Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Recent Developments in Detection of Central

Serous Retinopathy through Imaging and
Artificial Intelligence Techniques – A Review
Syed Ale Hassan1, Shahzad Akbar1, Amjad Rehman2, Tanzila Saba2, Hoshang Kolivand3,
Saeed Ali Bahaj4
1
Riphah College of Computing, Riphah International University, Faisalabad Campus, Pakistan
2
Artificial Intelligence & Data Analytics Lab (AIDA), Prince Sultan University, Riyadh, Saudi Arabia.
3
School of Computer Science and Mathematics, Liverpool John Moores University, Liverpool, UK
3
School of Computing and Digital Technologies, Staffordshire University, Staffordshire, UK
4
MIS Department College of Business Administration, Prince Sattam bin Abdulaziz University, Alkharj, Saudi Arabia.
Corresponding authors: Shahzad Akbar ([email protected]) and Amjad Rehman ([email protected])

ABSTRACT Central Serous Retinopathy (CSR) or Central Serous Chorioretinopathy (CSC) is a significant
disease that causes blindness and vision loss among millions of people worldwide. It transpires as a result of
accumulation of watery fluids behind the retina. Therefore, detection of CSR at early stages allows preventive
measures to avert any impairment to the human eye. Traditionally, several manual methods for detecting CSR
have been developed in the past; however, they have shown to be imprecise and unreliable. Consequently,
Artificial Intelligence (AI) services in the medical field, including automated CSR detection, are now possible
to detect and cure this disease. This review assessed a variety of innovative technologies and researches that
contribute to the automatic detection of CSR. In this review, various CSR disease detection techniques,
broadly classified into two categories: a) CSR detection based on classical imaging technologies, and b) CSR
detection based on Machine/Deep Learning methods, have been reviewed after an elaborated evaluation of
29 different relevant articles. Additionally, it also goes over the advantages, drawbacks and limitations of a
variety of traditional imaging techniques, such as Optical Coherence Tomography Angiography (OCTA),
Fundus Imaging and more recent approaches that utilize Artificial Intelligence techniques. Finally, it is
concluded that the most recent Deep Learning (DL) classifiers deliver accurate, fast, and reliable CSR
detection. However, more research needs to be conducted on publicly available datasets to improve
computation complexity for the reliable detection and diagnosis of CSR disease.

INDEX TERMS Central Serous Retinopathy, Deep Learning, Fundus Images, Machine Learning, Optical
Coherence Tomography Images

I. INTRODUCTION Since any damage to the retina may have severe

The retina is located behind the eyeball near the optic nerve ramifications to our ocular abilities. A typical schematic
and comprises a thin layer of tissue [1]. It obtains the focused diagram of the human eye has been depicted in Figure 1.
light from the eye-lens, converts it into neural signals, and
imparts signs to the brain for visual recognition. The retina
processes light using a layer of photoreceptor cells. These are
light-sensitive cells responsible for detecting visual
characteristics, such as color and light intensity.
Subsequently, the data accumulated by the photoreceptor
cells are sent to the brain through the optic nerve for optical
recognition. Therefore, the retina plays a crucial role in
image processing for the human brain recognizes and
FIGURE 1: Anatomy of the Human Eye, depicting various parts
distinguishes various surrounding objects and names them. including the Retina [2]

VOLUME XX, 2017 1

 Author Name: Preparation of Papers for IEEE Access (February 2017)

In this review, various retinal disease diagnosis and leading to complete recovery. The statistics have shown that
detection techniques have been described, and then focused CSR usually affects one eye, but the damage to both eyes
on utilizing advanced imaging and Artificial Intelligence cannot be ruled out during the lifetime of any patient.
(AI) methods in a specific retinal disorder known as the Besides, in some cases, the patients may recover after a
Central Serous Retinopathy (CSR). while, without any treatment [31]. Moreover, CSR usually is
The retinal disease diagnosis and detection can be not correlated with affecting patient's quality of life [32].
performed through several imaging techniques such as According to many studies, the CSR patients’ ages range
fundus photography and Auto-Fluorescence, Optical between 7 to 83 years, with the most affected age group of
Coherence Tomography (OCT), and Angiography. Recently, 40-50 years.
in advanced Deep Learning (DL) investigations, every The interpretation of CSR in classical terms is referred to
patient is potentially a source of precious diagnostic data that as acute CSR. A patient with intense CSR may encounter
can be utilized in training new Machine Learning (ML) obscured vision, decrease in contrast sensitivity and shading
models for better treatment and diagnosis [3]–[5]. The most vision, metamorphopsia, and minor hyperopic move.
recognized use of artificial intelligence (AI) strategies in Traditionally there is a central serous separation of focal
retinal disease are the development of spots intricate to retina, sometimes with dull yellow stores and in a few cases
disease characteristics on color fundus photos. In this with serous RPE separation. In several cases of CSR, a
context, programmed segmentation and characterization of permanent sub-retinal fluid accumulates for three months or
anatomic and pathologic have been combined and applied in more: resulting in permanent visual symptoms. Such cases
retinal diseases [6], [7]. These techniques are effectively of CSR frequently experience a fluctuating degree of sub-
being utilized due to the emergence and popularity of the retinal fluid. However, most cases recover automatically,
OCT methodology since the early 2000s. whereas some patients may experience chronic CSR [33].
Consequently, it is now possible to conduct exact Nevertheless, their visual acuity remains typically steady.
detection of several features to perform automated The chronicity of these patients depends upon the time
segmentation of retinal diseases such as, sub-retinal and duration of CSR, and it usually takes 3 to 6 months in case
intraretinal fluid, drusen, color epithelial disintegration, and of acute CSR. Contrarily, in chronic CSR patients, the
topographical atrophy with equivalent exactness as human symptoms of morphological changes and an increased risk of
graders. Another research field is developing a profound- CNV have been observed [33].
settled image of the retinal vasculature, where OCTA Patients with CSC are generally of the age of 25 and 50, in
generates multiple consecutively generated images using which men are afflicted far more frequently than women.
OCT-B scans. In recent times some automated systems have These patients normally have symptoms such as the
been designed in the detection and treatment of retinal grumblings of unexpected beginning, contortion, and focal
disease mentioned as Diabetic Retinopathy [8]–[14], vision blurring. Visual acuity ranges between 6/5 and 6/60,
Hypertensive Retinopathy [15]–[18], Age-related Macular but its usual range is 6/9 to 6/12 [34]. In the case of CSR, it
Degeneration [19]–[23] and Glaucoma [24]–[29]. is generally a self-constraining disease with unconstrained
The automatic diagnosis of retinal disorders through the resolution having boundaries of 3–4 months. Historically
analysis of retinal images has become a significant practice data reveals that nearly half of the CSR patients may
in clinical systems. Using these automated techniques, the experience recurrences of the disease within a year, causing
physicians have extra optimized and accurate results. In the the patient to undergo various treatment procedures, which
past, manual methods were used for retinal disease detection, may last for three months in chronic CSR, recurrent CSR,
which were inefficient, time-consuming, and inaccurate. and first-time CSR patients. Some standard CSR treatments
Contrarily, computer-aided retinal disease recognition include the Micro Pulse Laser Treatment (MPLT), the
systems are cost-effective, objective, accurate, user-friendly, Transpupillary Thermo treatment (TTT), the Photodynamic
and fast. In addition, they do not largely rely on the ability of therapy (PDT), and the Intravitreal anti-Vascular Endothelial
an experienced ophthalmologist to examine the various Growth Factor (anti-VEGF). These treatment methods are
scanning images to detect the disease [29]. This review based on the following points:
article focuses on a specific retinal disease that is commonly i. A majority of the population automatically
known as Central Serous Retinopathy (CSR) or Central recuperates within 4 to 6 months without requiring
Serous Chorioretinopathy (CSC). It was first reported by any specific medicaments.
Von Graefe as "Relapsing central Lueitic Retinitits" [30]. ii. If a patient’s CSR lasts for a year, then some
The CSR is among one of the major eye diseases, which is treatment may be required.
caused due to the collection of fluids behind the retina, which iii. In rare cases, if CSR lasts for more than a year, an
can severely damage eyesight due to the presence of a ophthalmologist may opt for specific treatments such
delicate tissue layer constituting retina. Therefore, early- as RPE detachment or bullous retinal detachment
stage detection can essentially allow for taking the [35].
appropriate preventive measures to restore vision, thereby

2 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

There have been approximately 40 articles published fluorescein angiography. , The advancement of recent
related to deep learning and medical imaging. Among these scanning strategies, has transformed the comprehension
40 articles, only 29 articles are associated with Central regarding the pathophysiology of CSC, and thus the
Serous Retinopathy (CSR). These articles, along with their treatment has thoroughly changed. This review paper
methodologies, advantages, and limitations, are discussed in explains existing comprehension about physiopathology and
this review. As mentioned above, CSR detection is hazardous components and multimodal imaging-based
performed through various imaging technologies. We briefly highlights of CSR.
summarize them below. 2) FLUORESCEIN ANGIOGRAPHY (FA)
Fluorescein Angiography (FA) is an effective imaging
A. Imaging Technologies for Detection of CSR technique in which fluorescent dye is injected into blood
The traditional imaging technologies for CSR detection vessels of patient’s eyes in order to capture their clear
include Fundus Photography, Fluorescein Angiography images. The main objective of this technique is to highlight
(FA), and Optical Coherence Tomography (OCT). These the blood vessels to form a clear and visible image. The
technologies are discussed in the following subsections. patient is normally prescribed with primary care before
1) THE FUNDUS PHOTOGRAPHY initiating the FA procedure to ensure a satisfactory blood
Fundus Photography is a process of obtaining the retinal red- stream in the veins. In addition, the physicians propose
free image and is considered an alternative to OCT imaging. primary care to analyze further the eye issues, including
This technique is based on the statistical approach and macular degeneration, diabetic retinopathy, or Central
requires advancement in contrast to its forerunner, color Serous Retinopathy (CSR). The retinal images before and
photography film. Similarly, the digital image of retina after fluorescein angiograms are depicted in Figures 4 and 5,
provides quick towering-resolution and consistent image, respectively [36].
and it is accessible instantly and manageable for the
development of an image. Moreover, Fundus photography is
regularly employed for ailment records and clinical
examination, along with potential usage for tolerant training
and telehealth. Additionally, the images generated through
Fundus techniques can incorporate average and extensive
views [3]. Figures 2 and 3 depict retinal scans obtained via
fundus photography [36], [37]. In Figure 2, the normal
indications of a healthy eye have been depicted, whereas, in
FIGURE 4: Retinal image before the Fluorescein Angiograms (FA)
Figure 3, the dark spot of a blister of fluid caused by CSR
disease has been shown.

FIGURE 5: Retinal image after the Fluorescein Angiograms (FA)

FIGURE 2: Normal Retinal Fundus scan 3) OPTICAL COHERENCE TOMOGRAPHY (OCT)

Optical Coherence Tomography (OCT) appeared as an
advanced automation technology used for detection and
diagnosis of different diseases developed in the early 1990s.
Above mentioned approach is similar to ultrasound imaging.
However, instead of making use of sound it uses the light.
The combination of catheters and endoscopes with OCT
produces high-resolution imaging of the organ system. OCT
can provide tissue images intangible and situ form.
Additionally, two essential modalities of OCT are named as
FIGURE 3: Retinal Fundus scan with CSR (i) Time-Domain (TD), and (ii) Spectral Domain (SD) [37].
In CSC, multimodal-imaging methods incorporate OCT Time domain is used to detect the interference patterns in
with improved deep imaging, Indocyanine green creating a picture, whereas the Spectral Domain SD fuses a
angiography, fundus auto fluorescence, OCTA, and
3 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

spectrometer to build examining time. SD can get pictures testing goals. In the following list, we briefly describe two
more rapidly than TD due to its effective imaging method. publicly-available datasets along with their accessibility.
The physiologic segment gets the visual light beam 1) OCTID DATASET
partitioned into the pigmented part (pigment epithelium) and The OCT imaging database is an open-source OCT imaging
the nervous part. The layers of retina are named as database, which is accessible at the website of University of
Ganglionic cell layer, inner plexiform layer, Nerve fiber Waterloo, Canada. This database contains 102 MH, 55
layer, Outer plexiform layer, External limiting membrane, AMD, 107 DR, as well as 206 NO retinal images. The
Internal limiting membrane, Outer nuclear layer, Retinal interface of the dataset is user-friendly, and all the diverse
pigment epithelium, Inner nuclear layer, Photoreceptor outer groups have been classified in the form of discrete datasets.
segment, Interface between IS and OS, Photoreceptor inner The users can easily access the required datasets via a unique
segment and Outer Photoreceptors[38]. These layers are DOI link. Additionally, the images can be observed and
present in retina OCT as shown in Figure 6. downloaded in the form of folders or as zip files [41].
2) ZHANG’S LAB
Zhang’s lab [42] contains the largest dataset of labeled OCT
images. There are thousands of publicly available datasets of
OCT images. It has been referred to specifically in
identifying medical diagnosis and treatable disease by
image-based deep learning. At Zhang’s lab, various images
are available for both training and testing purposes of various
Machine Learning ventures in retinal disease detection.
This review article provides a comprehensive review of
FIGURE 6: Normal retinal OCT image [33] RPE: Retinal Pigment
several state-of-the-art technologies, which make use of
Epithelium, IS/OS: Interface between IS and OS, IS Photoreceptor Inner
Artificial Intelligence (AI) algorithms for automatic
Segment, ILM: Inner Limiting Membrane, ELM: External Limiting
detection of CSR disease in patients. Most of these
Membrane, OS: Photoreceptor Outer Segment, OPL: Outer Plexiform
algorithms lead to the evolution of new models that are
Layer, IPL: Inner Plexiform Layer, OPR: Outer Photoreceptor/RPE
trained on proprietary datasets that are also discussed in this
complex, INL: Inner Nuclear Layer, ONL: Outer Nuclear Layer, GCL:
study. These models have assisted in development of many
Ganglionic Cell Layer, NFL: Nerve Fiber Layer.
commercially-available products used by hospitals and
In Figure 6, the layers of the OCT image of retina do not ophthalmologists to automatically detect CSR in patients. To
indicate any abnormality or disease [39]. Normally a person the best of our knowledge, no such study is available in
with perfect vision has the same OCT as depicted in Figure literature for summarization of such technologies for CSR
6. In Figure 7, the sub retinal fluid and intra-retinal fluid disease detection.
(OCT Scan) are linked with the CSR disease [40]. Besides, The remainder of this article has been organized as
the patient suffering from CSR produces a different OCT follows: Section II describes a detailed review of the
scan, as depicted in Figure 7. literature. It also contains the methodology to organize this
review article by discussing the content inclusion and
exclusion criteria. Furthermore, it includes case studies as
well as Artificial Intelligence (AI) techniques in imaging,
using machine learning and deep learning. Section III
contains an analytical discussion based on the literature
review and establishes certain research findings and open
questions for researchers. Finally, Section IV provides a
FIGURE 7: Retinal OCT image with CSR disease detailed conclusion.

B. OCT Imaging Datasets II. LITERATURE REVIEW

In this study, two publicly available datasets associated with In this section, a comprehensive literature review has been
CSR are analyzed. A typical dataset contains a range of compiled based on the following steps:
records, and these CSR datasets consist of the repositories of
OCT and fundus images. Several researchers normally i. Identification of relevant and advanced research
access and utilize these publicly accessible datasets, which articles.
can be easily retrieved using their specific links. In various ii. Articles emphasizing on the basic understanding the
experimental studies, a portion of the all-out images is research problem.
normally downloaded from these datasets in order to train iii. A comprehensive search strategy that identifies the
new Machine Learning models and algorithms and achieve research problem.

4 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

iv. Extraction of the desired data from the selected 1) SEARCHING STUDY
research articles. In this phase, the objective of CSR disease study is defined
v. Validation and fact-checking of the collected data. and elaborated. Based on a defined selection criterion for
vi. Representation of data for better visualization and automatic CSR detection, a number of suitable CSR papers
readability. have been selected for further analysis.
The literature review articles have been searched from the 2) DETERMINING LITERATURE
following reputable and well-recognized scientific After an extensive search of relevant articles, the next phase
publishers and organizations: was to determine the valuable articles for this study. For that
i. The American Academy of Ophthalmology, purpose, all the repetitive articles were eliminated based on
(http://www.aaojournal.org/) their citations, authors’ integrity, and institutional
ii. The JAMA Network, affiliations. After studying the text of each article, the eye
(http://jamanetwork.com/journals/jama) physicians were consulted for settling differences and a
iii. Investigative Ophthalmology and Visual Science, better understanding of the CSR disease.
(http://iovs.arvojournals.org/) 3) CSR DATA DERIVATION
iv. American journal of ophthalmology, Various advanced CSR detection algorithms based on AI,
(https://www.ajo.com/) ML, and DL were studied and analyzed in this phase. This
v. Digital library of IEEE Xplore, process provided a list of state-of-the-art algorithms based on
(http://ieeexplore.ieee.org) inclusion and exclusion criteria.
vi. Springer Nature Link, (http://link.springer.com/) 4) QUALITY EVALUATION & RESULTS
vii. Elsevier, (https://www.elsevier.com/) SUMMARIZATION
viii. Science Direct, (http://www.sciencedirect.com/) Finally, all the discoveries have been summarized after
These scientific databases have been accessed to evaluating the quality of the content.
accumulate peer-reviewed journal articles, review articles, In the following subsections, an extensive review of various
and conference papers. Figure 8 illustrates the technologies based on advanced research articles is
methodological flowchart of this review with different provided. First, the content inclusion/exclusion criteria are
phases. detailed. Then, the criterion of journal selection is briefly
In Figure 8 the methodological phases through which this described, which is followed by goal of this review.
review article has been arranged and presented for the
researchers, scientists, physicians, and practitioners to get 5) CONTENT INCLUSION CRITERIA
CSR treatment insights through detection using artificial "Detection of CSR in OCT images using artificial
intelligence techniques. There are four different phases of the intelligence” is set as the sign indication for literature search.
review process, as shown in Figure 8. The abstracts, titles, and methodologies of each article have
been checked for the research prompt and facts. All those
articles were selected which have a clear research domain of
identification and detection of CSR through retinal images.
Another criterion was set to check the use of data sciences
and Artificial Intelligence in CSR disease detection.
6) CONTENT EXCLUSION CRITERIA
The aim of this review is to study and organize algorithms
and techniques that have been used for the segmentation of
CSR from retinal images. In this way, the articles having
algorithms but not relevant to CSR have not been considered
in this review. Additionally, the articles were selected based
on the repute and the impact fact of the journal. In the case
of conference/symposium papers, their proceedings and
ranking of the conference were considered for this review.
Another criterion was to weigh the articles on the basis of
their citations and institutional affiliations. Finally, the
contents of similar papers have not been considered to
maintain the integrity in the quality of the review article.
7) SELECTION OF JOURNAL ARTICLES
The research articles were selected by cautious investigation
for its consideration as well as prohibition standards. The
selected articles were copied in the system library and
FIGURE 8: Flow chart of systematic review process
spared, the terminology used to spare articles represent title,
5 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

publishing year, publisher, author first and last initials. Such to detect CSR at an early stage. Two recent studies confirmed
description facilitated in arranging the research articles and the findings of different PEDs on OCT in patients with CSR.
their proficient recovery as per inquiries. References of the Hence, FA has additionally exhibited zones of diffuse
considerable number of papers have been taken and spared epitheliopathy in patients suffering from constant CSR [47].
in EndNote reference library. Based on this selection Furthermore, E.Costanzo et al. [48] had compared the
criterion, a database of articles was set up which included OCTA with multimodal imaging, which resulted in CSR
articles related to basic understanding of retinal disease detection. The authors have performed experiments of both
based on studies of human anatomy, the CSR disease and its techniques using the same dataset of 33 eyes of 32 patients.
variants, and the recent studies of utilization of AI techniques The population and protocols in this study have been
for automatic detection of CSR. In the following sections, analyzed to find the results of both techniques.
these studies are discussed in detail. Consequently, the resultant dataset of both techniques has
8) GOALS OF THIS REVIEW been naive compared to the previous studies in the treatment
In this review, the sequel of precise and complete study of of eye diseases. This procedure used the description and
CSR algorithms has been explored for automatically classification of abnormal choroidal vessels. The
segmenting process in retinal OCT and fundus images. The experimental analysis declares OCTA had all the earmarks
significant goal of the review is to analyze the development of being a promising method since it kept away from the
of CSR algorithms and recognition of CSR from fundus weight of intravenous injections of dye which probably
images. The analysis of optimistic algorithms has some would complicate genuine symptoms. The findings of this
applicable clinical attributes for diagnosis, treatment, and study verified that OCTA technique was eligible for the
prevention. This review article presents the novel insights in detection of an unusual choroidal vessel specimen among all
using AI techniques for CSR. As far as we could possibly of the cases. Similar studies using the OCTA method had the
know, there is no such review in literature to give a thorough option to identify the distinctive neovascular network,
review of such technologies. The significance of this review usually CNV, in 58% of eyes with lingering CSR [49].
is to present recent pattern recognition technologies, based Another similar study resulted in excessive specificity and
on the methodologies in pre-processing and segmentation sensitivity of 100% in OCTA images [50].
from OCT and fundus images that have already detected the F. Rupesh Agrawal et al. [51] proposed early detection of
CSR disease. CSR with the help of an image binarization. For this purpose,
a sample of 78 eyes from 39 patients was extracted for the
9) DETECTION OF CSR BASED ON IMAGING experiment. The significant vascularity records with intense
TECHNOLOGIES CSR compared and clear solid fellow eyes were observed.
In this section, several imaging technologies for CSR The choroidal vascular changes in CSR of [52], [53] had
detection are discussed. In each case, the relevant research resulted in similar observations. The CVI was observed
studies are pointed out, along with their brief descriptions critical, and LA was lower in the fellow eyes compared to
and possible shortcomings. the age-coordinated healthy subjects. The stromal zone had
M. M. Teussink et al. [43] had differentiated OCT not been significant for exactness as the age-coordinated
angiography features of a patient with acute CSR selection of healthy subjects was required. Subsequently, the
considering FA as well as Indo Cyanine Green Angiography general vascularity index was present in the fellow eyes
(ICGA). In conducting experiment, the research team compared to the age-matched healthy eyes that generally
included 225 eye images along with 100 FA and ICGA make the eyes prone to leakage. The authors also stated that
images. Additionally, 125 and 30 OCT angiography images they could not use normalization before binarization.
of patients were recorded. The results of chronic CSR S. Weng Xiang et al. [54] proposed the OCTA to tackle
reflected the irregular choriocapillaris flow patterns. Several CSR disease. This research was conducted on seventy
researchers have studied the combination of both FA and patients and seventy-five eyes were utilized to perform
ICGA [44], [45]. However, this study has been confined to a OCTA on them. Two readers analyzed the obtained images
modest number of patients. Further improvements in and CNV was evaluated at outer retina. The authors
knowledge regarding vascular changes require prospective concluded that the CSR disease (along with the CNV
and longitudinal studies. disease) mostly occurs in patients over the age of 50 years
J. Y. Shin Velthoven et al. [46] had used ophthalmoscope [55]. The authors collected samples of 70 eyes of patients as
to detect CSR from OCT scans. The CSR detection was dataset for their experiments. The advanced mediums for
normally depending on FA or biomicroscopy. The sample photography were used which had fundus FA. To detect CSR
data of scanned eye images of 38 patients for testing in images, the OCTA appeared to be the best algorithm and
purposes was considered. The OCT ophthalmoscope usually stated that no other technique could perform the detection
provides additional knowledge related to the zone involved similar to this technique. The results of the experiments
and is practically identical to the statistics proposed for FA. represent 97.77% accuracy, approximately 100% sensitivity,
Therefore, Ophthalmoscope proves to be the best technique and a specificity level of 93.33%. The authors have

6 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

compared their technique with those of Quaranta-El iii. Sub-retinal fluid volume measurement via manual
Maftouhi et al. [49], who claimed that 12 eyes suffering segmentation.
from chronic CSC displayed the characteristic of the iv. Obtaining sub-retinal fluid contour and volume from
choroidal hyper permeability except authentication of CNV. the Balloon Snake algorithm.
Michael R et al. [56] proposed a rational technique for The experimental results using the Balloon Snake
OCT image scan. This technique develops high resolution algorithm were reported to be approximately 30 minutes
tomography for ocular tissues. The sample dataset of 16 faster as compared to the manual detection methods. The
patients for examination purpose was taken and results were researchers have introduced different strategies for
correlated with slit-lamp biomicroscopy and fundus measuring the volume of retinal fluid such as manual, semi-
photography [57]. Likewise, optical imaging tomography manual technique or automatic commercially-available
has provided a quantitative assessment for the accumulation software products [65]. The techniques have been precise
of subcutaneous fluid, which can be used to achieve irregular and regeneratable that may require human collaborations to
goals of solid separation with high efficiency [58]. develop the segmentation line. The Balloon Snake-based
In addition to the above research, Z. W. Jelena Novosel et algorithm is capable of measuring a retinal fluid volume
al. [54] proposed the locally-adaptive and loosely-coupled within 5 minutes. Using this methodology in clinical
approach to detect retinal fluids in CSR disease. The dataset examinations is found to be extremely reliable, accurate, and
from an in-process study on CSR in a hospital in Germany fast.
was collected. The results were compared by medical doctors K. Gao, et al. [37] had proposed instinctive technique
manually and by using automatic trained segmentation. which depends on graph theory and preceding B-scan
Furthermore, the authors compared their algorithm with information to automatically segment retinal layers in CSR
other similar researches in a subjective manner. A few diseases. The segmentation process includes the estimation
methodologies for the division of fluid-associated of the external nuclear layer boundaries, inner limiting
abnormalities reported the TPR ranging between 86 to 96%, membrane, photoreceptor inner segments, abrasion of eye,
[59], [60] that is similar to the authors attained 95% TPR. and pigmented layer of retina and RPE-choroid on graph
Overall, their algorithm presented a flexible solution for search model. The region of flexible search was developed
retinal fluid segmentation owning high accuracy with calculation of the thickness. The experimental results
qualitatively. had shown that it meant entire thickness in contrast with
M.Wu Xiang et al. [61] proposed the max-flow manual segmentation of (3.68±2.96 μm) as well as
optimization algorithm based on a three-dimensional (3-D) (5.84±4.78 μm) respectively. Recent researchers also
continuous method. This method covered two of the worked on automatic methods, but these strategies were
Neurosensory Retinal Detachment (NRD), together with the dependent upon intense irregularities and breaks in the
Pigment Epithelial Detachment (PED). The proposed retinal layers because of outlines brought about by retinal
method is based on a probability map using a random veins. Future direction has been proposed to classify the
forecast model. The authors used 37 retinal OCT images for border of CSC zone, which constitutes of different layers to
training and testing purposes. The experimental results have measure CSC locale also breakdown variety of every retinal
shown that 92.1%, 0.53%, 94.7%, and 93.3% constitute the layer in various infections.
Dice Similarity Coefficient (DSC), for segmentation with Similarly, SG Odaibo et al. [66] had proposed an
NRD as compared to 92.5%, 0.14%, 80.9%, and 84.6%, with optimized technique to detect CSR in OCT images through
PED segmentation. Various methods were proposed to cloud-based mobile Artificial Intelligence (AI) platform.
conduct the classification of CSR but all of them had been This technique played a vital role in assisting physicians with
observed as 2-D images which classified retinal disorder [62] less cost and high-accuracy results. The algorithm is capable
with a few restrictions that need to be enhanced in future to classify fluid and non-fluid OCT images. Interestingly,
investigations. There were extremely little fluid sectors used this algorithm was integrated as an application in IOS
as a rarity in the proposed methodology for classification of operating system for a test run. After testing and
retinal fluids. experimenting, the authors reported the sensitivity levels
H. J. Chien et al. [63] proposed the Balloon Snake between 82.5% & 97%, whereas the specificity range was
Algorithm implemented in MATLAB tool. This algorithm reported between 52% & 100%. Moreover, the authors
was utilized for the segmentation of fluid from OCT retinal proposed to test large amount of dataset in their future
images in terms of ophthalmology [64]. The primary experiments [56].
objective of this research work was to detect edges and M. LU HE et al. [67] derived comparison of the
volume of retinal fluid and provide the best possible OCT multicolored imaging (MC) with the Color Fundus
images. The following steps have been referred for CSR Photography (CFP) and employed the result achiever
disease detection. technique to detect retinal fluid in CSR. The authors took the
i. Patient selection and data collection. sample of 75 eye scans of 69 patients for experimental setup.
ii. OCT scan protocol. The examination of all the samples for the clarity of vision,

7 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

intraocular pressure, slit lamp, CFP (Digital Fundus Camera, patients for testing. Their results determined that the OCTA
VISUCAM 200; Carl Zeiss Meditec AG, Jena, Germany), provided a better solution among all other techniques such
FFA, MC image, as well as SD-OCT was conducted. The as, FA, SD-OCT, and ICGA, because these techniques are
comparison of results indicated that MC plays a more not diagnostic. Therefore, the OCTA proved to detect CNV
influential role of detection compared to the CFP. The MC in CSR by a direct visualization and resulted in attaining
method attained 92.0% detection rate compared to the CFP. 100% sensitivity and specificity. Furthermore, a number of
Many other studies have also shown that MC is useful for recent researches have reported that OCTA has the best
detecting fundus diseases [68]. However, the limitation of ability to detect the choriocapillaris vascular density [70].
MC is that there were no regular and continuous session Table 1 presents a summary of the aforementioned CSR
visits arranged for the patients. Furthermore, it has been detection techniques in an ascending order. The summaries
observed that the artifacts may affect the quality of images, of the intricate peer-reviewed articles comprise of algorithms
thereby compromising disease detection. used, number of datasets considered for the study and the
G Cennamo et al. [69] used OCTA for the examination of effectiveness of the proposed results. Essentially, Table 1
the Choriocapillary Vascular Density (CNV) in CSR disease. provides a quick glance of CSR imaging technologies for the
The authors considered twelve OCTA eye scans from twelve readers of this review.

TABLE 1
SUMMARY OF CSR DETECTION BASED ON IMAGING TECHNOLOGIES

Reference # Year Authors Dataset Key Technique Results

Moderately higher (median JI,
[43] 2015 M. M. Teussink et al. 2 OCT eye scans images OCTA 0.74) as well as moderate
(median JI, 0.52)
J. Y. Shin Velthoven et Optomap-AF Visual acuteness (VA) 0.2
[46] 2015 73 OCT eye scans
al. (range 0–1.3).
33 OCTA eye scans of 32 OCTA with multimodal
[48] 2015 E. Costanzo et al. Mean BCVA was 0.23 ± 0.25
consecutive patients. imaging
F. Rupesh Agrawal et OCT scans of 78 eyes of Image binarization Mean best-corrected optical
[51] 2016
al. 39 patients acuteness 0.31 ± 0.38
Optical-coherence 97.77%, 100%, 93.33%
70 OCTA eye scans of
[54] 2016 S. Weng Xiang et al. tomographic angiography Accuracy, sensitivity and
patients
(OCTA) specificity respectively.
[56] slit-lamp bio microscopy
2016 Michael R et al. 16 patients OCT images and fundus photography Efficiency
Z. W. Jelena Novosel et Dice coefficient, TPR and FPR
[59] 2016 74 retinal OCT pictures
al. 0.96,95%and 1%, respectively
max-flow optimization
[61] 2017 M.Wu Xiang et al. 37 retinal SD-OCT 95 % Accuracy.
algorithm
Performance: 30 min faster
[63] 2019 H. J. Chien et al. 20 OCT scan images
Balloon Snake Algorithm than manual segmentation
Suggests absolute thickness
K. Gao, et al. Graph theory and B-scan differences parallel to manual
[37] 2019 200 B-scan images
segmentation 3.68±2.96 μm
and 5.84±4.78 μm
Cloud-based-mobile
90 time-domain OCT 82.5% to 97% sensitivity and
[66] 2019 SG Odaibo et al. Artificial Intelligence (AI)
images specificity from 52% to 100%
platform
75 SD-OCT eyes scans of Multicolored imaging (MC)
[67] 2020 M. LU HE et al. 69 patients including and Color Fundus 92.0% as a detection rate.
disease eyes Photography (CFP)
Optical-Coherence CVD of controls
OCTA images scans of 12 Tomography Angiography resulted in significant higher in
[69] 2020 G Cennamo et al.
eyes Group 2 at baseline (whole,
parafovea and fovea p<0.05).

10) ARTIFICIAL INTELLIGENCE BASED DETECTION OF supervised learning, the targeted variables are predicted from
CSR a prearranged indicator (known as the Objective Function)
Artificial Intelligence (AI) techniques are divided into three for simplicity. Taking advantage of arrangement of factors,
different phases: training, validation and testing for generation of the Objective Function guides contribution to
arrangement of images [71]. The most advanced AI anticipated targets. The dataset training phase keeps iterating
techniques include Machine Learning (ML) and its more until the model attains an ideal degree of precision. The
recent variant known as Deep Learning (DL). Machine examples of supervised learning techniques are Regression
Learning (ML) is further categorized into supervised, & Logistic Regression, KNN, Decision Tree, and Random
unsupervised and reinforcement learning. In case of Forest, etc. In unsupervised learning, the prediction of the

8 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

objective or desired variable is carried out. It is employed for image and have the option to separate one from the other
the collecting population in various groups, which is broadly based on their significance [73].
used for fragmenting purposes. The examples of
unsupervised learning include K-means and Apriori
algorithm. In reinforcement learning, the machine is
normally trained to sort and categorize explicit choices. It is
programmed to strategically learn itself when introduced to
a certain circumstance and up-skill reliably with trial and
error experimentation. This allows the algorithm to
FIGURE 10: CNN architecture for classification of CSR
experience and attempt to catch the ideal data to attain the
The development and advancement of the CNN declares the
precise business choices. An example of reinforcement
capacity to learn automatically specific to a training dataset
learning includes Markov Decision Process. Figure 9 shows
as per the predictive modeling problem. The image
the methodology of Machine Learning (ML) model in which
classification can be the best example of CNN and their
fundus retinal image is used for classification of CSR.
outcomes can easily be determined to specific features
anywhere on input images [74]. Figure 10 depicts the
methodology of CNN architecture in which fundus retinal
image is used for classification of CSR. It is evident that the
input eye scan image goes through multiple convolutional
layers which have the self-learning ability to focus on the
CSR patch. After the final CNN layer, the model can
accurately segment the image and determine the exact
FIGURE 9: ML model for classification of CSR
location and extent of the CSR patch. Similar to this patch,
The process starts with inputting a retinal image, which is
numerous other features related to various anomalies are
preprocessed for any anomaly detection or better
focused in different DL models. In the following paragraphs,
visualization for the next stages. After preprocessing, the
the use of various ML/DL algorithms in automatic detection
next stage is feature extraction in which various parts and
of CSR disease have been reviewed, along with brief
features of the eye scan are extracted. Then specific parts of
description of the results and shortcomings.
the scan are selected for training or testing purposes. If the
Hassan B et al. [75] had proposed the automated and
image is inputted specifically for the training purpose, it is
rational algorithm to detect CSR through OCT images. In
processed through various algorithmic stages to boost up the
this methodology, the input OCT scan has been preprocessed
ML model, thereby obtaining a particular pre-classified CSR
by de-noising and enhancing the image and then ILM and
related image. Contrarily, if the image was initially inputted
choroid are segmented by applying optimal thresholding.
for testing purposes, then it goes through the predefined
Feature extraction has been performed and the training
algorithmic stage to test the model for its validation. In this
dataset is passed through SVM classifier and the required
way, the model is first trained using a large dataset of images,
classification is achieved. The algorithm detects images by
and then tested using more data. Once the algorithmic model
applying a Support Vector Machine (SVM). In this
goes through hundreds of thousands of images, it is trained
mechanism, five different characteristics were extracted
to determine any CSR diseases in future inputted images.
from a dataset gathered from different hospitals. The
Deep Learning (DL) is an advanced version of ML based
performance of the SVM algorithms has been observed
on neural networks capable of unsupervised learning from an
better than others as proposed in [76]. The authors algorithm
unstructured and unlabeled dataset. In recent studies, DL
truly distinguished 88 out of 90 subjects attaining 97.77%
algorithms are vastly used by researchers and scientists in
accuracy, 100% sensitivity and 93.33% specificity [75].
almost every field including image recognition and
A. M. Syed et al. [77] proposed the rational technique to
automated disease detection. In ophthalmology, DL has been
overcome the CSR disease problem. The methodology
applied to fundus images, optical intelligibility tomography
consists of three phases. In the first phase, the robust
and visual fields, accomplishing robust classification
reconstruction of 3-D OCT retinal surfaces was planned. In
execution in the discovery of CSR [72]. Figure 10 shows the
the second phase 2 feature sets are formulated, one is
methodology of deep learning model in which fundus retinal
thickness profile and the other is cyst fluid. The retinal
image is used for classification of CSR.
subject is classified using SVM classifier. The lodged
11) CONVOLUTIONAL NEURAL NETWORK (CNN)
algorithm worked on multiple OCT images for the detection,
The Convolutional Neural Network (CNN) is a deep
whereas several researchers worked on OCT and fundus
learning (DL) architecture that requires unstructured data
images [78], [79]. The sample size of 18 OCT images was
(such as scanned images) as inputs, that allocate learning
used as a dataset. After conducting the experiments and
weights and biases to different viewpoints/entities in the
calculations, authors achieved the accuracy of 98.88%,

9 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

sensitivity of 100%, and specificity of 96.66%. Furthermore, scans into little blob areas dependent on local convexity by
the authors reported that their proposed algorithm is faster amassing the log-scale-standardized convolution responses
and more robust in detecting CSR compared to the state-of- of every individual gLoG filter. The feature vectors were
the-art researches. Finally, the authors claimed to utilize the sustained in the direction of unsupervised clustering
proposed algorithm for diagnosing other macular disorders algorithm. The testing results showed excellent performance
such as, AMD and glaucoma. as compared to the other manual methods. Authors achieved
S. Khalid et al. [80] proposed a decision support system 95.15%, 93.65% and 94.35% as a mean true positive volume
that instinctively aimed to dredge up CSR from retinal fraction, positive predictive and dice similarity coefficient
images using the SVM classifier. The methodology works by respectively. Various state-of-the-art models were proposed
preprocessing i.e. sparse de-noise the input image and then in [88]–[91] for detecting sub-retinal fluids in CSR, most of
segmentation is applied. This segmentation results into which are retinal layer segmentation dependent. Therefore,
retinal layers whose retinal thickness profile is generated. the introduced methodology can deliver precise division
Features extracted are passed to the SVM classifier. outcomes that are exceptionally reliable with the argument
The dataset contained 90 OCT scans of 73 patients for without retinal layer separation. This technique may provide
experimental purpose. The system ratings of 99.92%, 100%, dependable NRD segmentations to SD-OCT images and can
and 99.86% attained accuracy, sensitivity, and specificity be helpful for clinical analysis. The future direction of the
respectively [80]. In addition, the authors stated that any proposed strategy essentially centers on the expansion of 3-
other intricate research works had utilized their OCT pictures D segmentation.
to distinguish retinal variations from the norm [81], [82]. DL plays a vital role in detecting sub-retinal fluids from
D. Xiang et al. [83] proposed the random forecast classifiers SD-OCT images and improving its boundaries day by day. J
incorporating a live wire algorithm. This algorithm intends De Fauw et al. [92] had proposed a self-activating and self-
to find surfaces in optical coherence tomography. The learning technique to segment retinal fluid from SD-OCT
sample SD-OCT images of 24 spectral domains suffering CSR using a DL algorithm known as ‘deep segmentation
with CSR from various hospitals were used to conduct network’. This network uses a 3-D U-Net architecture to
experiments. The authors used both manual and automated interpret abraded OCT into tissue map. This algorithm was
methods to detect the CSR. The authors claimed that the trained on a very large dataset of 14,884 tissue-maps with
results obtained from their algorithm outperformed the committed diagnosis and was later tested on 887 SD-OCT
contemporary methods. Even though, the proposed strategy scanned images. On the other hand, several researchers have
has been utilized in OCT imaging with CSR, the underlying worked on a single end-to-end black-box network that needs
algorithm could be efficiently utilized in different origins of indefinite number of OCT scans [10]. This algorithm uses a
OCT pictures. B. Hassan et al. [84] proposed the CAD-based group of diverse segmentation and classification models; (i)
fully automated algorithm using the KNN classifier to detect CNV (ii) MRO (iii) Normal (iv) CSR (V) Mac hole. This
CSR. The authors utilized a dataset of 80 retinal subjects to also incorporates four DL phases: the digital OCT scans,
conduct the experimental study. The proposed algorithm was segmentation network, tissue segmentation map, and
assumed to extract retinal layers through 3D features. The classification network. The 'black-box' issue is distinguished
proposed algorithm supports practitioners and doctors while as a hindrance to the use of DL within healthcare systems.
detecting CSR efficiently and accurately. The results had The system structure intently coordinates the clinical
achieved 100% efficiency in diagnosing the CSR. The dynamic procedure, distinguishing decisions independently
numerous strategies eventually resulted in the acquisition of from the subsequent referrals based on the scan. This practice
eye checks as it may be like OCT imaging and has emerged permits a clinician to investigate and envision an
as one of the best modern methods [85], [86]. Despite its interpretable division instead of merely giving a conclusion
good features, the algorithm can be modified more to attain and referral recommendation. The proposed algorithm has
the exact images with precision. In addition, this achieved a total error rate (1- accuracy) of 5.5% AUC on the
methodology can equally be reached out to other retinal referral decisions. In [93], K. Gao et al. proposed the
pathologies in the future. innovative image-to-image, double-branched, fully-in-
Most of the CSR detection methods commonly impart the convolutional networks (DA-FCN) for segmentation of
process of segmentation. However, the authors Z Ji et al. [87] retina fluid from SD-OCT images. Many researchers worked
proposed an algorithm that detects CSR without on the same disease but with other methods like Montuoro et
segmentation of retinal layers named as ‘Hessian-based al., proposed a 3D graph search technique for the detection
Aggregate comprehensive Laplacian of Gaussian of CSR [94]. First the dataset is extended by applying
algorithm’. The ensuing framework is now commonly used mirroring technique and then shallow coarse is learnt by
for the automatic segmentation of Neurosensory Retinal designing double branched structures. In the end, area loss is
Detachments (NRD). The 23 samples of longitudinal SD- joint with softmax loss in order to learn features that are more
OCT scans were collected for the testing phase. The useful.
execution and implementation process require filtration of B-

10 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

The proposed technique was prospective to provide a proposed algorithm is dependent on DL may evaluate CSR
measurable evaluation of NRD just as PED. The sample of portrayed on color fundus in a moderately solid manner. The
52 SD-OCT test images dataset was acquired from 35 eye authors claimed that their results showed that the AUC was
scans of 35 patients. The results depicted that 94.3% of true 0.927 (95% CI: 0.895-0.959) and the p-value was less than
positive value fraction, 94.3% dice similarity efficient, and 0.001. Moreover, the cutoff threshold was 0.5 and the
96.4% positive predicative value was attained. Moreover, the accuracy was 85.7%. The restricted dataset was used for
proposed method had excessive achievements in TPVF and creating model, but the trials showed the special quality of
DSC compared to the other related methods given in the profound learning innovation. Future translational exertion
literature. These attainments can be enhanced and improved is alluring to make such a tool accessible to the clinical
to develop features that are more distinguished by practice for beginning CSR screening results in early stage
introducing loss factors, which can segment retinal fluid diagnosis. Several researchers have worked on color fundus
efficiently and precisely. images but these images were observed to be related to
In addition to the aforementioned causes of CSR, diabetic retinopathy and diabetic macula edema [10] but very
Maculopathy plays a vital role in the blindness of the human limited work have been conducted on CSR. The combination
eye. It has been diagnosed by automated systems like OCT of DL algorithm and fundus photography can be very
imaging, and it may act as harmful object for the human eye. efficient for the ophthalmologists in their clinical diagnosis,
B. Hassan et al. [95] had proposed to develop a rational which may reduce unnecessary usage of OCT-images and
technique for segmentation and grading according to medical fluorescein photography. Finally, it must be pointed out that
criteria. An algorithm called Support Vector Machine this study contains some limitations which include the less
(SVM) was used, which works in six different phases: (i) variety of images and the effects of image artifacts on CSC
data acquisition (ii) preprocessing (iii) retinal layer assessment Y. Ruan et al. [99] proposed the Fully
segmentation (iv) retinal fluid detection (v) feature set Convolutional Networks (FCN) with multiphase level set
formulation, and (vi) classification. The clinical review of named (FCN-MLS) for segmenting the boundaries in retinal
CSR and ME alongside the seriousness examination and 3D images. The system proposed by the authors has been
profiling utilizing OCT images were also presented. partitioned into three main phases: (i) Pre-processing (ii)
The dataset of 30 OCT test images was collected from FCN for layer boundaries segmentation (iii) Distance
seventy-three different patients. Several researchers have regularization level set (DRLS). The sample size of 10 SD-
worked on retinal fluid segmentation like computer aided OCT test datasets was collected, thereby making it 120 B-
diagnosis method by B. Hassan et al. [95] for self-activating scan images. The experimental results obtained a gross
identification of idiopathic CSR. The proposed framework average difference of absolute boundary location and the
first performs preprocessing one the image, then retinal layer gross average difference of boundary thickness, which were
segmentation is performed using Tensor Graph and then 5.88 ± 2.38μm and 5.81 ± 2.19μm, respectively. Various
retinal fluid is detected. Features extracted are passed to researchers have worked on retinal layer segmentation using
SVM classifier and disease grading is performed. The Artificial Intelligence (AI) and it worked as the state-of-art
precision of the proposed work was reported as the true methodology [100]. TJN Rao et al. [101] proposed the
affirmative rate as well as the true negative rate of 97.78%, segmentation algorithm based on neural networks to detect
96.77%, and 100%, respectively. The technology of OCT CSR. The capacity of a convolutional neural system to
imaging, when integrated with ML algorithms, turns out to highlight the separation of unpretentious spatial variety that
be advantageous. R V Teja et al. [96] had proposed the converts retinal fluid into division task [102]. This method
integrated random forest classifier and DeepLab algorithm will find and segment the retinal fluid from OCT images and
for the detection of CSR. The sample sizes of 768 B-scan reduce the impact of noise at background. It is based on two
tests were collected. The phases of classification and phases: (i) pre-processing phase and (ii) fluid segmentation
segmentation were determined through a number of famous phase. Dataset for experiments consists of 15 OCT images
frameworks known as the XGBoost classifier, the random obtained from CSR patients. The experimental results with
forest classifier and the DeepLab. The testing score on OCT dice rate of 91%, precision 93%, and recall 89% were
images was achieved as the mean dice of 86.23% as reported. Furthermore, the authors claimed that the
compared to the manual delineations made by the trained development of hybridization of neural network and feature
experts. Recent studies [97] also show that DeepLab set extraction techniques would be used as a future work for
another cutting edge on a several benchmark semantic image better segmentation of CSR.
segmentation datasets. Y. Zhen et al. [98] proposed the S.A.E. Hassan et al. [6] proposed a method for fully
mechanized methodology based on DL architecture termed automated Central Serous Retinopathy (CSR) detection
as InceptionV3 to recognize CSR portrayed on color fundus using deep CNN. In the developed framework, the pre-
images. The sample size for training and testing was based processing phase enhanced the image quality and eliminated
on 2504 OCT-images that were pre-processed and noise. The three pre-trained classifiers (AlexNet, ResNet-18,
normalized. The results and experiments concluded that the and GoogleNet) have been used for classification. The

11 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

outcomes indicate that AlexNet achieved the highest leakage points and UNet was used for optic disk (OD) and
accuracy 99.64%, precision 98.91%, recall 100%, and f1- macula region segmentation. The results showed that the
score 99.45% on publicly available OCT images dataset. detection was perfectly matched 60.7% in the test set. After
Pawan et al. [103] proposed an enhanced SegCaps utilizing eliminated method, the accurate detection cases
architecture based on Capsule Networks for the increased 93.4% and the dice on lesion level was 0.949,
segmentation of SRF using CSCR OCT images. The respectively. Future work includes the improvement of
developed method was outperformed in UNet architecture Computational Complexity.
and reduced parameters to be trained by 54.21%. Table 2 presents a list of the research papers considered for
Furthermore, it was also reduced the computation this review paper that particularly uses Machine/Deep
complexity of SegCaps by 37.85%, with comparison to other learning technologies for CSR detection. This table
method's performance. comprises scholar name, year of publication, proposed
Chen et al. [104] presented a study consisting of 2104 FFA algorithm to solve the problem, suggested methodology for
images from FFA sequences of 291 eyes, including 137 right the problem, dataset used, and outcomes respective to
eyes and 154 left eyes from 262 patients. The attention gated sensitiveness, specificity, accuracy, and region fall under
Network (AGN) has been used for the segmentation of ROC curve.

TABLE 2
SUMMARY OF CSR DETECTION BASED ON MACHINE/DEEP LEARNING IN CSR DETECTION

Reference # Year Authors Dataset Key Technique Results

SVM classifier 97.77%, 100%, 93.33% accuracy,
[75] 2016 Hassan B et al. 34 OCT scans sensitivity and specificity
respectively
98.88%, 100%, and 96.66%
SVM classifier accuracy, sensitivity and specificity
[77] 2016 A. M. Syed et al 90 OCT volumes
respectively

Multi layered SVM classifier 99.92%, 100%, and 99.86%

90 OCT scans of 73
[80] 2017 S. Khalid et al. Accuracy, sensitivity and specificity
patients
respectively.
Sum of mean differences of absolute
Random Forest and Live wire boundary and thickness positioning
6 macula-centered SD-
[83] 2018 D. Xiang et al. algorithm differentiated to manual
OCT images
segmentation results: 3.68 ± 2.96 μm
and 5.84 ± 4.78 μm.
CAD-based fully automated
80 CSR retinal Value of k = 5, achieving 99.8 %
[84] 2018 B. Hassan et al. algorithm using the KNN
subjects accuracy 100 % Efficiency
classifier
Sum of true positive volume fraction
Z Ji et al. 23 longitudinal SD- Hessian-based Aggregate (95.15%), positive predicative value
[87] 2018
OCT scans. comprehensive Laplacian of (93.65%) and dice similarity
Gaussian algorithm coefficient (94.35%).
14,884 training tissue DL algorithm known as ‘deep Mean error rate (1- accuracy) 5.5%
J De Fauw et al. segmentation network
[92] 2018 maps and 887 SD- AUC on referrals decision.
OCT scans.
52 SD-OCT images DA-FCN (fully convolutional True positive value fraction 94.3%,
K. Gao et al. networks)
[93] 2019 dataset of 35-eyes of dice similarity efficient 94.3%, and
35 subjects positive predicative value 96.4%.
30 OCT images of SVM classifier Accuracy, true positive rate and true
[95] 2019 B. Hassan et al. negative rate of 97.78%, 96.77%,
seventy-three patients
100% separately
R V Teja et al. Random forest classifier and
[96] 2019 OCT images DeepLab algorithm Average Dice Sore of 86.23%
DL architecture termed as AUC 0.927, p-value < 0.001
[98] 2019 Y. Zhen et al. 2504 Fundus images
InceptionV3 cutoff threshold 0.5, accuracy 85.7%
Gross average difference of absolute
Y. Ruan et al FCN-MLS boundary location and boundary
[99] 2019 10 SD-OCT subjects
thickness were (5.88 ± 2.38μm) and
(5.81 ± 2.19μm) respectively.
15 OCT volumes got Segmentation algorithm based on Dice rate 91%, precision 93%, and
[101] 2019 TJN Rao et al. neural networks to detect CSR
from CSR patients recall 89%

12 VOLUME XX, 2017

 Author Name: Preparation of Papers for IEEE Access (February 2017)

S. A. E. Hassan et AlexNet, ResNet and GoogleNet 99.64% classification accuracy of

[6] 2021 309 OCT images
al. AlexNet
SegCaps (Capsule Network) Reduction of trainable parameters by
[103] 2021 Pawan, S. J., et al. OCT images 54.21% and reduction of
computational complexity by
37.85%.
Chen, Menglu, et U-Net and Attention Gated Accurate detection cases increased to
[104] 2021 291 FFA images Network (AGN)
al. 93.4%

III. DISCUSSION scanned images fed to them for further analysis. Therefore,
After a thorough study of 29 relevant articles, section II the reliability and size of datasets used to train the models
looked at several CSR disease detection methodologies and must be major criteria for quantifying the accuracy of the
algorithms. There were two broad categories for this results. Tables 1 & 2 give the size of datasets as well as the
approach: a) Detection of CSR based on classical imaging relevant accuracy of the results, and subsequently, a more
technologies, and b) Detection of CSR based on accurately reported result based on smaller dataset should be
Machine/Deep Learning techniques. Table 1 & 2 highlight considered as less reliable compared to a less accurate result
the key information about the relevant publications regarding based on a large dataset.
both of these categories. This information includes the Another important parameter is to check the complexity and
authors’ list, the size of datasets, the underlying algorithms, comprehensiveness of the algorithms used for detection. The
and results summary. The Artificial Intelligence (AI) comprehensiveness of algorithm means that it should take
techniques mentioned in this review are primarily based on into account every type of anomalies for CSR detection. The
ML & DL approaches. The ML approach is further classified DL algorithms have greater complexity due to their usage of
into three sub-categories, namely, supervised learning Artificial Neural Networks (ANNs), which are composed of
(Regression, Decision Tree, Random Forest, KNN, and multiple layers through which the data is processed and
Logistic Regression etc.), the unsupervised learning (Apriori transformed. Hence, they are more accurate at the cost of
algorithm, K-means) and the Reinforcement Learning their complexity. The downside of these algorithms is that
(Markov Decision Process). From a careful analysis, it can they are normally computationally intensive in nature, and
be determined that the utilization of the classical ML they require a lot of processing time during the training and
approach has yielded excellent results in accuracy, testing phases. Their cost of reliability and accuracy is
reliability, and speed for a given sample dataset to diagnose computational time, especially if they are trained and tested
CSR. over a huge dataset.
On the other hand, the DL technique is quite beneficial for Another major limitation in most of these ML/DL models
fundus image sorting procedures, and it yields better results is that they have been trained & tested on
accuracy results than the classical ML approach. However, proprietary datasets that are not publicly available. This
the results largely depend on the size of datasets and the raises an important question of their integrity and
complexity of the underlying algorithms. This review authenticity. There are very limited publicly available
provides a qualitative and quantitative assessment of the datasets; in fact, there are only two such datasets. Most of the
algorithms in both aforementioned categories (see Tables 1 models are trained on private datasets, which are
& 2 for summary). Furthermore, this meta-analysis of proprietary to their owner institutions. Therefore, more
previous researches has identified several algorithms such as public datasets should be available for the research
OCTA algorithms and SVMs. The results are described in community to train & test their AI models. This practice will
terms of accuracy, sensitivity, and error rates. From a careful not only encourage the researchers to access the datasets
analysis, it can be determined that the KNN classifier (in the easily, but it will also help to standardize the models for the
category of ML) has proposed to attain the best accuracy, CSR disease detection. In this manner, the research
followed by the SVM classifier has attained high accuracy community can develop a generic standard detection model.
and double-branched algorithm as well as area Constraint
Fully Convolutional Networks (DA-FCN) with moderate
accuracy. In case of DL, the Mobile Artificial Intelligence
Platform reported to have attained maximum accuracy,
followed by Robust Reconstruction of 3-D OCT Retinal
Surfaces, which attained 98.88% accuracy. Finally, the Split-
Spectrum Amplitude-Decorrelation Angiography (SSADA)
methods attained slightly moderate accuracy. The most
significant measure to ascertain the reliability of results is the
size of datasets in all cases. Some of the researchers claim
that their results portray an accurate detection of the CSR
disease based on their models trained under specific datasets.
In most cases, the datasets are not large enough to properly
FIGURE 11: Classifiers used in the literature for CSR detection
train the models to accurately detect any future anomalies in
13 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

Figure 12 depicts the classification accuracy attained by of vision loss and blindness as a result of CSR. These
ML/DL classifiers and figure 13 shows total no of images methods were categorized into classical imaging and ML/DL
used in the literature with respect to each classifier. based techniques. Furthermore, this study encompassed the
102.00% retinal datasets, strategies proposed for CSR assessment and
100.00% accuracy. The additional retinal neovascularization,
98.00% hemorrhages, miniaturized scale aneurysm, and exudates
96.00% were also been highlighted for dataset of CSR detection. The
94.00% roles of assessment measurements for Computer Aided
92.00% Diagnosis (CAD) frameworks were deeply examined. The
90.00% significance of each ML and DL methodology was
ascertained to perform a better analysis of CSR for research
community and ophthalmologists. A number of recent
research articles were studied and examined to compare their
respective algorithms, usage of their datasets and the
FIGURE 12: Comparison of literature based on classification accuracy
accuracy of their results. A detailed analysis showed that the
advanced ML/DL models are being employed for accurate,
350 309 291 309309 reliable, and rapid detection of CSR. The algorithm
300 introduced by Hassan et al [84] is the most promising
250
200 algorithm as it gives 99.8% accuracy. This open source CSR
150 90 90 73 80 detection model will provide agility, sustainability flexibility
100 70
37 and cost effectiveness, thus it can serve the mankind in a
50 6 10 20
0 much better way.
Balloon…
Max flow…

CAD-based…
Live-wire…
Multi-layered…

Detection of CSR is evolving, using Machine

FCN-MLS
U-Net
OCTA

SVM

Res-Net
GoogleNet
Classifier

Optomap-AF
AlexNet

Learning/Deep Learning Algorithms and imaging

technologies. However, it is still mattered of controversy,
and to date studies that have been published so far depends
on the old methodologies and are unrestrained. Failure to
FIGURE 13: Comparison of literature based on validation data images detect and diagnose CSR may result in complete vision loss.
IV. CONCLUSION Although, researches showing promising results but more
Traditionally, the CSR disease due to retinal fluids was research is required on publically available dataset and
detected by experienced ophthalmologists after manually improving the computational complexity. Therefore, further
examining the scanned images of retina. However, this studies are warranted to address this issue, improving the
approach is time-consuming, inaccurate and unreliable. With patient’s care.
the recent advancements in technology, the CSR has been This paper provides a comprehensive review than existing
automatically detected with the aid of various AI-based review articles on existing methodologies to detect CSR
models. Although the optimal detection of CSR is still a big through ML/DL techniques. Additionally, the limitations of
challenge for the researchers, these intelligent techniques these methodologies as well as the key suggestions have
have become more popular and effective compared to the been mentioned to cope up with the limitations for the
manual detection approaches. The recent literature found assistance of research community. Moreover, an open
that numerous intelligent algorithms on OCT images are question for future research can be to create an open-source
utilized for the CSR detection, and the experts of CSR detection model based on Deep Learning technologies,
segmentation procedures carefully control this process. which is accessible to the research community to test it on
However, it is still subjected to human error. Therefore, the their proprietary datasets or publicly available ones.
latest technology relies on AI based algorithms, which
include ML & DL. These algorithms form accurate detection CONFLICT OF INTEREST
models, which are trained & tested on the legacy datasets. The authors of this review paper have no conflict of interest.
After proper testing and tuning, the researchers and
physicians for the automatic CSR detection and diagnosis ACKNOWLEDGEMENT
utilize these models. In recent times, these models are This work was supported by the Artificial Intelligence &
becoming more and more accurate and reliable. Furthermore, Data Analytics Lab (AIDA), Prince Sultan University,
there are numerous commercially available products using Riyadh, Saudi Arabia in collaboration with Riphah Artificial
AI based models. Moreover, the performance of these Intelligence Research (RAIR) Lab, Riphah International
models must be precisely analyzed with the standard University, Faisalabad Campus, Pakistan. Authors also
measurements and benchmarks. appreciate Prince Sultan University support for paying APC
In this study, a detailed review of automatic detection of for this publication.
CSR using AI techniques has been presented. These
automatic detection methods assist in detecting the disease
14 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

REFERENCES [10] V. Gulshan et al., “Development and validation of a

deep learning algorithm for detection of diabetic
retinopathy in retinal fundus photographs,” JAMA -
[1] C. E. Willoughby, D. Ponzin, S. Ferrari, A. Lobo, K. Journal of the American Medical Association, 2016,
Landau, and Y. Omidi, “Anatomy and physiology of doi: 10.1001/jama.2016.17216.
the human eye: Effects of mucopolysaccharidoses [11] A. Tufail et al., “An observational study to assess if
disease on structure and function - a review,” automated diabetic retinopathy image assessment
Clinical and Experimental Ophthalmology. 2010, software can replace one or more steps of manual
doi: 10.1111/j.1442-9071.2010.02363.x. imaging grading and to determine their cost-
[2] A. A. Dahl and T. R. Gest, “Retina Anatomy,” effectiveness,” Health Technology Assessment, vol.
Medscape Reference. pp. 1–4, 2015, doi: 20, no. 92, pp. 1–72, Dec. 2016, doi:
”https://www.nvisioncenters.com/education/eye- 10.3310/hta20920.
structure-anatomy/. [12] R. Rajalakshmi, R. Subashini, R. M. Anjana, and V.
[3] M. U. Akram, S. Akbar, T. Hassan, S. G. Khawaja, Mohan, “Automated diabetic retinopathy detection
U. Yasin, and I. Basit, “Data on fundus images for in smartphone-based fundus photography using
vessels segmentation, detection of hypertensive artificial intelligence,” Eye, vol. 32, no. 6, pp. 1138–
retinopathy, diabetic retinopathy and papilledema,” 1144, Jun. 2018, doi: 10.1038/s41433-018-0064-9.
Data in Brief, 2020, doi: 10.1016/j.dib.2020.105282. [13] M. Bhaskaranand et al., “Automated Diabetic
[4] S. Akbar, T. Hassan, M. U. Akram, U. U. Yasin, and Retinopathy Screening and Monitoring Using
I. Basit, “AVRDB: Annotated dataset for vessel Retinal Fundus Image Analysis,” Journal of
segmentation and calculation of arteriovenous Diabetes Science and Technology, vol. 10, no. 2, pp.
ratio,” in Proceedings of the 2017 International 254–261, Mar. 2016, doi:
Conference on Image Processing, Computer Vision, 10.1177/1932296816628546.
and Pattern Recognition, IPCV 2017, 2017. [14] T. A. Soomro, J. Gao, T. Khan, A. F. M. Hani, M. A.
[5] S. Akbar, M. Sharif, M. U. Akram, T. Saba, T. U. Khan, and M. Paul, “Computerised approaches
Mahmood, and M. Kolivand, “Automated for the detection of diabetic retinopathy using retinal
techniques for blood vessels segmentation through fundus images: a survey,” Pattern Analysis and
fundus retinal images: A review,” Microscopy Applications. 2017, doi: 10.1007/s10044-017-0630-
Research and Technique. 2019, doi: y.
10.1002/jemt.23172. [15] S. Akbar, M. U. Akram, M. Sharif, A. Tariq, and U.
[6] S. A. E. Hassan, S. Akbar, S. Gull, A. Rehman, and ullah Yasin, “Arteriovenous ratio and papilledema
H. Alaska, “Deep Learning-Based Automatic based hybrid decision support system for detection
Detection of Central Serous Retinopathy using and grading of hypertensive retinopathy,” Computer
Optical Coherence Tomographic Images,” in 2021 Methods and Programs in Biomedicine, vol. 154, pp.
1st International Conference on Artificial 123–141, Feb. 2018, doi:
Intelligence and Data Analytics (CAIDA), 2021, pp. 10.1016/j.cmpb.2017.11.014.
206–211, doi: [16] S. Akbar, M. U. Akram, M. Sharif, A. Tariq, and S.
10.1109/CAIDA51941.2021.9425161. A. Khan, “Decision support system for detection of
[7] T. A. Soomro et al., “Deep Learning Models for hypertensive retinopathy using arteriovenous ratio,”
Retinal Blood Vessels Segmentation: A Review,” Artificial Intelligence in Medicine, vol. 90, pp. 15–
IEEE Access. 2019, doi: 24, Aug. 2018, doi: 10.1016/j.artmed.2018.06.004.
10.1109/ACCESS.2019.2920616. [17] S. Akbar, M. U. Akram, M. Sharif, A. Tariq, and U.
[8] M. D. Abràmoff et al., “Improved automated ullah Yasin, “Decision Support System for Detection
detection of diabetic retinopathy on a publicly of Papilledema through Fundus Retinal Images,”
available dataset through integration of deep Journal of Medical Systems, vol. 41, no. 4, p. 66,
learning,” Investigative Ophthalmology and Visual Apr. 2017, doi: 10.1007/s10916-017-0712-9.
Science, vol. 57, no. 13, pp. 5200–5206, Oct. 2016, [18] D. Hua et al., “Retinal Microvascular Changes in
doi: 10.1167/iovs.16-19964. Hypertensive Patients with Different Levels of
[9] A. A. van der Heijden, M. D. Abramoff, F. Verbraak, Blood Pressure Control and without Hypertensive
M. V van Hecke, A. Liem, and G. Nijpels, Retinopathy,” Current Eye Research, pp. 1–8, Jun.
“Validation of automated screening for referable 2020, doi: 10.1080/02713683.2020.1775260.
diabetic retinopathy with the IDx-DR device in the [19] H. Bogunovic et al., “Prediction of Anti-VEGF
Hoorn Diabetes Care System,” Acta Treatment Requirements in Neovascular AMD
Ophthalmologica, vol. 96, no. 1, pp. 63–68, Feb. Using a Machine Learning Approach,” Investigative
2018, doi: 10.1111/aos.13613.
15 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

Opthalmology & Visual Science, vol. 58, no. 7, p. https://www.slideshare.net/AnkushWeginwar/centr

3240, Jun. 2017, doi: 10.1167/iovs.16-21053. al-serous-retinopathy-73712617 (accessed Sep. 30,
[20] M. Rohm et al., “Predicting Visual Acuity by Using 2020).
Machine Learning in Patients Treated for [31] “Central serous retinopathy.”
Neovascular Age-Related Macular Degeneration,” https://www.medicalnewstoday.com/articles/32060
Ophthalmology, vol. 125, no. 7, pp. 1028–1036, Jul. 6#overview (accessed Sep. 30, 2020).
2018, doi: 10.1016/j.ophtha.2017.12.034. [32] M. B. Breukink et al., “Chronic central serous
[21] M. Treder, J. L. Lauermann, and N. Eter, chorioretinopathy: long-term follow-up and vision-
“Automated detection of exudative age-related related quality of life,” Clinical Ophthalmology, vol.
macular degeneration in spectral domain optical Volume 11, pp. 39–46, Dec. 2016, doi:
coherence tomography using deep learning,” 10.2147/OPTH.S115685.
Graefe’s Archive for Clinical and Experimental [33] “Central Serous Retinopathy overview,” 2013.
Ophthalmology, vol. 256, no. 2, pp. 259–265, 2018, [Online]. Available: https://www.rcophth.ac.uk/wp-
doi: 10.1007/s00417-017-3850-3. content/uploads/2014/08/Focus-Winter-2013.pdf.
[22] C. González‐Gonzalo et al., “Evaluation of a deep [34] G. Manayath, R. Ranjan, S. Karandikar, V. Shah, V.
learning system for the joint automated detection of Saravanan, and V. Narendran, “Central serous
diabetic retinopathy and age‐related macular chorioretinopathy: Current update on management,”
degeneration,” Acta Ophthalmologica, vol. 98, no. 4, Oman Journal of Ophthalmology, vol. 11, no. 3, p.
pp. 368–377, Jun. 2020, doi: 10.1111/aos.14306. 200, Sep. 2018, doi: 10.4103/ojo.OJO_29_2018.
[23] E. Vaghefi, S. Hill, H. M. Kersten, and D. Squirrell, [35] M. A. Abouammoh, “Advances in the treatment of
“Multimodal Retinal Image Analysis via Deep central serous chorioretinopathy,” Saudi Journal of
Learning for the Diagnosis of Intermediate Dry Age- Ophthalmology. 2015, doi:
Related Macular Degeneration: A Feasibility 10.1016/j.sjopt.2015.01.007.
Study,” Journal of Ophthalmology, vol. 2020, pp. 1– [36] “Central Serous Retinopathy before and after
7, Jan. 2020, doi: 10.1155/2020/7493419. Angiograms.” .
[24] T. R. V. Bisneto, A. O. de Carvalho Filho, and D. M. [37] K. Gao, W. Kong, S. Niu, D. Li, and Y. Chen,
V. Magalhães, “Generative adversarial network and “Automatic retinal layer segmentation in SD-OCT
texture features applied to automatic glaucoma images with CSC guided by spatial characteristics,”
detection,” Applied Soft Computing, vol. 90, p. Multimedia Tools and Applications, vol. 79, no. 7–
106165, May 2020, doi: 8, pp. 4417–4428, Feb. 2020, doi: 10.1007/s11042-
10.1016/j.asoc.2020.106165. 019-7395-9.
[25] D. C. Hood, Z. Z. Zemborain, E. Tsamis, and C. G. [38] M. Cardoso, T. Arbel, A. Melbourne, and H.
De Moraes, “Improving the Detection of Glaucoma Bogunovic, “Fetal, Infant and Ophthalmic Medical
and Its Progression: A Topographical Approach,” Image Analysis: International Workshop, FIFI 2017,
Journal of Glaucoma, vol. 29, no. 8, pp. 613–621, and 4th International Workshop, OMIA 2017, Held
Aug. 2020, doi: 10.1097/IJG.0000000000001553. in Conjunction,” 2017.
[26] P. Kaur and P. K. Khosla, “Artificial Intelligence [39] “Normal Coloured Retinal OCT image.”
Based Glaucoma Detection,” Springer, Singapore, https://www.ceenta.com/patientresources/the-retina
2020, pp. 283–305. (accessed Oct. 01, 2020).
[27] A. Sarhan, J. Rokne, and R. Alhajj, “Approaches for [40] “Central Serous Retinopathy Coloured Retinal OCT
Early Detection of Glaucoma Using Retinal Images: image.”https://imagebank.asrs.org/Content/imageba
A Performance Analysis,” Springer, Cham, 2020, %0Ank/ped-WITH-
pp. 213–238. SRF.jpg/imagefull;max$643,0.ImageHandler
[28] V. Mohammadzadeh et al., “Macular imaging with (accessed Oct. 01, 2020).
optical coherence tomography in glaucoma,” Survey [41] P. Gholami, P. Roy, M. K. Parthasarathy, and V.
of Ophthalmology, vol. 65, no. 6, pp. 597–638, 2020, Lakshminarayanan, “OCTID: Optical coherence
doi: 10.1016/j.survophthal.2020.03.002. tomography image database,” Computers and
[29] F. Li et al., “Deep learning-based automated Electrical Engineering, vol. 81, 2020, doi:
detection of glaucomatous optic neuropathy on color 10.1016/j.compeleceng.2019.106532.
fundus photographs,” Graefe’s Archive for Clinical [42] D. S. Kermany et al., “Identifying Medical
and Experimental Ophthalmology, vol. 258, no. 4, Diagnoses and Treatable Diseases by Image-Based
pp. 851–867, 2020, doi: 10.1007/s00417-020- Deep Learning,” Cell, vol. 172, no. 5, pp. 1122-
04609-8. 1131.e9, 2018, doi: 10.1016/j.cell.2018.02.010.
[30] “Central Serous Retinopathy.” [43] M. M. Teussink et al., “Oct angiography compared
16 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

to fluorescein and indocyanine green angiography in Ophthalmologica, vol. 91, no. 5, pp. e358–e362,
chronic central serous chorioretinopathy,” Aug. 2013, doi: 10.1111/aos.12059.
Investigative Ophthalmology and Visual Science, [54] S. Weng, L. Mao, S. Yu, Y. Gong, L. Cheng, and X.
2015, doi: 10.1167/iovs.15-17140. Chen, “Detection of choroidal neovascularization in
[44] N. Kitaya et al., “Features of abnormal choroidal central serous chorioretinopathy using optical
circulation in central serous chorioretinopathy,” coherence tomographic angiography,”
British Journal of Ophthalmology, 2003, doi: Ophthalmologica, vol. 236, no. 2, pp. 114–121,
10.1136/bjo.87.6.709. 2016.
[45] A. Scheider, J. E. Nasemann, and O. E. Lund, [55] R. F. Spaide et al., “Central Serous
“Fluorescein and indocyanine green angiographies Chorioretinopathy in Younger and Older Adults,”
of central serous choroidopathy by scanning laser Ophthalmology, vol. 103, no. 12, pp. 2070–2080,
ophthalmoscopy,” American Journal of Dec. 1996, doi: 10.1016/S0161-6420(96)30386-2.
Ophthalmology, 1993, doi: 10.1016/S0002- [56] C. Wong et al., “Optical Coherence Tomography of
9394(14)73524-X. Central Serous Chorioretinopathy,” American
[46] J. Y. Shin, H. J. Choi, J. Lee, and M. Choi, “Fundus Journal of Ophthalmology, vol. 120, no. 1, pp. 65–
autofluorescence findings in central serous 74, 1995, doi: 10.1016/S0002-9394(14)73760-2.
chorioretinopathy using two different confocal [57] H. (Ed. . Min, Stereo Atlas of Vitreoretinal Diseases.
scanning laser ophthalmoscopes : correlation with Springer Nature, 2019.
functional and structural status,” 2015, doi: [58] E. A. Swanson et al., “In vivo retinal imaging by
10.1007/s00417-015-3244-3. optical coherence tomography,” Optics Letters,
[47] N. E. Gross, R. F. Spaide, D. L. L. Costa, S. J. 1993, doi: 10.1364/ol.18.001864.
Huang, J. M. Klancnik, and A. Aizman, [59] J. Novosel, Z. Wang, H. De Jong, M. Van
“INDOCYANINE GREEN ANGIOGRAPHY – Velthoven, K. A. Vermeer, and L. J. Van Vliet,
GUIDED PHOTODYNAMIC THERAPY FOR “LOCALLY-ADAPTIVE LOOSELY-COUPLED
TREATMENT OF CHRONIC CENTRAL LEVEL SETS FOR RETINAL LAYER AND
SEROUS A Pilot Study,” pp. 288–298, 2003. FLUID SEGMENTATION IN SUBJECTS WITH
[48] E. Costanzo et al., “Optical Coherence Tomography CENTRAL SEROUS RETINOPATHY Rotterdam
Angiography in Central Serous Chorioretinopathy,” Ophthalmic Institute , 1 Rotterdam Eye Hospital ,
Journal of Ophthalmology, 2015, doi: Rotterdam , The Netherlands Quantitative Imaging
10.1155/2015/134783. Group , Departmen,” pp. 702–705, 2016.
[49] M. Q. Maftouhi, A. El Maftouhi, and C. M. Eandi, [60] X. Xu, K. Lee, L. Zhang, M. Sonka, and M. D.
“Chronic Central Serous Chorioretinopathy Imaged Abràmoff, “Stratified Sampling Voxel Classification
by Optical Coherence Tomographic Angiography,” for Segmentation of Intraretinal and Subretinal Fluid
American Journal of Ophthalmology, pp. 1–8, 2015, in Longitudinal Clinical OCT Data,” IEEE
doi: 10.1016/j.ajo.2015.06.016. Transactions on Medical Imaging, 2015, doi:
[50] M. Antonio et al., “Association of Choroidal 10.1109/TMI.2015.2408632.
Neovascularization and Central Serous [61] M. W. U. Englin et al., “Three-dimensional
Chorioretinopathy With Optical Coherence continuous max flow optimization-based serous
Tomography Angiography,” vol. 02116, no. 8, pp. retinal detachment segmentation in SD-OCT for
899–906, 2015, doi: central serous chorioretinopathy,” vol. 8, no. 9, pp.
10.1001/jamaophthalmol.2015.1320. 4257–4274, 2017.
[51] R. Agrawal, J. Chhablani, K.-A. Tan, S. Shah, C. [62] D. C. Fernández, “Delineating fluid-filled region
Sarvaiya, and A. Banker, “CHOROIDAL boundaries in optical coherence tomography images
VASCULARITY INDEX IN CENTRAL SEROUS of the retina,” IEEE Transactions on Medical
CHORIORETINOPATHY,” Retina, vol. 36, no. 9, Imaging, 2005, doi: 10.1109/TMI.2005.848655.
pp. 1646–1651, Sep. 2016, doi: [63] H. Chien, C. Chang, H. Chien, and C. Chang,
10.1097/IAE.0000000000001040. “Seminars in Ophthalmology Application of the
[52] C. Brandl, H. Helbig, and M. A. Gamulescu, Balloon Snake in the Volume Measurement of
“Choroidal thickness measurements during central Subretinal Fluid in Central Serous
serous chorioretinopathy treatment,” International Chorioretinopathy Application of the Balloon Snake
Ophthalmology, 2014, doi: 10.1007/s10792-013- in the Volume Measurement of Subretinal Fluid in
9774-y. Central Serous Chorioretinopathy,” Seminars in
[53] L. Yang, J. B. Jonas, and W. Wei, “Choroidal vessel Ophthalmology, vol. 0, no. 00, pp. 1–6, 2019, doi:
diameter in central serous chorioretinopathy,” Acta 10.1080/08820538.2019.1640749.
17 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

[64] M. Wei, Y. Zhou, and M. Wan, “A fast snake model [74] “Convolutional Layers overview.”
based on non-linear diffusion for medical image https://machinelearningmastery.com/convol%0Auti
segmentation,” Computerized Medical Imaging and onal-layers-for-deep-learning-neuralnetworks/
Graphics, 2004, doi: (accessed Oct. 01, 2020).
10.1016/j.compmedimag.2003.12.002. [75] B. Hassan, G. Raja, T. Hassan, and M. Usman
[65] J. Ruiz-Medrano, M. Pellegrini, M. G. Cereda, M. Akram, “Structure tensor based automated detection
Cigada, and G. Staurenghi, “Choroidal of macular edema and central serous retinopathy
Characteristics of Acute and Chronic Central Serous using optical coherence tomography images,”
Chorioretinopathy Using Enhanced Depth Imaging Journal of the Optical Society of America A, vol. 33,
Optical Coherence Tomography,” European Journal no. 4, p. 455, Apr. 2016, doi:
of Ophthalmology, vol. 27, no. 4, pp. 476–480, Jun. 10.1364/JOSAA.33.000455.
2017, doi: 10.5301/ejo.5000796. [76] B. Nicholson, J. Noble, F. Forooghian, and C.
[66] S. G. Odaibo, M. MomPremier, R. Y. Hwang, S. Meyerle, “MAJOR REVIEW Central Serous
Yousuf, S. Williams, and J. Grant, “Mobile artificial Chorioretinopathy : Update on Pathophysiology and
intelligence technology for detecting macula edema Treatment,” Survey of Ophthalmology, vol. 58, no.
and subretinal fluid on OCT scans: Initial results 2, pp. 103–126, 2013, doi:
from the DATUM alpha Study,” arXiv preprint 10.1016/j.survophthal.2012.07.004.
arXiv:1902.02905, 2019. [77] A. M. Syed, T. Hassan, M. U. Akram, S. Naz, and S.
[67] L. He, C. Chen, Z. Yi, X. Wang, J. Liu, and H. Khalid, “Automated diagnosis of macular edema and
Zheng, “CLINICAL APPLICATION OF central serous retinopathy through robust
MULTICOLOR IMAGING IN CENTRAL reconstruction of 3D retinal surfaces,” Computer
SEROUS CHORIORETINOPATHY,” Retina, vol. Methods and Programs in Biomedicine, vol. 137, pp.
40, no. 4, pp. 743–749, Apr. 2020, doi: 1–10, Dec. 2016, doi: 10.1016/j.cmpb.2016.09.004.
10.1097/IAE.0000000000002441. [78] N. F. Mokwa, T. Ristau, P. A. Keane, B. Kirchhof,
[68] L. Reznicek et al., “Scanning laser ‘en face’ retinal S. R. Sadda, and S. Liakopoulos, “Grading of Age-
imaging of epiretinal membranes,” Saudi Journal of Related Macular Degeneration: Comparison
Ophthalmology, 2014, doi: between Color Fundus Photography, Fluorescein
10.1016/j.sjopt.2014.03.009. Angiography, and Spectral Domain Optical
[69] G. Cennamo, C. Comune, F. Mirra, P. Napolitano, Coherence Tomography,” Journal of
D. Montorio, and G. de Crecchio, “Choriocapillary Ophthalmology, vol. 2013, pp. 1–6, 2013, doi:
vascular density in central serous chorioretinopathy 10.1155/2013/385915.
complicated by choroidal neovascularization,” [79] H. S. Sandhu et al., “Automated diabetic retinopathy
Photodiagnosis and Photodynamic Therapy, vol. 29, detection using optical coherence tomography
p. 101604, Mar. 2020, doi: angiography: a pilot study,” British Journal of
10.1016/j.pdpdt.2019.101604. Ophthalmology, vol. 102, no. 11, pp. 1564–1569,
[70] G. Cennamo, D. Montorio, F. Mirra, C. Comune, A. Nov. 2018, doi: 10.1136/bjophthalmol-2017-
D’Alessandro, and F. Tranfa, “Study of vessel 311489.
density in adult-onset foveomacular vitelliform [80] S. Khalid, M. U. Akram, T. Hassan, A. Nasim, and
dystrophy with optical coherence tomography A. Jameel, “Fully Automated Robust System to
angiography,” Photodiagnosis and Photodynamic Detect Retinal Edema, Central Serous
Therapy, vol. 30, p. 101702, Jun. 2020, doi: Chorioretinopathy, and Age Related Macular
10.1016/j.pdpdt.2020.101702. Degeneration from Optical Coherence Tomography
[71] P. Aggarwal, “Machine learning of retinal pathology Images,” BioMed Research International, 2017, doi:
in optical coherence tomography images,” Journal 10.1155/2017/7148245.
of Medical Artificial Intelligence, vol. 2, pp. 20–20, [81] T. Hassan, M. U. Akram, B. Hassan, A. Nasim, and
Sep. 2019, doi: 10.21037/jmai.2019.08.01. S. A. Bazaz, “Review of OCT and fundus images for
[72] D. S. W. Ting et al., “Artificial intelligence and deep detection of Macular Edema,” in IST 2015 - 2015
learning in ophthalmology,” British Journal of IEEE International Conference on Imaging Systems
Ophthalmology, vol. 103, no. 2, pp. 167–175, Feb. and Techniques, Proceedings, 2015, doi:
2019, doi: 10.1136/bjophthalmol-2018-313173. 10.1109/IST.2015.7294517.
[73] “Convolutional Neural Networks overview.” [82] S. L. Y. J, C. Madhavi, and M. A. Kumar, “Visual
https://towardsdatascience.com/acomprehensive- outcome of central serous retinopathy,” vol. 3, no. 8,
guide-to-convolutionalneural-networks-the-eli5- pp. 1885–1888, 2015.
way-%0A3bd2b1164a53 (accessed Oct. 01, 2020). [83] D. Xiang, G. Chen, F. Shi, W. Zhu, and X. Chen,
18 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

“Automatic Retinal Layer Segmentation Based on 10.1016/j.cmpb.2019.04.027.

Live Wire for Central Serous Retinopathy,” 2017, [94] A. L. M. Ontuoro, S. E. M. W. Aldstein, B. I. S. G.
pp. 118–125. Erendas, and U. R. S. C. Rfurth, “Joint retinal layer
[84] B. Hassan, R. Ahmed, B. Li, A. Noor, and Z. ul and fluid segmentation in OCT scans of eyes with
Hassan, “A comprehensive study capturing vision severe macular edema using unsupervised
loss burden in Pakistan (1990-2025): Findings from representation and auto-context,” vol. 8, no. 3, pp.
the Global Burden of Disease (GBD) 2017 study,” 182–190, 2017.
PLoS ONE, 2019, doi: [95] B. Hassan and T. Hassan, “Fully automated
10.1371/JOURNAL.PONE.0216492. detection, grading and 3D modeling of maculopathy
[85] S. H. Freiesleben, J. Soelberg, N. T. Nyberg, and A. from OCT volumes,” in 2019 2nd International
K. Jäger, “Determination of the wound healing Conference on Communication, Computing and
potentials of medicinal plants historically used in Digital systems (C-CODE), Mar. 2019, pp. 252–257,
Ghana,” Evidence-based Complementary and doi: 10.1109/C-CODE.2019.8680996.
Alternative Medicine, 2017, doi: [96] R. V. Teja et al., “Classification and Quantification
10.1155/2017/9480791. of Retinal Cysts in OCT B-Scans: Efficacy of
[86] T. Hassan, M. Usman Akram, B. Hassan, A. M. Machine Learning Methods,” in Proceedings of the
Syed, and S. A. Bazaz, “Automated segmentation of Annual International Conference of the IEEE
subretinal layers for the detection of macular Engineering in Medicine and Biology Society,
edema,” Applied Optics, vol. 55, no. 3, p. 454, Jan. EMBS, 2019, doi: 10.1109/EMBC.2019.8857075.
2016, doi: 10.1364/AO.55.000454. [97] L.-C. Chen, G. Papandreou, F. Schroff, and H.
[87] Z. J. B, Q. C. B, M. Wu, S. Niu, and W. Fan, Beyond Adam, “Rethinking Atrous Convolution for
Retinal Layers : A Large Blob Detection for Semantic Image Segmentation,” Jun. 2017, [Online].
Subretinal Fluid Segmentation in SD-OCT Images, Available: http://arxiv.org/abs/1706.05587.
vol. 1. Springer International Publishing. [98] Y. Zhen et al., “Assessment of central serous
[88] Y. Zheng, J. Sahni, C. Campa, A. N. Stangos, A. Raj, chorioretinopathy (CSC) depicted on color fundus
and S. P. Harding, “Computerized Assessment of photographs using deep Learning,” Jan. 2019,
Intraretinal and Subretinal Fluid Regions in Spectral- [Online]. Available:
Domain Optical Coherence Tomography Images of http://arxiv.org/abs/1901.04540.
the Retina,” American Journal of Ophthalmology, [99] Y. Ruan et al., “Multi-phase level set algorithm
vol. 155, no. 2, pp. 277-286.e1, Feb. 2013, doi: based on fully convolutional networks (FCN-MLS)
10.1016/j.ajo.2012.07.030. for retinal layer segmentation in SD-OCT images
[89] T. Wang et al., “Label propagation and higher-order with central serous chorioretinopathy (CSC),”
constraint-based segmentation of fluid-associated Biomedical Optics Express, vol. 10, no. 8, p. 3987,
regions in retinal SD-OCT images,” Information Aug. 2019, doi: 10.1364/BOE.10.003987.
Sciences, vol. 358–359, pp. 92–111, Sep. 2016, doi: [100] K. McDonough, I. Kolmanovsky, and I. V. Glybina,
10.1016/j.ins.2016.04.017. “A neural network approach to retinal layer
[90] A. Lang et al., “Automatic segmentation of boundary identification from optical coherence
microcystic macular edema in OCT,” Biomedical tomography images,” in 2015 IEEE Conference on
Optics Express, vol. 6, no. 1, p. 155, Jan. 2015, doi: Computational Intelligence in Bioinformatics and
10.1364/BOE.6.000155. Computational Biology (CIBCB), Aug. 2015, pp. 1–
[91] Y. Xu et al., “Dual-stage deep learning framework 8, doi: 10.1109/CIBCB.2015.7300299.
for pigment epithelium detachment segmentation in [101] T. J. Narendra Rao, G. N. Girish, A. R. Kothari, and
polypoidal choroidal vasculopathy,” Biomedical J. Rajan, “Deep Learning Based Sub-Retinal Fluid
Optics Express, vol. 8, no. 9, p. 4061, Sep. 2017, doi: Segmentation in Central Serous Chorioretinopathy
10.1364/BOE.8.004061. Optical Coherence Tomography Scans,” in 2019
[92] J. De Fauw et al., “Clinically applicable deep 41st Annual International Conference of the IEEE
learning for diagnosis and referral in retinal disease,” Engineering in Medicine and Biology Society
Nature Medicine, vol. 24, no. 9, pp. 1342–1350, Sep. (EMBC), Jul. 2019, pp. 978–981, doi:
2018, doi: 10.1038/s41591-018-0107-6. 10.1109/EMBC.2019.8857105.
[93] K. Gao et al., “Computer Methods and Programs in [102] J. Long, E. Shelhamer, and T. Darrell, “Fully
Biomedicine Double-branched and area-constraint Convolutional Networks for Semantic
fully convolutional networks for automated serous Segmentation.” in Proceedings of the IEEE
retinal detachment segmentation in SD-OCT conference on computer vision and pattern
images,” vol. 176, pp. 69–80, 2019, doi: recognition. Oct. 2015, pp. 3431-3440, doi:
19 VOLUME XX, 2017
 Author Name: Preparation of Papers for IEEE Access (February 2017)

10.1109/CVPR.2015.7298965.
[103] Pawan, S. J., et al. "Capsule Network–based
architectures for the segmentation of sub-retinal
serous fluid in optical coherence tomography images
of central serous chorioretinopathy." Medical &
Biological Engineering & Computing 59.6 (2021):
1245-1259, May 2021, doi: 10.1007/s11517-021-
02364-4
[104] Chen, Menglu, et al. "Automatic detection of
leakage point in central serous chorioretinopathy of
fundus fluorescein angiography based on time
sequence deep learning." Graefe's Archive for
Clinical and Experimental Ophthalmology (2021):
1-11, Mar 2021, doi: 10.1007/s00417-021-05151-x.

20 VOLUME XX, 2017