated module and the pyramid module were integrated
to form the segmentation network. Wu et al. [111] proposed the scale
and context-sensitive network for the segmentation of vessels. The
model consists of three modules: scale-aware feature aggregation (SFA),
adaptive feature fusion (AFF), and multi-level semantic supervision
(MSS). The SFA adjusts the receptive eld dynamically to extract fea-
tures. The AFF guides the fusion between features efciently. Satha-
nanthavathi and Indumathi [112] proposed the enhanced encoder
atrous UNet (EEA-UNet) for retinal vessel segmentation. The images
were preprocessed with the CLAHE and resized to 512 ×512.
Post-processing was done by morphological operations to remove iso-
lated false positives. The EEA-UNet is an asymmetric contraction and
expansion path that replaces all the convolutions as the atrous convo-
lution to increase the receptive eld. The contracting has 5 blocks
containing 2 atrous convolutions, batch normalization, pooling, and
ReLU layer. The atrous convolution reduces the image size without
losing the signicant features in it [113].
Yin et al. [114] proposed a U-shaped deep learning fusion network
(DF-Net) for vessel image segmentation. The method involves 4 stages:
multiscale fusion, U-shaped network, feature fusion, and classier
fusion. The original image was multiscaled by the image pyramid [115]
and constructed on a multiscale input integrated into the encoder path
for information fusion. The encoder of the network encompasses 2
convolution layers, coupled with max-pooling employing ReLU activa-
tion. The decoder, on the other hand, entails 2 convolutional layers,
succeeded by up-sampling with the number of feature maps halved. The
vessel fusion module is attached to the decoder and enhanced with the
corresponding output features. This network is combined with the
Frangi lter and a deep neural network was trained. The research by
Tang et al. [116] adopts the multiscale method. The authors developed
multi-scale channel importance sorting (MSCS) for vessel segmentation.
First, the CLAHE algorithm was used to enhance the image before it was
fed to the network. The MSCS is an encoder-decoder that consists of 3
encoders and 2 decoder blocks. Each encoder block consists of
multi-scale, channel importance, and a convolution layer. Between the
encoder and decoder, the spatial attention mechanism was used instead
of the traditional skip connection to read the output and characterize the
encoder generating the attention coefcients. The research by Ref. [117]
proposed the cascaded attention residual network (AReN-UNet). The
AReN-UNet features a unique aggregated residual module, which in-
corporates concatenated max and average pooling, along with a shared
multilayer perceptron (MLP) [118] concluded with a sigmoid layer. Shi
et al. [119] developed the multiscale dense network (MD-Net) that
makes good use of the multi-scale features and the encoder features. This
network is preprocessed with the CLAHE algorithm. A residual atrous
spatial pyramid pooling (Res-ASPP) was blended into the error frame-
work and the dense multi-level fusion merges the features in the encoder
and decoder. A squeeze and extraction (SE) block was applied to the
concatenated layer for effective feature channels. The Res-ASPP has 12
layers all of which are convolution layers with varying dimensions and
sizes. The multi-level fusion mechanism and SE block perform the fusion
procedure in the network.
Tchinda et al. [120,121] used the combination of edge detection and
neural network to segment vessels in RF images. The method used
feature vectors with eight characteristic pixels. The feature vectors
include 1) image gradient obtained with edge detection (Prewitt, Sobel,
Canny, and Gaussian [122,and123]]), 2) the Laplacian of Gaussian l-
ter, 3) morphological transformation (erosion, dilation, and top hat
ltering [124]). The cascaded feed-forward network was used for seg-
mentation. Gegundez-Arias et al. [125] proposed the simplied UNet for
the segmentation of RF images. A combination of the residual block and
batch normalization in the upsampling and downsampling layers pro-
duces the required segmentation results. The simplied UNet has 10
blocks consisting of 1 CONV_ReLU1 layer, 1 convolution layer, 3 Block2
layers, 2 Block I1, and 3 Block I2 layers. Skip connections were used to
link Block I1 and Block I2 together. Maji et al. [126] combined the
attention-based neural network with transfer learning. This research
Fig. 8. Multiscale CNN [102].
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
10
used the optimized learning method to classify and grade RF images. The
attention mechanism adds attention to pixels near the vessel. The
Gaussian lter was used to normalize color balance and illumination,
then the data was augmented. The softmax graded the health risk with
0 as bad and 2 as good. Sangeethaa and Maheswari [127] used
morphological process, thresholding, edge detection, and adaptive his-
togram for segmentation, while CNN was used for classication. After
preprocessing and segmentation, the image was fed into the trained CNN
for classication (either normal or diseased).
Muthusamy et al. [128] used the CNN recurrent network (CNN-RRN)
to segment retinal images. The image was rst preprocessed with the
median lter for denoising and smoothing. Then, the image was resized
with dual-tree complex wavelet transform and then the classication
was done using renewal networks (CNN with recurrent neural network
concept [129]). Shi et al. [130] developed a graph-based convolutional
feature aggregation network for segmenting retinal vessels. The model
consists of multilevel feature extraction, and a graph-based high and low
aggregation model. In this network, features are extracted at different
levels and passed to high and low aggregation models for preprocessing
and segmentation. Song Guo [131] proposed the cascaded guided
network for retinal vessel segmentation. This network consists of three
branches each with a scale of different sizes. Each of the branches
consists of connected multiscale multidirectional feature learning
modules to learn under the low-resolution feature and segments se-
mantic maps as output. Ren et al. [132], combined the U-Net network
and the BiFPN [133] for the segmentation of retinal vessel images. This
network is an encoder-decoder network and the BiFPN was used as the
skip between the encoder and decoder. Hang et al. [134], proposed the
improved U-Net with the skip layer replaced with a multilayer
connection (NOLnet). The NOL block consists of transpose convolutions,
batch normalization, and concatenation layers. Xu & Fan [135] pro-
posed the dual-channel asymmetric convolution neural network. This
network consists of two parts (mainsegment Net and Finsesegment Net).
The original image was divided into two thick feature maps, with each
feature map used as input to the two networks.
Zhong et al. [136]; developed a multiscale dilated convolutional
neural network (MMDC-Net) to capture information from receptive
elds. They perform a cascading procedure on the U-Net and used a
multilayer fusion (MLF) model to fuse features and lter noise. A width
attention CNN (WA_net) proposed by Ref. [137] for retinal vessel seg-
mentation was used to decompose multiple channels of feature maps
into categories. Deng & Ye [138] proposed a multi-scale attention
mechanism residual deformable convolution that combines the M-sha-
ped network with the pulse-coupled neural network. Similarly, Huang
et al. [139], proposed the cascaded self-attention U-shaped network. The
network used the U-Net and the residual self-attention model to segment
retinal fundus images. The network progressed from ne to coarse
segmentation before producing the nal mask. Pavani et al. [140],
proposed the multistage dual-path interactive renement network. This
network follows a single encoder and dual decoder style. Data
augmentation with model implementation was used to boost the
network. Karlsson & Hardarson [141] proposed a CNN based on several
interconnected U-Nets. The network was simultaneously connected and
sandwiched with global features and exceptional linear units [142]. Zhu
et al. [143], proposed an inception-based U-Net that consists of two
parts: inception downsampling (connected to the encoder) and incep-
tion upsampling (connected to the decoder). A copy framework was
added to the decoder to help rene the network. The network proposed
by Yin et al. [144], used the U-Net, feature fusion, and classication
fusion network. The network was a multiscale network used to fuse all
features after segmentation. The segmentation outputs from each
encoder were downscaled and combined to yield the nal outcome.
Khan et al. [145] introduced a novel approach called the
multi-resolution contextual network for the segmentation of vessels in
RF. This network leverages multiscale features to capture contextual
dependencies through a bi-directional context learning framework. Liu
et al. [146] conducted a study utilizing the dual attention Res2Unet
(DA-Res2UNet) for retinal vessel segmentation. In this method, the
convolution layer in the traditional UNet was replaced with the Res2-
Block and a DropBlock. The spatial attention and dual attention mech-
anisms were key components of this approach. Comparative analysis
demonstrated a segmentation accuracy of 97% when compared to other
state-of-the-art methods. In a separate study, Liu et al. [147] proposed
the attention augmented Wasserstein generative adversarial network
(AA-WIGAN) for retinal vessel segmentation. The network employed an
advanced UNet as the generator and a discriminator. Experimental re-
sults exhibited promising outcomes with a segmentation accuracy of
97%. For further details and a comprehensive comparison of different
methods, please refer to Tables 5 and 6.
4. Performance measures
The evaluation metrics used to assess the performance of various
CNN methods in the segmentation and classication of retina fundus
images include the following.
1) True Positive (TP): The number of correctly identied positive in-
stances or correctly segmented/classied abnormalities in the image.
2) False Positive (FP): The number of incorrectly identied positive
instances or incorrectly segmented/classied abnormalities in the
image.
3) False Negative (FN): The number of incorrectly missed positive in-
stances or incorrectly unsegmented/unclassied abnormalities in the
image.
4) True Negative (TN): The number of correctly identied negative
instances or correctly classied background/non-abnormal regions
in the image.
5) Intersection over union (IOU): measures the degree of overlap be-
tween the ground truth (GT) and the segmented prediction, quanti-
fying it within a range of 0–1. This overlap assessment, commonly
referred to as the Intersection over Union (IOU), is also synonymous
with the Jaccard index.
IOU =GT ∩Prediction
GT ∪Prediction (1)
6) Recall: this method characterizes the relevant portion of the positive
segmentation prediction in relation to the ground truth. It aligns with
the concept of sensitivity, indicating the proportion of positive cases
accurately classied by the method.
Recall =Tp
Tp +Fn (2)
7) Precision: This metric quanties the positive predictions in relation
to the ground truth.
Precision =Tp
Tp +Fp (3)
8) Accuracy (ACC): This measurement computes the sum of true posi-
tive and true negative, divided by the total number of positives and
negatives.
ACC =Tp +Tn
Tp +Tn +Fp +Fn (4)
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
11
9) Specicity: This term denotes the ratio of negative cases that were
accurately classied.
Specificity =Tn
Tn +Fp (5)
10) Area under the curve (AUC): his metric characterizes the classi-
er’s capability to distinguish between classes. A higher value of
the Area Under the Curve (AUC) indicates superior performance.
AUC =∫
b
a
F(x).dx (6)
Table 5
CNN used for Retinal vessel.
Author Ref Year CNN Name Inspiration for research procedure
Budak et al. [63] 2020 Densely connected/concatenated multi-encoder-decoder
CNN
Feedforward CNN Segmentation
Tang et al. [65] 2019 Multi-proportion channel ensemble model Ensemble model Segmentation
Zhao et al. [67] 2020 RCNN-based junction renement network Masked RCNN-model [148] Segmentation
Yuan et al. [71] 2021 Shallow U-Net UNet Segmentation
Sun et al. [74] 2021 Multi-path cascaded UNet UNet Segmentation,
Zhao et al. [69] 2021 Nested U-shaped attention network UNet Segmentation
Guo et al. [78] 2019 Bottom-top and top-bottom short connection deep
supervised network
Deep supervised network Segmentation
Guo et al. [79] 2018 Multiple deep CNN Deep CNN [149] Segmentation
Noh et al. [150] 2019 Scale-space approximated CNN DRIU [151] Segmentation
Zhuo et al. [80] 2020 Size-invariant and dense connectivity network DenseNet Network [82] Segmentation
Hervella et al. [83] 2020 Multi-instance heat map regression DNN Segmentation
P. M Samuel & T. Veeramalai [152] 2021 Vessel Specic Skip chain CNN Fully convolutional networks Segmentation
D. Li & S. Rahardja [86] 2021 Attention-based before-activation residual U-Net Modied UNet Segmentation
Lin et al. [87] 2021 Multi-path scale network HR-Net [153] Segmentation
Tian et al. [89] 2020 Multi-path CNN UNet Segmentation
I. Atli & O. S. Gedik [90] 2021 Sine-Net CNN Fully CNN Segmentation
Guo et al. [92] 2018 CNN with reinforcement sample learning. Reinforcement learning Segmentation
Wu et al. [94] 2020 A network followed network. Deep CNN [93] Segmentation
Hemelings et al. [95] 2019 Fully convolutional network. UNet Segmentation
Boudegga et al. [97] 2021 RV-Net UNet, AlexNet [154], VGG Segmentation
Wang et al. [100] 2015 Features and ensemble learning CNN, and RFs Segmentation and
classication
Hu et al. [102] 2018 Multiscale CNN Richer convolutional features
[155]
Segmentation
Zhou et al. [103] 2021 Equilibrium GAN UNet, GAN [156] Segmentation
Yang et al. [105] 2021 Improved UNet UNet Segmentation
Jin et al. [106] 2019 DUNet UNet, Deformable convNet
[157]
Segmentation
Soomro et al. [107] 2019 Strided FCNN SegNet [158] Segmentation
Chala et al. [109] 2021 Improved deep CNN DCNN Segmentation
Sun et al. [110] 2021 CIEU-Net UNet Segmentation
Wu et al. [111] 2021 SCS-Net UNet Segmentation
Sathananthavathi &
Indumathi
[112] 2021 EEA UNet UNet Segmentation
Yin et al. [114] 2021 DF-Net UNet Segmentation
Tang et al. [116] 2020 MSCS UNet Segmentation
Rahman et al. [117] 2021 Cascaded AReN-UNet UNet Segmentation
Shi et al. [119] 2021 MD-Net SegNet, PSPNet [159], UNet Segmentation
Tchinda et al. [120] 2021 Classical edge detection and neural network Articial neural network Segmentation
Gegundez-Arias et al. [125] 2021 Simplied UNet UNet Segmentation
Maji & Sekh [126] 2020 Tradition method with CNN CNN Classication
Sangeethaa & Maheswari [127] 2018 Trained CNN CNN Segmentation and
Classication
Muthusamy & Tholkapiyan [128] 2019 CNN-RNN CNN Segmentation and
Classication
Shi et al. [130] 2022 Graph-based network Multilevel method Segmentation
Guo [131] 2022 Cascade guided network Multiscale method Multiscale methods
Ren et al. [132] 2022 Combined U-Net UNet Segmentation
Huang et al. [134] 2022 Improved U-Net UNet Segmentation
Xu & Fan [135] 2022 Dual semantic CNN CNN Segmentation
Zhong et al. [136] 2022 MMDC-Net UNet Segmentation
Deng et al. [138] 2022 Multiscale attention mechanism Attention network Segmentation
Huang et al. [139] 2022 Cascaded self-attention UNet Segmentation
Pavani et al. [140] 2022 Decoder/encoder network UNet Segmentation
Karlsson & Hardarson [141] 2022 CNN based U-Net UNet Segmentation
Zhu et al. [143] 2022 Inception U-Net UNet Segmentation
Yin et al. [144] 2022 Multiple networks UNet Segmentation and
classication
Renyuan Liu et al. [146] 2023 dual attention Res2Unet UNet Segmentation
Meilin Liu et al. [147] 2023 AA-WIGAN UNet Segmentation
Khan et al. [145] 2023 Multiscale context Network Reccurent Network Segmentation
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
12
Where a and b are the upper and lower limits on the curve, and F(x) is
the equation of the curve.
11) Dice measure (DM): This measurement evaluates the degree of
overlap between the predicted segmentations and the ground
truth. Higher values indicate accurate predictions.
DM =2∗/prediction ∩GT/
/Prediction/+/GT/(7)
12) F1 scores: measures the function of precision and recall.
F1=2∗Precision ∗Recall
Precision +Recall (8)
13) Mathew correction coefcient (MCC): measures the number of
true and false positives and negatives.
MCC =Tp ∗Tn −Fp ∗Fn
(Tp +Fp)(Tp −Fn)(Tn +Fp)(Tn +Fn)
√(9)
These metrics provide insights into the accuracy and effectiveness of
the CNN methods in detecting and classifying abnormalities in retina
fundus images.
5. Datasets
This section provides an overview of the datasets utilized for seg-
menting and classifying retina fundus images. The datasets are pre-
sented in the order of their usage, including DRIVE [160], STARE [161],
CHASE-DB1 [162], HRF [163], MESSIDOR [164], IOSTAR [165],
ORIGA [166], REFUGE [166], DB1 [167], DB0 [168], Kaggle dataset,
DRISHTI-GS [169], NIVE [170], RIM-ONEr3 [171], DRIONS-DB [172],
RITE [173], WIDE [174], SYNTHE [175], LES-AV [176], RIGA [177],
DUKE, DCA, EIARG1, and AV-INSPIRE. Fig. 9 illustrates the list of
datasets with the highest usage, while datasets with a usage count of less
than 2 were excluded from the graph. The prominent datasets in terms of
usage are DRIVE, STARE, CHASE-DB1, and HRF. It is evident that a
valuable dataset should possess relevance, usability, and high-quality
attributes [178]. Therefore, we posit that datasets with greater usage
exemplify these aforementioned characteristics. In summary, the ma-
jority of databases discussed in this study are meticulously curated re-
positories of retinal fundus images. These databases represent
invaluable assets for medical research, fostering the creation of sophis-
ticated diagnostic tools and therapies to enhance both ocular health and
patient well-being.
6. Analyses
The authors conducted a comprehensive search across various online
repositories to identify research papers focused on the classication and
segmentation of retina fundus images using CNN. Initially, a total of 300
papers were gathered through this search process. Subsequently, these
papers underwent a screening process, resulting in a selection of 170
relevant papers. The authors then specically focused on papers related
to arteries, vessels, and optic discs/cups, as illustrated in Fig. 10. From
the narrowed scope, a total of 80 research papers were included in this
review, specically addressing the classication and segmentation of
veins, optic discs, and blood vessels using CNN. Among these papers, 60
were obtained from ScienceDirect, 15 from Springer, and 5 from other
sources. While the authors made efforts to include all relevant research
articles within the scope of their study, it is possible that some
Table 6
Pros and cons of CNN used for Retinal vessel.
Advantages Disadvantages Accuracy
range
Data management is executed
with a high degree of
prociency, potentially leading
to enhanced accuracy.
The network’s substantial depth
equips it to manage intricate
representations adeptly.
It is skillfully constructed with
residual blocks to mitigate the
vanishing gradient issue.
This design facilitates reduced
gradients, bias, and improved
data comprehension.
Increased parameter count can
lead to information loss,
consequently elongating
network processing time.
Inability to execute
translational invariant
procedures.
Prone to generating deceptive
results.
Demands substantial
computational resources.
0.79–0.97
Fig. 9. Datasets used for classication and segmentation of retina fundus images.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
13
interesting studies may have been inadvertently overlooked.
For a more in-depth examination of the data origins and the
dispersion of publications across various years, refer to Fig. 11. To
explore the contrasts in accuracy between optic disc (OD) and optic cup
(OC) segmentation, consult Fig. 12. Observing Fig. 12 reveals that the
spectrum of achieved accuracies spans from a minimum of 86% to a
maximum of 99%. Notably, a signicant proportion of accuracy scores
cluster within the range of 93%–96%. These visual representations serve
to augment our understanding of the origins of the incorporated studies
as well as the chronological distribution of research ndings within the
eld.
This review encompassed a total of 2 papers focusing on arteries and
veins, 15 papers on optic disc and cup, and 58 papers on retina vessels.
Five additional research papers published in 2023 have utilized CNN for
disease detection, employing the entire retinal fundus image for anal-
ysis. Among these, 4 studies employed a combination of CNN methods
with traditional approaches, as indicated in Tables 6 and 2 Notably,
reference [126] solely concentrated on classication, whereas refer-
ences [127,128], and [100] presented segmentation and classication
approaches specically for retina fundus images. Fig. 12 provides an
analysis of the accuracy metric specically for the optic disc and cup.
However, for a comprehensive overview of accuracy and other perfor-
mance measures, referring to the tables is recommended, as they contain
additional information on various metrics across different studies.
Fig. 10. Approach for the review process.
Fig. 11. Number of papers published yearly.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
14
7. Conclusions and future directions
The objective of this study is to provide a comprehensive review of
CNN methods used for the segmentation and classication of retinal
fundus images. These CNN algorithms have demonstrated impressive
performance, achieving approximately 99% accuracy in certain cases. In
the eld of medical practice, these CNN methods have provided valuable
insights that clinicians have utilized to draw conclusions about specic
pathologies. They have aided in disease diagnosis and the prediction of
ophthalmic conditions in patients. The overall performance of CNN al-
gorithms in this domain has positioned them as promising alternatives to
traditional methods. Notably, the IDX (Intact Dilated Extraction) tech-
nique has already gained approval for practical use in medium and large
healthcare facilities, physician groups, and medical management rms
[179,180,181]. As CNN algorithms continue to improve in stability and
efciency, their relevance is expected to be widely accepted, opening
doors for their application in other critical areas of healthcare delivery.
7.1. Observations
•From the reviewed papers, it is evident that the majority of CNN
methods for retinal fundus image segmentation are built upon the
UNet, attention mechanisms, and adversarial networks. This pref-
erence can be attributed to the strong performance of these net-
works, particularly the UNet, in medical image tasks. Considering
that retinal fundus image segmentation requires a network that
accurately addresses the task, leveraging the UNet architecture is a
suitable choice.
•The utilization of preprocessing methods has contributed to
improved accuracy. Many researchers have employed techniques
such as Contrast Limited Adaptive Histogram Equalization (CLAHE)
for image enhancement and various lters for denoising. While the
use of different denoising methods has not signicantly enhanced
results, the application of CLAHE for image enhancement has
demonstrated performance improvements.
•Conversion of color retinal fundus images from RGB format to LAB
and subsequent conversion back to RGB after preprocessing has been
a common practice. This conversion facilitates a conducive envi-
ronment for preprocessing procedures.
•Acquiring a large dataset is a challenge in medical image analysis due
to limited availability. To address this, most CNN methods have
employed data augmentation techniques as a preprocessing step to
expand the dataset.
•Multilevel, multiscale, and multidirectional networks have been
prevalent in papers published in 2022. These networks involve the
combination of multiple encoder and decoder frameworks.
•Retinal vessel segmentation has emerged as the most commonly
studied application of CNN methods for segmentation and classi-
cation, likely due to the prominent visibility of retinal vessels in
fundus images.
•Recent trends involve incorporating attention mechanisms, Bilinear
Feature Pyramid Networks (BIFPN), or other networks into the skip
connections of the UNet architecture. Many CNN networks proposed
in 2022 have adopted this approach.
7.2. Limitations and future directions
Although CNN algorithms have made signicant contributions to
fundus image segmentation and classication, there are still challenges
that need to be addressed when applying CNN to retinal fundus (RF)
images.
1) Limited availability of labeled quality data: The quality of the
labeled data directly inuences the performance of CNN algorithms,
and obtaining a sufcient amount of accurately labeled data for RF
images remains a challenge. One reason is that experts involved in
labeling have busy schedules and may not have enough time to
accomplish the task. For instance, achieving reliable labeling for
breast ultrasound images typically requires the participation of three
experts who perform the labeling and make majority-voting de-
cisions to obtain the nal labeled image. This process is demanding,
time-consuming, and requires signicant effort. To address this
issue, researchers can explore methods like the ones described in
Ref. [182], which involve a combination of algorithms and human
supervision to generate labeled data. Additionally, weakly super-
vised methods can be employed to leverage the available labeled
data and expand its usage in the context of unlabeled data. These
approaches offer potential solutions to mitigate the scarcity of
labeled data in RF image analysis. unlabeled data.
Fig. 12. Accuracy comparison for optic disc and cup.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
15
2) Performance of CNN methods in a real-life situation: Many of the
CNN methods evaluated in this review lack testing in real-life sce-
narios, leaving uncertainty about their optimal performance when
applied to practical applications. To address this concern, it is crucial
for analysts and computer vision engineers to establish close
collaboration with clinicians. By working hand-in-hand, these
interdisciplinary teams can ensure a gradual and informed deploy-
ment of CNN algorithms in real-life settings. This collaborative
approach allows for the integration of clinical expertise, validation
studies, and iterative improvements, thereby enhancing the reli-
ability and effectiveness of CNN algorithms when used in real-life
applications.
3) Unavailability of standard measuring metrics: The use of diverse
evaluation metrics for reporting quantitative results is evident in the
research, as these metrics capture different aspects of performance.
Consequently, authors have the exibility to choose any metric from
the available list to quantify their ndings. While this review has
addressed the issue to some extent by aggregating two common ac-
curacies from the studies, there is still room for further efforts to
minimize the proliferation of metrics. We propose the adoption of a
standardized procedure that selects three or four metrics as a
benchmark for the segmentation and classication of RF images. By
establishing this standardized set of metrics, uniformity can be
achieved in the evaluation process, allowing for better comparison
and interpretation of results across different studies.
4) CNN high computation space: The resource-intensive nature of
CNN algorithms is widely recognized, often necessitating a signi-
cant computational infrastructure. Inadequate computational re-
sources can pose challenges to achieving higher accuracies in
segmentation tasks. To address this issue, we call for the develop-
ment of additional cloud computing platforms, such as Collab, that
offer ample computational space. These platforms should be acces-
sible to researchers at minimal or no cost, facilitating the execution
of their applications and alleviating the burden of limited compu-
tational resources. By expanding the availability of such cloud
computing platforms, researchers can harness the necessary
computational power to enhance the performance of CNN algorithms
in segmentation tasks without incurring substantial expenses.
Author contributions
All authors contributed to the study’s conception and design. Mate-
rial preparation and data collection were performed by AEI. Analysis
was done by AEI, TI., AG. The manuscript was written and polished by
AEI.
Funding
No funding applicable to this research.
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Data availability
No data was used for the research described in the article.
Acknowledgements
We would like to sincerely express our deep gratitude to our
esteemed mentors: Professor David A. Wolk from the Penn Memory
Center at the University of Pennsylvania, Professor Sandhitsu Das from
PICSL at the University of Pennsylvania, Professor Jayaram K. Udupa
and Professor Drew A. Torigian from the MIPG at the University of
Pennsylvania, and Professor Stanislav Makhanov at SIIT Thammassat
University. We also extend our heartfelt thanks to the management of
Alex Ekwueme Federal University Nigeria, as well as the anonymous
referees of the review, for their invaluable remarks and signicant
contributions.
References
[1] World Health Organization, Blindness and Vision Impairment, 2022. https
://www.who.int/news-room/fact-sheets/detail/blindness-and-visual-impairmen
t. June 19 2023.
[2] S. Sengupta, A. Singh, H.A. Leopold, T. Gulati, V. Lakshminarayanan, Ophthalmic
diagnosis using deep learning with fundus images - a critical review, Artif. Intell.
Med. 102 (2020 Jan), 101758, https://doi.org/10.1016/j.artmed.2019.101758.
Epub 2019 Nov 22. PMID: 31980096.
[3] M.T. Nicolela, J.R. Vianna, Pearls of Glaucoma Management, in: Optic Nerve:
Clinical Examination, Springer, Berlin, Heidelberg, 2016, pp. 17–26.
[4] T. Aziz, A.E. Ilesanmi, C. Charoenlarpnopparut, Efcient and accurate
hemorrhages detection in retinal fundus images using smart window features,
Appl. Sci. 11 (2021) 6391, https://doi.org/10.3390/app11146391.
[5] A.S. Krolewski, J.H. Warram, L.I. Rand, A.R. Christlieb, E.J. Busick, C.R. Kahn,
Risk of proliferative diabetic retinopathy in juvenile-onset type I diabetes: a 40-yr
follow-up study, Diabetes Care 9 (5) (1986) 443–452.
[6] T. Kauppi, V. Kalesnykiene, J.K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo,
H. K¨
alvi¨
ainen, J. Pietil¨
a, DIARETDB0: evaluation database and methodology for
diabetic retinopathy algorithms, Machine Vision and Pattern Recognition
Research Group, Lappeenranta University of Technology: Lappeenranta, Finland
73 (2006) 1–17.
[7] Age-Related Eye Disease Study Research Group, The relationship of dietary
carotenoid with vitamin A, E, and C intake with age-related macular
degeneration in a case-control study, Arch. Ophthalmol. 125 (9) (2007)
1225–1232.
[8] Age-Related Eye Disease Study Research Group, Risk factors associated with age-
related nuclear and cortical cataracts, Ophthalmol. Times 108 (8) (2001)
1400–1408.
[9] U.S. Department, Of Health and Human Services, Ofce of the Surgeon General,
The Health Consequences of Smoking: A Report of the Surgeon General,
Washington, D.C, 2004.
[10] M.J. Fowler, Microvascular and macrovascular complications of diabetes, Clin.
Diabetes 26 (2) (2008) 77–82.
[11] ETDRSR. Group, Grading diabetic retinopathy from stereoscopic color fundus
photographs—an extension of the modied Airlie House classi- cation: ETDRS
report number 10, Ophthalmol. Times 98 (5) (1991) 786–806.
[12] C.J. Rudnisky, B.J. Hinz, M.T. Tennant, A.R. de Leon, M.D. Greve, High-
resolution stereoscopic digital fundus photography versus contact lens
biomicroscopy for the detection of clinically signicant macular edema,
Ophthalmol. Times 109 (2) (2002) 267–274.
[13] Md Mohaimenul Islam, Hsuan-Chia Yang, Tahmina Nasrin Poly, Wen-Shan Jian,
Yu-Chuan (Jack) Li, Deep learning algorithms for detection of diabetic
retinopathy in retinal fundus photographs: a systematic review and meta-
analysis, Comput. Methods Progr. Biomed. 191 (2020), 105320.
[14] C. L Srinidhi, P. Aparna, J. Rajan, Recent advancements in retinal vessel
segmentation, J. Med. Syst. 41 (2017) 70, https://doi.org/10.1007/s10916-017-
0719-2.
[15] https://www.invision2020.com/retinal-imaging-history/. (Accessed 10 October
2021).
[16] Ademola Enitan Ilesanmi, Oluwagbenga Paul Idowu, Stanislav S. Makhanov,
Multiscale superpixel method for segmentation of breast ultrasound, Comput.
Biol. Med. 125 (2020), 103879.
[17] Yang Chen, Guirong Weng, An active contour model based on local pre-piecewise
tting image, Optik (2021), 168130.
[18] Daniel Lacker, Kavita Ramanan, Ruoyu Wu, Locally interacting diffusions as
Markov random elds on path space,Stochastic Processes and their Applications
140 (2021) 81–114.
[19] O. Ronneberger, P. Fischer, T. Brox, U-net: convolutional networks for biomedical
image segmentation, in: Conference on Medical Image Computing and Computer
Assisted Intervention (MICCAI), 2015.
[20] N. Siddique, S. Paheding, C.P. Elkin, V. Devabhaktuni, U-net and its variants for
medical image segmentation: a review of theory and applications, IEEE Access 9
(2021) 82031–82057, https://doi.org/10.1109/ACCESS.2021.3086020.
[21] Rohit Thanki, A deep neural network and machine learning approach for retinal
fundus image classication, Healthcare Analytics 3 (2023), 100140.
[22] M.S.B. Phridviraj, Raghuram Bhukya, Sujatha Madugula, Aakunuri Manjula,
Swathy Vodithala, Mohammed Sharfuddin Waseem, A bi-directional Long Short-
Term Memory-based Diabetic Retinopathy detection model using retinal fundus
images, Healthcare Analytics 3 (2023), 100174.
[23] Kamesh Sonti, Ravindra Dhul, A new convolution neural network model ‘‘KR-
NET’’ for retinal fundus glaucoma classication, Optik 283 (2023), 170861.
[24] V.M. Raja Sankari, U. Snekhalatha, Ashok Chandrasekaran, Prabhu Baskaran,
Automated diagnosis of Retinopathy of prematurity from retinal images of
preterm infants using hybrid deep learning techniques, Biomed. Signal Process
Control 85 (2023), 104883.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
16
[25] L. Yanhong, S. Ji, Y. Lei, B. Guibin, Y. Hongnian, ResDO-Unet, A deep residual
network for accurate retinal vessel segmentation from fundus images, Biomed.
Signal Process Control 79 (1) (2023), https://doi.org/10.1016/j.
[26] Lei Wang, Juan Gu, Yize Chen, Yuanbo Liang, Weijie Zhang, Jiantao Pu,
Hao Chen, Automated segmentation of the opt ic disc from fundus images using
an asymmetric deep learning network, Pattern Recogn. 112 (2021), 107810.
[27] Yinghua Fu, Jie Chen, Li Jiang, Dongyan Pan, Xuezheng Yue, Yiming Zhu, Optic
disc segmentation by U-net and probability bubble in abnormal fundus images,
Pattern Recogn. 117 (2021), 107971.
[28] H. Narasimhaiyer, A. Can, B. Roysam, V. Stewart, H.L. Tanenbaum, A. Ma-
jerovics, H. Singh, Robust detection and classication of longitudinal changes in
color retinal fundus images for monitoring diabetic retinopathy, IEEE Trans.
Biomed. Eng. 53 (6) (2006) 1084–1098.
[29] Xianjing Meng, Xiaoming Xi, Yang Lu, Guang Zhang, Yilong Yin, Xinjian Chen,
Fast and effective optic disk localization based on convolutional neural network,
Neurocomputing 312 (2018) 285–295.
[30] Xin Yuan, Lingxiao Zhou, Shuyang Yu, Miao Li, Xiang Wang, Xiujuan Zheng,
A multi-scale convolutional neural network with context for joint segmentation of
optic disc and cup, Artif. Intell. Med. 113 (2021), 102035.
[31] Ademola Enitan Ilesanmi, Utairat Chaumrattanakul, Stanislav S. Makhanov,
A method for segmentation of tumors in breast ultrasound images using the
variant enhanced deep learning, Biocybern. Biomed. Eng. 41 (2) (2021) 802–818.
[32] H. Fu, J. Cheng, Y. Xu, D.W.K. Wong, J. Liu, X. Cao, Joint optic disc and cup
segmentation based on multi-label deep network and polar transformation, IEEE
Trans. Med. Imag. 37 (7) (2018) 1597–1605.
[33] Lei Wang, Liu Han, Yaling Lu, Hang Chen, Jian Zhang, Jiantao Pu, A coarse-to-
ne deep learning framework for optic disc segmentation in fundus images,
Biomed. Signal Process Control 51 (2019) 82–89.
[34] Jen Hong Tan, U. Rajendra Acharya, Sulatha V. Bhandary, Chua Chua Kuang,
Segmentation of optic disc, fovea and retinal vasculature using a single
convolutional neural network, Journal of Computational Science 20 (2017)
70–79.
[35] H.N. Veena, A. Muruganandham, T. Senthil Kumaran, A novel optic disc and optic
cup segmentation technique to diagnose glaucoma using deep learning
convolutional neural network over retinal fundus images, Journal of King Saud
University – Computer and Information Sciences 34 (2022) 6187–6198.
[36] Neha Mathur, Shruti Mathur, Divya Mathur, A novel approach to improve Sobel
edge detector, Proc. Comput. Sci. 93 (2016) 431–438.
[37] Lei Zhang, Lang Zou, Chuanyu Wu, Jiangming Jia, Jianneng Chen, Method of
famous tea sprout identication and segmentation based on improved watershed
algorithm, Comput. Electron. Agric. 184 (2021), 106108.
[38] Rakhshanda Imtiaz, M. Tariq, Khan, Syed saud naqvi, muhammad arsalan, syed
junaid nawaz, screening of glaucoma disease from retinal vessel images using
semantic segmentation, Comput. Electr. Eng. 91 (2021), 107036.
[39] Zhe Xie, Tonghui Ling, Yuanyuan Yang, Rong Shu, Brent J. Liu, Optic disc and
cup image segmentation utilizing contour-based transformation and sequence
labeling networks, J. Med. Syst. 44 (2020) 96.
[40] B. Xu, N. Wang, T. Chen, M. Li, Empirical Evaluation of Rectied Activations in
Convolutional Network, 2015, pp. 1–5, arXiv: 1505.00853vol. 2.
[41] A. Viterbi, Optimum detection and signal selection for partially coherent binary
communication, IEEE Trans. Inf. Theor. 11 (2) (Apr. 1965) 239–246.
[42] A.J. Viterbi, A personal history of the Viterbi algorithm, IEEE Signal Process. Mag.
23 (4) (Jul. 2006) 120–142.
[43] Sandip Sadhukhan, Arpita Sarkar, Debprasad Sinha, Goutam Kumar Ghorai,
Gautam Sarkar, Ashis K. Dhara, Attention based fully convolutional neural
network for simultaneous detection and segmentation of optic disc in retinal
fundus images, World academy of science, Engineering and Technology
International Journal of Medical and Health Sciences 14 (8) (2020).
[44] R. Priyanka, S.J. Grace Shoba, A. Brintha Therese, Segmentation of optic disc in
fundus images using convolutional neural networks for detection of glaucoma,
International Journal of Advanced Engineering Research and Science (IJAERS) 4
(Issue 5) (2017).
[45] F. Hoppner, F. Klawonn, A contribution to convergence theory of fuzzy c-means
and derivatives, IEEE Trans. Fuzzy Syst. 11 (5) (2003) 682–694.
[46] C. Raja, N. Vinodhkumar, An efcient segmentation of optic disc using
convolution neural network for glaucoma detection in retinal images, European
Journal of Molecular & Clinical Medicine 7 (Issue 03) (2020).
[47] A.Z.M.E. Chowdhury, G. Mann, W.H. Morganb, A. Vukmirovic, A. Mehnert,
F. Sohel, Msganet-Rav, A multiscale guided attention network for artery-vein
segmentation and classication from the optic disc and retinal images, Journal of
Optometry 15 (2022) S58S69, https://doi.org/10.1016/j.optom.2022.11.001.
[48] ´
A.S. Hervella *, J. Rouco, J. Novo, M. Ortega, End-to-end multi-task learning for
simultaneous optic disc and cup segmentation and glaucoma classication in eye
fundus images, Appl. Soft Comput. 116 (2022), 108347.
[49] Indresh Kumar Gupta, Abha Choubey, Siddhartha Choubey, Mayy optimization
with deep learning enabled retinal fundus image classication model, Comput.
Electr. Eng. 102 (2022), 108176.
[50] K. Zervoudakis, S. Tsafarakis, A mayy optimization algorithm, Comput. Ind.
Eng. 145 (2020), 106559.
[51] H.N. Veena, A. Muruganandham, T. Senthil Kumaran, A novel optic disc and optic
cup segmentation technique to diagnose glaucoma using deep learning
convolutional neural network over retinal fundus images, Journal of King Saud
University – Computer and Information Sciences 34 (2022) 6187–6619.
[52] Hao Xiong, Sidong Liu, Roneel V. Sharan, Enrico Coiera, Shlomo Berkovsky,
Weak label based Bayesian U-Net for optic disc segmentation in fundus images,
Artif. Intell. Med. 126 (2022), 102261.
[53] Souvik Maiti, Debasis Maji, Ashis Kumar Dhara, Gautam Sarkar, Automatic
detection and segmentation of optic disc using a modied convolution network,
Biomed. Signal Process Control 76 (2022), 103633.
[54] Rajarshi Bhattacharya, Rukhshanda Hussain, Agniv Chatterjee, Dwipayan Paul,
Saptarshi Chatterjee, Debangshu Dey, Py-Net, Rethinking segmentation
frameworks with dense pyramidal operations for optic disc and cup segmentation
from retinal fundus images, Biomed. Signal Process Control 85 (2023), 104895.
[55] A. Septiarini, H. Hamdani, E. Setyaningsih, E. Junirianto, F. Utaminingrum,
Automatic method for optic disc segmentation using deep learning on retinal
fundus images, Healthc Inform Res 29 (2) (2023 Apr) 145–151, https://doi.org/
10.4258/hir.2023.29.2.145. Epub 2023 Apr 30. PMID: 37190738; PMCID:
PMC10209731.
[56] H. Fu, J. Cheng, Y. Xu, D. Wong, J. Liu, X. Cao, Joint optic disc and cup
segmentation based on multi-label deep network and polar transformation, IEEE
Trans. Med. Imag. 37 (7) (2018) 1597–1605.
[57] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to
document recognition, Proc. IEEE 86 (11) (1998) 2278–2324.
[58] F. Yu, V. Koltun, Multi-scale context aggregation by dilated convolutions, in: 4th
International Conference on Learning Representations, ICLR 2016 - Conference
Track Proceedings, 2016.
[59] Fantin Girarda, Conrad Kavalec, Farida Cheriet, Joint segmentation and
classication of retinal arteries/veins from fundus images, Artif. Intell. Med. 94
(2019) 96–109.
[60] Jos
′
e Morano,
′
Alvaro S. Hervella, Jorge Novo, Jos
′
e Rouco, Simultaneous
segmentation and classication of the retinal arteries and veins from color fundus
images, Artif. Intell. Med. 118 (2021), 102116.
[61] C. Wang, Z. Zhao, Y. Yu, Fine retinal vessel segmentation by combining Nest U-
net and patch-learning, Soft Comput. 25 (2021) 5519–5532, https://doi.org/
10.1007/s00500-020-05552-w.
[62] M. Kaur, A. Kamra, Detection of retinal abnormalities in fundus image using
transfer learning networks, Soft Comput. (2021), https://doi.org/10.1007/
s00500-021-06088-3.
[63] F. Girard, C. Kavalec, F. Cheriet, Joint segmentation and classication of retinal
arteries/veins from fundus images, Artif. Intell. Med. 94 (2019) 96–109, https://
doi.org/10.1016/j.artmed.2019.02.004.
[64] Ümit Budak, Zafer C¨
omert, Musa Çıbuk, Abdulkadir S¸ engür, Dccmed-Net,
Densely connected and concatenated multi Encoder-Decoder CNNs for retinal
vessel extraction from fundus images, Med. Hypotheses 134 (2020), 109426.
[65] Tanga Peng, Qiaokang Lianga, Xintong Yana, Dan Zhang, Gianmarc Coppola,
Wei Sun, Multi-proportion channel ensemble model for retinal vessel
segmentation, Comput. Biol. Med. 111 (2019), 103352.
[66] J. Odstrcilik, et al., Retinal vessel segmentation by improved matched ltering:
evaluation on a new high-resolution fundus image database, IET Image Process. 7
(4) (2013) 373–383.
[67] Zhao He, Yun Sun, Huiqi Li, Retinal vascular junction detection and classication
via deep neural networks, Comput. Methods Progr. Biomed. 183 (2020), 105096.
[68] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
2016, pp. 770–778.
[69] Ruohan Zhao, Li Qin, Jianrong Wu, Jane You, A nested U-shape network with
multi-scale upsample attention for robust retinal vascular segmentation, Pattern
Recogn. 120 (2021), 107998.
[70] T.-Y. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid
networks for object detection, in: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2017, pp. 2117–2125.
[71] Maciej Czepita, Anna Fabija ´nska, Image processing pipeline for the detection of
blood ow through retinal vessels with subpixel accuracy in fundus images,
Comput. Methods Progr. Biomed. 208 (2021), 106240.
[72] S. Klein, Staring, M.K. Murphy, M.A. Viergever, J.P.W. Pluim, Elastix: a toolbox
for intensity based medical image registration, IEEE Trans. Med. Imag. 29 (1)
(2010) 196–205, https://doi.org/10.1109/TMI.2009.2035616.
[73] L.D. Hubbard, R.J. Brothers, W.N. King, et al., Methods for evaluation of retinal
microvascular abnormalities associated with hypertension/sclerosis in the
atherosclerosis risk in communities study, Ophthalmol. Times 106 (12) (1999)
2269–2280 Dec, https://doi.org/10.1016/s0161-6420(99)90525-0.
[74] Gang Sun, Xiaoyan Liu, Xuefei Yu, Multi-path cascaded U-net for vessel
segmentation from fundus uorescein angiography sequential images, Comput.
Methods Progr. Biomed. 211 (2021), 106422.
[75] Michael Yeung, Evis Sala, Carola-Bibiane Schonlieb, Leonardo Rundo, Focus U-
Net, A novel dual attention-gated CNN for polyp segmentation during
colonoscopy, Comput. Biol. Med. 137 (2021), 104815.
[76] Lei Sang, Min Xu, Shengsheng Qian, Xindong Wu, Knowledge graph enhanced
neural collaborative ltering with residual recurrent network, Neurocomputing
454 (2021) 417–429.
[77] J.V. Soares, J.J. Leandro, R.M. Cesar, H.F. Jelinek, M.J. Cree, Retinal vessel
segmentation using the 2-D Gabor wavelet and supervised classication, IEEE
Trans. Med. Imag. 25 (9) (2006) 1214–1222.
[78] Song Guo, Kai Wang, Hong Kang, Yujun Zhang, Yingqi Gao, Tao Li, Bts-Dsn,
Deeply supervised neural network with short connections for retinal vessel
segmentation, Int. J. Med. Inf. 126 (2019) 105–113.
[79] Yanhui Guo, Ümit Budak, Abdulkadir ¸S engür, A novel retinal vessel detection
approach based on multiple deep convolution neural networks, Comput. Methods
Progr. Biomed. 167 (2018) 43–48.
[80] Zhongshuo Zhuo, Jianping Huang, Ke Lu, Daru Pan, A size-invariant
convolutional network with dense connectivity applied to retinal vessel
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
17
segmentation measured by a unique index, Comput. Methods Progr. Biomed. 196
(2020), 105508.
[81] Y. Luo, H. Cheng, L. Yang, Size-invariant fully convolutional neural network for
vessel segmentation of digital retinal images, in: Signal and Information
Processing Association Annual Summit and Conference (APSIPA), 2016 Asia-
Pacic, IEEE, 2016, pp. 1–7.
[82] G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected
convolutional networks, CVPR 1 (2017) 3.
[83] ´
Alvaro S. Hervella, Jorge Novo Jos´
eRouco, ManuelG. Penedo, Marcos Ortega,
Deep multi-instance heatmap regression for the detection of retinal vessel
crossings and bifurcations in eye fundus images, Comput. Methods Progr.
Biomed. 186 (2020), 105201.
[84] S. Pearl Mary, V. Thanikaiselvan, Unied adaptive framework for contrast
enhancement of blood vessels, Int. J. Electr. Comput. Eng. 10 (2020) 767–777,
https://doi.org/10.11591/ijece.v10i1.pp767-777.
[85] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale
image recognition, in: 3rd Int. Conf. Learn. Represent. ICLR 2015 - Conf, Track
Proc., 2015.
[86] Di Li, Susanto Rahardja, BseresU-Net, An attention-based before-activation
residual U-Net for retinal vessel segmentation, Comput. Methods Progr. Biomed.
205 (2021), 106070.
[87] Zefang Lin, Jianping Huang, Yingyin Chen, Xiao Zhang, Wei Zhao, Yong Li,
Ligong Lu, Meixiao Zhan, Xiaofei Jiang, Xiong Liang, A high resolution
representation network with multi-path scale for retinal vessel segmentation,
Comput. Methods Progr. Biomed. 208 (2021), 106206.
[88] C. Meng, K. Sun, S. Guan, Q. Wang, R. Zong, L. Liu, Multiscale dense
convolutional neural network for DSA cerebrovascular segmentation,
Neurocomputing 373 (2020) 123–134.
[89] Chun Tian, Tao Fang, Yingle Fan, Wei Wu, Multi-path convolutional neural
network in fundus segmentation of blood vessels, biocybernetic and biomedical
engineering 40 (2020) 583–595.
[90] Ibrahim Atli, Osman Serdar Gedik, Sine-Net: a fully convolutional deep learning
architecture for retinal blood vessel segmentation, Engineering Science and
Technology, Int. J. 24 (2021) 271–283.
[91] X. Bai, F. Zhou, B. Xue, Image enhancement using multi-scale image features
extracted by top-hat transform, Opt Laser. Technol. 44 (2) (2012) 328–336.
[92] Yanhui Guo, Ümit Budak, J. Lucas, Vespa, Elham Khorasani, Abdulkadir S
¸engür,
A retinal vessel detection approach using convolution neural network with
reinforcement sample learning strategy, Measurement 125 (2018) 586–591.
[93] L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A.L. Yuille, Deeplab: semantic
image segmentation with deep convolutional nets, atrous convolution, and fully
connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (4) (2018) 834–848,
https://doi.org/10.1109/tpami.2017.2699184.
[94] Yicheng Wu, Yong Xia, Yang Song, Yanning Zhang, Weidong Cai, NFN+: a novel
network followed network for retinal vessel segmentation, Neural Network. 126
(2020) 153–162.
[95] Ruben Hemelings, Bart Elend, Ingeborg Stalmans, Karel Van Keer, Patrick De
Boever, Matthew B. Blaschko, Artery–vein segmentation in fundus images using a
fully convolutional network, Comput. Med. Imag. Graph. 76 (2019), 101636.
[96] J.V.B. Soares, J.J.G. Leandro, R.M. Cesar, H.F. Jelinek, M.J. Cree, Retinal vessel
segmentation using the 2-d gabor wavelet and supervised classication,
IEEETrans. Med. Imaging 25 (2006) 1214–1222.
[97] Henda Boudegga, Yaroub Elloumi, Mohamed Akil, Mohamed Hedi Bedoui,
Rostom Kachouri, Asma Ben Abdallah, Fast and efcient retinal blood vessel
segmentation method based on deep learning network, Comput. Med. Imag.
Graph. 90 (2021), 101902.
[98] Z. Jiang, H. Zhang, Y. Wang, S.-B. Ko, Retinal blood vessel segmentation using
fully convolutional network with transfer learning, Comput. Med. Imag. Graph.
68 (September) (2018) 1–15, https://doi.org/10.1016/j.
compmedimag.2018.04.005.
[99] P. Liskowski, K. Krawiec, Segmenting retinal blood vessels with deep neural
networks, IEEE Trans. Med. Imag. 35 (11) (2016) 2369–2380, https://doi.org/
10.1109/TMI.2016.2546227. November.
[100] Shuangling Wang, Yilong Yin, Guibao Cao, Benzheng Wei, Yuanjie Zheng,
Hierarchical retinal blood vessel segmentation based on feature and ensemble
learning, Neurocomputing 149 (2015) 708–717.
[101] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32.
[102] Kai Hu, Zhenzhen Zhang, Xiaorui Niu, Yuan Zhang, Chunhong Cao, Fen Xiao,
Xieping Gao, Retinal vessel segmentation of color fundus images using multiscale
convolutional neural network with an improved cross-entropy loss function,
Neurocomputing 309 (2018) 179–191.
[103] Yukun Zhou, Zailiang Chen, Hailan Shen, Xianxian Zheng, Rongchang Zhao,
Xuanchu Duan, A rened equilibrium generative adversarial network for retinal
vessel segmentation, Neurocomputing 437 (2021) 118–130.
[104] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, et al.,
Attention is all you need, in: International Conference on Learning
Representations, Long Beach, California, USA, 2017.
[105] Lei Yang, Huaixin Wang, Qingshan Zeng, Yanhong Liu, Guibin Bian, A hybrid
deep segmentation network for fundus vessels via deep-learning framework,
Neurocomputing 448 (2021) 168–178.
[106] Qiangguo Jin, Zhaopeng Meng, Tuan D. Pham, Qi Chen, Leyi Wei, Ran Su, Dunet,
A deformable network for retinal vessel segmentation, Knowl. Base Syst. 178
(2019) 149–162.
[107] Touque Ahmed Soomro, Ahmed J. Ab, Junbin Gao, Olaf Hellwich,
Lihong Zheng, Manoranjan Paul, Strided fully convolutional neural network for
boosting the sensitivity of retinal blood vessels segmentation, Expert Syst. Appl.
134 (2019) 36–52.
[108] T.A. Soomro, A.J. A, J. Gao, O. Hellwich, M.U. Khan, M. Paul, L. Zheng,
Boosting sensitivity of a retinal vessel segmentation algorithm with convolutional
neural network, in: International Conference on Digital Image Computing:
Techniques and Applications (DICTA), 2017.
[109] Chala Mohamed, Benayad Nsiri, My Hachem El yous Alaoui,
Abdelmajid Soulaymani, Abdelrhani Mokhtari, Brahim Benaji Chandler, An
automatic retinal vessel segmentation approach based on Convolutional Neural
Networks, Expert Syst. Appl. 184 (2021), 115459.
[110] Muyi Sun, Kaiqi Li, Xingqun Qi, Hao Dang, Guanhong Zhang, Contextual
information enhanced convolutional neural networks for retinal vessel
segmentation in color fundus images, J. Vis. Commun. Image Represent. 77
(2021), 103134.
[111] Huisi Wu, Wei Wang, Jiafu Zhong, Baiying Lei, Zhenkun Wen, Jing Qin, Scs-Net,
A scale and context sensitive network for retinal vessel segmentation, Med. Image
Anal. 70 (2021), 102025.
[112] V. Sathananthavathi, G. Indumathi, Encoder enhanced atrous (EEA) unet
architecture for retinal blood vessel segmentation, Cognit. Syst. Res. 67 (2021)
84–95.
[113] Y. Fisher, V. Koltun, Multi-scale context aggregation by dilated convolutions, in:
Proceeding of Computer Vision and Pattern Recognition ICLR 2016, 2016 arXiv:
1511.07122v3 [cs.CV] 30 Apr 2016.
[114] Pengshuai Yin, Hongmin Cai, Qingyao Wu, Df-Net, Deep fusion network for multi-
source vessel segmentation, Inf. Fusion 78 (2022) 199–208.
[115] H. Zhao, J. Shi, X. Qi, X. Wang, J. Jia, Pyramid scene parsing network, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
2017, pp. 2881–2890.
[116] Xianlun Tang, Bing Zhong, Jiangping Peng, Bohui Hao, Jie Li, Multi-scale channel
importance sorting and spatial attention mechanism for retinal vessels
segmentation, Applied Soft Computing Journal 93 (2020), 106353.
[117] Aamer Abdul Rahman, Birendra Biswal, P. Geetha Pavani, Shazia Hasan, M.V.
S. Sairam, Robust segmentation of vascular network using deeply cascaded AReN-
UNet, Biomed. Signal Process Control 69 (2021), 102953.
[118] R. Collobert, S. Bengio, Links between perceptrons, MLPs and SVMs, Proc. Int’l
Conf. on Machine Learning (ICML) (2004).
[119] Zhengjin Shi, Tianyu Wang, Zheng Huang, Feng Xie, Zihong Liu, Bolun Wang,
Jing Xu, Md-Net, A multi-scale dense network for retinal vessel segmentation,
Biomed. Signal Process Control 70 (2021), 102977.
[120] Beaudelaire Saha Tchinda, Daniel Tchiotsop, Michel Noubom, Valerie Louis-Dorr,
Didier Wolf, Retinal blood vessels segmentation using classical edge detection
lters and the neural network, Inform. Med. Unlocked 23 (2021), 100521.
[121] J.V. Soares, J.J. Leandro, R.M. Cesar, H.F. Jelinek, M.J. Cree, Retinal vessel
segmentation using the 2-d gabor wavelet and supervised classication, IEEE
Trans. Med. Imag. 25 (9) (2006) 1214–1222.
[122] D. Ziou, S. Tabbone, Edge detection techniques- an overview, International
Journal of Pattern Recognition and Image Analysis 8 (4) (1998) 537–559.
[123] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal.
Mach. Intell. 8 (1986) 679–714.
[124] R.C. Gonzalez, R.E. Woods, S.L. Eddins, Digital Image Processing Using MATLAB,
Pearson/Prentice Hall, Upper Saddle River, New Jersey, 2004.
[125] Manuel E. Gegundez-Arias, Diego Marin-Santos, Manuel J. Isaac Perez-Borrero,
Vasallo-Vazquez, new deep learning method for blood vessel segmentation in
retinal images based on convolutional kernels and modied U-Net model,
Comput. Methods Progr. Biomed. 205 (2021), 106081.
[126] Debasis Maji, Arif Ahmed Sekh, Automatic grading of retinal blood vessel in deep
retinal image diagnosis, J. Med. Syst. 44 (2020) 180.
[127] S.N. Sangeethaa, P. Uma Maheswari, An intelligent model for blood vessel
segmentation in diagnosing DR using CNN, J. Med. Syst. 42 (2018) 175.
[128] Revathi Priya Muthusamy, S. Vinod, M. Tholkapiyan, Automatic detection of
abnormalities in retinal blood vessels using DTCWT, GLCM feature extractor and
CNN-rnn classier, Int. J. Recent Technol. Eng. 8 (Issue 1S4) (2019).
[129] Retinal vessel segmentation via A coarse-to-ne convolutional neural network, in:
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2018.
[130] C. Shi, C. Xu, J. He, Y. Chen, Y. Cheng, Q. Yang, H. Qiu, Graph-based convolution
feature aggregation for retinal vessel segmentation, Simulat. Model. Pract. Theor.
121 (2022), 102653.
[131] Song Guo, Csgnet, Cascade semantic guided net for retinal vessel segmentation,
Biomed. Signal Process Control 78 (2022), 103930.
[132] Kan Ren, Longdan Chang, Minjie Wan, Guohua Gu, Qian Chen, An improved U-
net based retinal vessel image segmentation method, Heliyon 8 (2022), e11187.
[133] M. Tan, R. Pang, Q.V. Le, EfcientDet: Scalable and Efcient Object Detection,
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),
2020, pp. 10778–10787.
[134] J. Han, Y. Wang, H. Gong, Fundus Retinal Vessels Image Segmentation Method
Based on Improved U-Net, https://doi.org/10.1016/j.irbm.2022.03.001.
[135] Yanan Xu, Yingle Fan, Dual-channel asymmetric convolutional neural network for
an efcient retinal blood vessel segmentation in eye fundus images, Biocybern.
Biomed. Eng. 42 (2022) 695–706.
[136] Zhong Xiang, Hongbin Zhang, Guangli Li, Donghong Ji, Do you need sharpened
details? Asking MMDC-Net: multi-layer multi-scale dilated convolution network
for retinal vessel segmentation, Comput. Biol. Med. 150 (2022), 106198.
[137] Dora E. Alvarado-Carrillo, Oscar S. Dalmau-Cede˜
no, Width attention based
convolutional neural network for retinal vessel segmentation, Expert Syst. Appl.
209 (2022), 118313.
A.E. Ilesanmi et al.
Healthcare Analytics 4 (2023) 100261
18
[138] Deng Xiangyu, Ye Jinhong, A retinal blood vessel segmentation based on
improved D-MNet and pulse-coupled neural network, Biomed. Signal Process
Control 73 (2022), 103467.
[139] Zheng Huang, Ming Sun, Yuxin Liu, Jiajun Wu, Csaunet, A cascade self-attention
u-shaped network for precise fundus vessel segmentation, Biomed. Signal Process
Control 75 (2022), 103613.
[140] Geetha Pavani, Birendra Biswal, Tapan Kumar Gandhi, Multistage Dpiref-Net, An
effective network for semantic segmentation of arteries and veins from retinal
surface, Neuroscience Informatics 2 (2022), 100074.
[141] Robert Arnar Karlsson, Sveinn Hakon Hardarson, Artery vein classication in
fundus images using serially connected U-Nets, Comput. Methods Progr. Biomed.
216 (2022), 106650.
[142] D.A. Clevert, T. Unterthiner, S. Hochreiter, Fast and accurate deep network
learning by exponential linear units (ELUs), 4th International Conference on
Learning Representations, ICLR 2016 - Conference Track Proceedings (2016)
1–14.
[143] Zifan Zhu, An Qing, Zhicheng Wang, Li Qian, Hao Fang, Zhenghua Huang, Ilu-
Net, Inception-Like U-Net for retinal vessel segmentation, Optik - International
Journal for Light and Electron Optics 260 (2022), 169012.
[144] Pengshuai Yin, Hongmin Cai, Qingyao Wu, Df-Net, Deep fusion network for multi-
source vessel segmentation, Inf. Fusion 78 (2022) 199–208.
[145] Tariq M. Khan, Syed S. Naqvi, Robles-Kelly Antonio, Imran Razzak, Retinal vessel
segmentation via a Multi-resolution Contextual Network and adversarial learning,
Neural Network. 165 (2023) 310–320.
[146] Renyuan Liu, Tong Wang, Xuejie Zhang, Xiaobing Zhou, DA-Res2UNet:
Explainable blood vessel segmentation from fundus images, Alex. Eng. J. 68
(2023) 539–549.
[147] Meilin Liu, Zidong Wang, Li Han, Peishu Wu, Fuad E. Alsaadi, Nianyin Zeng, Aa-
Wgan, Attention augmented Wasserstein generative adversarial network with
application to fundus retinal vessel segmentation, Comput. Biol. Med. 158 (2023),
106874.
[148] K. He, G. Gkioxari, P. Doll´
ar, R. Girshick, Mask r-cnn, in: Proceedings of the IEEE
International Conference on Computer Vision, 2017, pp. 2961–2969.
[149] P. Liskowski, K. Krawiec, Segmenting retinal blood vessels with deep neural
networks, IEEE Trans. Med. Imag. 35 (2016) 2369–2380, https://doi.org/
10.1109/TMI.2016.2546227.
[150] Kyoung Jin Noh, Sang Jun Park, Soochahn Lee, Scale-space approximated
convolutional neural networks for retinal vessel segmentation, Comput. Methods
Progr. Biomed. 178 (2019) 237–246.
[151] K. Maninis, J. Pont-Tuset, P.A. Arbel´
aez, L.V. Gool, Deep retinal image
understanding, in: Proc. Of Medical Image Computing and Computer Assisted
Intervention, MICCAI), 2016, pp. 140–148, https://doi.org/10.1007/978-3-319-
46723-8_17.
[152] Pearl Mary Samuel, Thanikaiselvan Veeramalai, VSSC Net, Vessel specic skip
chain convolutional network for blood vessel segmentation, Comput. Methods
Progr. Biomed. 198 (2021), 105769.
[153] K. Sun, B. Xiao, D. Liu, J. Wang, Deep High-Resolution Representation Learning
for Human Pose Estimation, 2019, 09212 arXiv preprint arXiv:1902.
[154] A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classication with deep
convolutional neural networks, Commun. ACM 60 (6) (2017) 84–90, https://doi.
org/10.1145/3065386. May.
[155] Y. Liu, M.-M. Cheng, X. Hu, K. Wang, X. Bai, Richer convolutional features for
edge detection, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2017, pp. 5872–5881.
[156] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, et
al., Generative adversarial nets, in: Neural Information Processing Systems
Conference, Canada, Montreal, Quebec, 2014.
[157] J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, Y. Wei, Deformable convolutional
networks, in: International Conference on Computer Vision, 2017, pp. 764–773.
[158] Vijay Kendall Badrinarayanan, Cipolla Alex, Roberto, Segnet: a deep
convolutional encoder-decoder architecture for image segmentation, IEEE Trans.
Pattern Anal. Mach. Intell. 99 (2017), 1–1.
[159] H. Zhao, J. Shi, X. Qi, X. Wang, J. Jia, Pyramid scene parsing network, in: Proc
IEEE Comput Soc Conf Comput Vision Pattern Recognit, 2017, pp. 2881–2890.
Venice Italy.
[160] J. Staal, M.D. Abr`
amoff, M. Niemeijer, M.A. Viergever, B.V. Ginneken, Ridge-
based vessel segmentation in color images of the retina, IEEE Trans. Med. Imag.
23 (4) (2004) 501–509.
[161] A. Hoover, M. Goldbaum, Locating the optic nerve in a retinal image using the
fuzzy convergence of the blood vessels, IEEE Trans. Med. Imag. 22 (8) (2003)
951–958.
[162] Q. Li, B. Feng, L.P. Xie, P. Liang, H. Zhang, T. Wang, A cross-modality learning
approach for vessel segmentation in retinal images, IEEE Trans. Med. Imag. 35
(2016) 109–118.
[163] J. Orlando, E. Prokofyeva, M. Blaschko, A discriminatively trained fully
connected conditional random eld model for blood vessel segmentation in
fundus images, IEEE Trans. Biomed. Eng. 64 (2017) 16–27.
[164] L. Wang, H. Liu, Y. Lu, H. Chen, J. Zhang, J. Pu, A coarse-to-ne deep learn- ing
framework for optic disc segmentation in fundus images, Biomed. Signal Process
Control 51 (2019) 82–89.
[165] S. Abbasi-Sureshjani, I. Smit-Ockeloen, E. Bekkers, B. Dashtbozorg, B. ter Haar
Romeny, Automatic detection of vascular bifurcations and crossings in retinal
images using orientation scores, in: International Symposium on Biomedical
Imaging (ISBI), IEEE, 2016, pp. 189–192.
[166] J. Orlando, H. Fu, J. Breda, K. Keer, D. Bathula, A. Pinto, R. Fang, P. Heng, J. Kim,
J. Lee, J. Lee, X. Li, P. Liu, S. Lu, B. Murugesan, V. Naranjo, S. Phaye, S. Shankara-
narayana, A. Sikka, J. Son, A. Hengel, S. Wang, J. Wu, Z. Wu, G. Xu, Y. Xu, P. Yin,
F. Li, X. Zhang, Y. Xu, H. Bogunovic, REFUGE Challenge: a unied framework for
evaluating automated methods for glaucoma assessment from fundus
photographs, Med. Image Anal. 59 (2020), 101570.
[167] R.V.J.P.H. K¨
alvi¨
ainen , H. Uusitalo , Diaretdb1 diabetic retinopathy database and
evaluation protocol, Med. Image Underst. Anal. 20 07 (20 07) 61.
[168] T. Kauppi, V. Kalesnykiene, J.K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo,
H. K¨
alvi¨
ainen, J. Pietil¨
a, Diaretdb0: Evaluation Database and Methodology for
Diabetic Retinopathy Algorithms, Machine Vision and Pattern Recognition Re-
Search Group vol. 73, Lappeenranta University of Technology, Finland, 2006.
[169] J. Sivaswamy, S. Krishnadas, A. Chakravarty, G. Joshi, A.S. Tabish,
A comprehensive retinal image dataset for the assessment of glaucoma from the
optic nerve head analysis, JSM Biomed Imaging Data Papers 2 (1) (2015) 1004.
[170] R. Wang, L. Zheng, C. Xiong, C. Qiu, H. Li, X. Hou, B. Sheng, P. Li, Q. Wu, Retinal
optic disc localization using convergence tracking of blood vessels, Multimed.
Tool. Appl. 76 (22) (2016) 1–23.
[171] F. Fumero, S. Alay
′
on, J.L. Sanchez, J. Sigut, M. Gonzalez-Hernandez, Rim-one: an
open retinal image database for optic nerve evaluation, in: 24th International
Symposium on Computer-Based Medical Systems, CBMS), 2011, pp. 1–6, 2011.
[172] M. Abdullah, M. Fraz, S. Barman, Localization and Segmentation of Optic Disc in
Retinal Images Using Circular Hough Transform and Grow-Cut Algorithm, PeerJ,
2016.
[173] H. Kang, Y. Gao, S. Guo, X. Xu, T. Li, K. Wang, AVNet: a retinal artery/vein
classication network with category-attention weighted fusion, Comput. Methods
Progr. Biomed. 195 (2020), 105629.
[174] R. Estrada, C. Tomasi, S.C. Schmidler, S. Farsiu, Tree topology estimation, IEEE
Trans. Pattern Anal. Mach. Intell. 37 (8) (2015) 1688–1701.
[175] H. Zhao, H. Li, S. Maurerstroh, L. Cheng, Synthesizing retinal and neuronal
images with generative adversarial nets, Med. Image Anal. 49 (2018) 14–26.
[176] J.I. Orlando, J.B. Breda, K. Van Keer, M.B. Blaschko, P.J. Blanco, C.A. Bulant,
Towards a glaucoma risk index based on simulated hemodynamics from fundus
images, in: International Conference on Medical Image Computing and Computer-
Assisted Intervention, Springer, 2018, pp. 65–73.
[177] A.A. Almazroa, S. Alodhayb, E. Osman, E. Ramadan, M. Hummadi, M. Dlaim,
M. Alkatee, K. Raahemifar, V. Lakshminarayanan, Retinal Fundus Images for
Glaucoma Analysis: the RIGA Dataset, 2018.
[178] L.M. Koesten, E. Kacprzak, J.F.A. Tennison, E. Simperl, The trials and tribulations
of working with structured data: -a study on information seeking behavior, in:
Proceedings of the 2017 CHI Conference on Human Factors in Computing
Systems, ACM, New York, NY, USA, 2017, pp. 1277–1289, https://doi.org/
10.1145/3025453.3025838.
[179] Tao Li, Bo Wang, Chunyu Hu, Hong Kang, Hanruo Liu, Kai Wang, Huazhu Fu,
Applications of deep learning in fundus images: a review, Med. Image Anal. 69
(2021), 101971.
[180] Sourya Sengupta, Amitojdeep Singh, Henry A. Leopold, Tanmay Gulati,
Vasudevan Lakshminarayanan, Ophthalmic diagnosis using deep learning with
fundus images – a critical review, Artif. Intell. Med. 102 (2020), 101758.
[181] A.M. David, Y. Lou, E. Ali, C. Warren, A. Ryan, J.C. Folk, N. Meindert, Improved
automated detection of diabetic retinopathy on a publicly available dataset
through integration of deep learning, Invest. Ophthalmol. Vis. Sci. 57 (13) (2016)
5200.
[182] Jayaram K. Udupa, Dewey Odhner, Liming Zhao, Yubing Tong, M. Monica,
S. Matsumoto, Krzysztof C. Ciesielski, Alexandre X. Falcao,
Pavithra Vaideeswaran, Victoria Ciesielski, Babak Saboury,
Syedmehrdad Mohammadianrasanani, Sanghun Sin, Raanan Arens, Drew
A. Torigian, Body-wide hierarchical fuzzy modeling, recognition, and delineation
of anatomy in medical images, Med. Image Anal. 18 (2014) 752–771.
A.E. Ilesanmi et al.
