Comprehensive review of retinal blood vessel segmentation ...

Vol.:(0123456789)1 3

Network Modeling Analysis in Health Informatics and Bioinformatics (2021) 10:20 https://doi.org/10.1007/s13721-021-00294-7

REVIEW ARTICLE

Comprehensive review of retinal blood vessel segmentation and classification techniques: intelligent solutions for green computing in medical images, current challenges, open issues, and knowledge gaps in fundus medical images

Aws A. Abdulsahib1 · Moamin A. Mahmoud1 · Mazin Abed Mohammed2 · Hind Hameed Rasheed1,3 · Salama A. Mostafa4 · Mashael S. Maashi5

Received: 19 August 2020 / Revised: 4 February 2021 / Accepted: 13 February 2021 © The Author(s), under exclusive licence to Springer-Verlag GmbH, AT part of Springer Nature 2021

AbstractRecently, there has been an advancement in the development of innovative computer-aided techniques for the segmenta-tion and classification of retinal vessels, the application of which is predominant in clinical applications. Consequently, this study aims to provide a detailed overview of the techniques available for segmentation and classification of retinal vessels. Initially, retinal fundus photography and retinal image patterns are briefly introduced. Then, an introduction to the pre-processing operations and advanced methods of identifying retinal vessels is deliberated. In addition, a discussion on the validation stage and assessment of the outcomes of retinal vessels segmentation is presented. In this paper, the proposed methods of classifying arteries and veins in fundus images are extensively reviewed, which are categorized into automatic and semi-automatic categories. There are some challenges associated with the classification of vessels in images of the retinal fundus, which include the low contrast accompanying the fundus image and the inhomogeneity of the background lighting. The inhomogeneity occurs as a result of the process of imaging, whereas the low contrast which accompanies the image is caused by the variation between the background and the contrast of the various blood vessels. This means that the contrast of thicker vessels is higher than those that are thinner. Another challenge is related to the color changes that occur in the retina from different subjects, which are rooted in biological features. Most of the techniques used for the classification of the retinal vessels are based on geometric and visual characteristics that set the veins apart from the arteries. In this study, different major contributions are summarized as review studies that adopted deep learning approaches and machine learning techniques to address each of the limitations and problems in retinal blood vessel segmentation and classification techniques. We also review the current challenges, knowledge gaps and open issues, limitations and problems in retinal blood vessel segmentation and classification techniques.

Keywords Retinal blood · Retinal blood vessel segmentation and retinal blood vessels classification · Fundus images · Deep learning · Machine learning techniques

* Mazin Abed Mohammed [email protected]

Aws A. Abdulsahib [email protected]

Moamin A. Mahmoud [email protected]

Hind Hameed Rasheed [email protected]

Salama A. Mostafa [email protected]

Mashael S. Maashi [email protected]

1 College of Computing and Informatics, Universiti Tenaga Nasional, 43000 Kajang, Selangor, Malaysia

2 College of Computer Science and Information Technology, University of Anbar, Anbar 31001, Iraq

3 Department of Computer Engineering, Al-Esraa University College, Baghdad, Iraq

4 Faculty of Computer Science and Information Technology, Universiti Tun Hussein Onn Malaysia, 86400 Johor, Malaysia

5 Software Engineering Department, College of Computer and Information Sciences, King Saud University, Riyadh 11451, Saudi Arabia

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Network Modeling Analysis in Health Informatics and Bioinformatics (2021) 10:20

1 3

20 of 32

1 Introduction

Several techniques are used in the area of medical imag-ing for collecting quantitative data regarding the internal organs of humans. These techniques allow either nonin-vasive or in vivo gathering of quantitative information (Al-Fahdawi 2016, 2018; Abdulhay 2018). It is for this reason that these techniques are preferred in the research and diagnosis of pathologies in living tissues. The dif-ferent properties of the tissues in the human body can be visualized using a wide range of modalities, but the com-monly used modalities include photography with visible spectrum, ultrasound-based imaging, and multi-spectral imaging, which have the capability of improving the spec-tral resolution using wavelengths that are invisible to the human’s eyes, as well as providing relevant information regarding the composition of the photographed image. Other techniques used to visually examine the internal parts of a body include advanced mathematical principles such as MRI and tomographic approaches (Mohammed 2018; Mostafa 2019). The aforementioned techniques are just a subset of the many suggested imaging patterns employed in the collection of quantitative information on organs. Despite a wide range of imaging techniques, the number of approaches proposed for image processing is more than that of the imaging techniques. In this regard, the structural segmentation and characterization of medi-cal images is a very relevant and broad field, which is aimed at outlining the regions of interest. Within the con-text of medicine, this implies the segmentation of ana-tomical structures such as blood vessels and organs. The characterization procedure is aimed at providing a group of measurements that can be utilized in highlighting the properties present in the object under investigation. Sub-sequently, tissues, pathological and healthy features, etc. are distinguished using these measures (Arunkumar et al. 2018; Nayak 2008).

Ophthalmologic and cardiovascular diseases like glau-coma, diabetic retinopathy, macular degeneration, vein occlusion, arteriosclerosis, hypertension, and choroidal neovascularization can be rapidly diagnosed, screened and assessed by quantitatively analyzing the retinal fundus images, which is widely used. Of all the aforementioned diseases, diabetic retinopathy and macular degeneration are the two common causes of vision loss. A crucial step in the quantitative analysis of retinal fundus images is the segmentation of blood vessels, which involves the extrac-tion of clinically relevant features like length, tortuosity, blood vessel density, etc. from the segmented vascular tree. In addition, several applications have used the seg-mented vascular tree, some of which include the synthe-sis of retinal mosaic image, biometric identification, optic

disc identification temporal or multimodal image registra-tion, and fovea localization (Couper et al. 2002). Examples of retinal fundus images are shown in Fig. 1, which also shows the blood vessel structures that have been manually segmented. There are many limitations associated with the manual segmentation of vascular tree in retinal images and some of which include human error and time consum-ing. Human error is prone to occur in situations, in which the structure of the vessels is complex and the number of images is huge. Thus, it is important for eye specialists and ophthalmologists to have an automatic system that can help them in the segmentation and extraction of important clinical features from the retinal blood vessels during the early diagnosis of a wide range of retinal diseases and treatment assessment.

The main purpose of identifying and localizing the ves-sels of the retina is to distinguish the diverse vasculature structure tissues of retina, (which could be tight or wide) from the background of the fundus image as well as other retinal anatomical structures like abnormal lesions, macula, and optic disc. The attention of researchers has been con-tinuously focused on the area of retinal vessels identification because of the presence of non-invasively fundus imaging techniques and the important details that are obtainable from the vasculature structure for the recognition and diagnosis of a wide range of retinal pathology such as age-related macu-lar degeneration (AMD), hypertension, diabetic retinopathy (DR) and glaucoma.

Recently, there has been an advancement in the devel-opment of innovative computer-aided techniques in retinal vessels segmentation and their usage in clinical applications. This study aims to provide a detailed overview of the tech-niques available for retinal vessels segmentation. Initially, retinal fundus photography and retinal image patterns are briefly introduced. Then an introduction to the pre-process-ing operations and advanced techniques of identifying retinal vessels is deliberated. In addition, a discussion on the valida-tion and assessment of the outcomes of retinal vessel seg-mentation is presented. There are some challenges associ-ated with vessels’ classification of the retinal fundus images, which include the low-contrast accompanying the fundus image and the inhomogeneity of the background lighting. The inhomogeneity occurs as a result of the process of imag-ing, whereas the low contrast which accompanies the images is caused by the variation between the background and the contrast of the various blood vessels. This means that the contrast of thicker vessels is higher than those that are thin-ner. Another challenge is related to the color changes that occur in the retina from different subjects, which are rooted in biological features. Most of the techniques used for the classification of the retinal vessels are based on geometric and visual characteristics that set the veins apart from the arteries.



1 3

of 32 20

In this study, different major contributions are summa-rized as follows:

• We review studies that adopted deep learning approaches and machine learning techniques to address each of the limitations and problems in retinal blood vessel segmen-tation and classification techniques.

• We highlight the deep learning approaches and machine learning techniques role in retinal blood vessel segmenta-tion and classification techniques.

• We review the current challenges, knowledge gaps and open issues, limitations and problems in retinal blood vessel segmentation and classification techniques.

The paper is structured as follows: Sect. 2 describes the retinal quantification measures, which refers to the correla-tion amongst the cardiovascular, heart diseases, and retinal variations. Section 3 discusses the retinal fundus medical images. In Sect. 4, the retinal vessel segmentation techniques provide the categorization and description of the extant reti-nal blood vessel segmentation methodologies. A summary of findings of retinal vessels segmentation methods modern and approaches is discussed in Sect. 5. Section 6 describes the retinal vessel classification techniques. Finally, the con-clusion is made in Sect. 7.

2 Retinal quantification measures

With the aid of retinal quantification measures, the correla-tion amongst the cardiovascular, heart diseases, and retinal variations can be described (Patton 2006). The retinal quan-tification measurements also allow the investigation of the associations between the measurements and different clinical factors like blood pressure, body mass index (BMI) and age in many previous studies. These measures are applied as predictions in computerized diagnosis models. The quan-titative measurement of the retina involves measuring the structure of the retinal blood vessels, as well as angles at bifurcations, junctional exponents, measures of vascular tortuosity, length, fractal dimensions, diameter ratios, and AVR (Patton 2006). Basically, in this section, the different measures are described.

The junctional indicator alludes to the magnitude of x in the equation dx0 = dx1 + dx2, which denotes the root vessel’s diameters (d0), with its branches (d1, d2). Theo-retically, the indicator is estimated as a value of 3 in well-formed vascular network so that the magnitude of losses in the vascular structure can be reduced (Patton 2006). Esti-mating the junctional exponent is a challenging task, espe-cially when the diameter of the branches is bigger than the root vessel and also it is highly sensitive to errors that may

Fig. 1 An example of retinal fundus images along with their corresponding manually segmented blood vessel structures obtained from the DRIVE data set



1 3

20 of 32

occur during the measurement of the diameter. To address these issues, the authors in Chapman et al. (2002) develop a measured proportion of deviation of the junctional indi-cator. The measure is developed from the optimal value of the junctional exponent, � = [d30(d31 + d32)]1 = 3 = d0 which addresses the above-mentioned problems. In their study, they infer that there is an important variation between the � of healthy people and people with fringe vessel ailment. The angles that are situated amid the branches in the vessel’s branching location are referred to as the vascular bifurcation angles. In theory, the estimated optimal value of angle is 75°. Researchers have found a relationship between the angle and a wide range of medi-cal outcomes. For example, reduction in arteriolar angles is reported for hypertension, with increase in age and low birth-weight males. Furthermore, a relationship has been found between a lower density of vascular network and lower branching angles, but no relationship is found amongst peripheral vascular disease and bifurcation angles (Patton 2006). The vascular tortuosity, which is the meas-urement of the vessels’ curvature, has been employed in measuring the magnitude of health environments like the retinopathy of prematurity (ROP), in which case, a rise in arteriolar tortuosity is one of the earliest determinants of plus disease. Another characteristic of diabetic retinopathy is venous beading. The length or the diameter proportion is determined as the length from a specific vessel bifurcation to the midpoint in the previous bifurcation and isolated using the distance across the parent vessel in the bifurca-tion. It is used in measuring the constriction of the vessel and is presented as higher in hypertension.

The fractal geometry structure of the vessel through which the fractal dimension is assessed, which is based on the best junctional supporter x = 3, can be very near to 2 (ideally satisfying the existing space). Based on research, an estimate of the value of the fractal dimensions is 1.7, and it is also assumed that the dimension of arterioles is lower than the venular diameters (Patton 2006). The venular diam-eters and arteriolar and their ratios, the AVR, are the most commonly utilized measurements in terms of retinal vessel quantification. In typical cases, the estimation of a vessel’s diameter is made at the centre of the sides of the double-Gaussian cross-sectional profile, which reduces the image’s defocusing effect on the diameter estimation. The AVR is basically made up of the central retinal artery equivalent known as (CRAE) and the central retinal vein equivalent known as (CRVE), as well as the estimates of the arteriolar or venular diameter during the entrance of the vessels into the retina via the OD. The vein and arteries located away from the centre of the OD about 1 and 1.5 disc diameters. They are used for the iterative computation of the estimates. Research efforts have been made in terms of the formula to enable the calculation of the diameter equivalents.

Retinal vessels are effectively and non-invasively images utilizing fundus camera devices. Developing prove involving longitudinal prove that proposes morphological changes in retinal vessel parts are early physiomarkers of cardio-meta-bolic chance and result like any disease. Thus, information from expansive populace-based-related works is required to look at the nature of these morphological affiliations. Many systems have been used for retinal image investigation. Whereas these provide a number of retinal vessel lists, they are regularly limited within the range of investigation and quantitative measurement, and have restricted computeriza-tion, counting the capacity to recognize between venules and arterioles. Consequently, creating trustworthy approach for retinal image examination computer program and produc-ing a wealthy measurement of retinal vasculature in huge volumes of fundus cases.

3 Retinal fundus medical images

The retina is the layered part of the human eye which is sensitive to light and it receives images that are framed by the lens focal point and transferred to the brain via the optic nerve. An image of the retinal fundus is described with the inner lining of the ball involving the optic disc, retina, mac-ula and retinal vessel tree. The fundus is an internal part of the eye that can be seen through the pupil while the eye is being examined. By means of the fundus retinal images, the major vessel features as well as the vascular structure can be studied with several tools and applications (Zhang et al. 2010). A wide range of retinal cameras is available for capturing the images of the retina, depending on what is required for the medical research. There are two main types of cameras used for capturing the images of the reti-nal fundus and they are non-mydriatic and mydriatic retinal image cameras. To capture the retinal fundus image using the mydriatic camera, dilation drops must be applied to the patient’s eyes so that the retina can be dilated. This cam-era is often employed in situations when the patient’s pupil is ≤ 4 mm. On the other hand, the non-mydriatic camera is the most commonly used camera for capturing the images of the retinal fundus. Examples of some of the widely identi-fied non-mydriatic cameras are the Non-Mydriatic Topcon TRC-NW6S (Topcon America Corp., Paramus, EEUU) and Canon CR6-45NM (Canon USA, Lake Success, EEUU) as presented in Fig. 2. With the use of these cameras, ultra-high-resolution images can be instantly obtained by optom-etrists and ophthalmologists (Mohammed 2017a).

Fundus retinal images are almost classified into digital red-free photography, digital color fundus photography, and digital fluorescein angiography (see Fig. 3):

The digital fluorescence angiography is an imaging system, which involves image-taking through the dyestuff



1 3

of 32 20

tracing technique. It encompasses injecting sodium fluo-rescing into the circulation system. Subsequently, the emit-ted fluorescence helps to illuminate the retina with blue light at a wavelength of 490 nm which is photographed to attain what is known as angiography. After 12–24 h, the reappearance of the fluorescing dye is observed in the urine of the patient, which in turn leads to a yellow-green appearance. To detect pathologic changes like tumours, abnormal vessels, staining, or diabetic retinopathy, an evaluation of the dye patterns is conducted by the ophthal-mologist. Nevertheless, there are some high-risk adverse effects that are associated with the use of dye, which is also regarded as the most parasitical mechanism for the patient (Poplin 2018).

The digital color fundus retinal photography involves mounting a microscope’s customized camera with convo-luted mirrors as well as lenses to capture the image. The design of the high-powered lenses is made in a manner that allows the ophthalmologist to visualise the eye’s rear by making light more focused through the cornea, pupil and lens. With the use of fundus photography, the health condi-tion of the macula, optic nerve, vitreous, retina and blood vessels can be evaluated as seen in Fig. 4.

To capture the digital red-free photography, an unrealiz-able infrared light is used to obtain retinal illumination dur-ing collocation and focusing. The aim of this is to prevent the blinding white light experienced by the patient during the image capturing process. A white Zenon flash is used to

Fig. 2 Above are two kinds of cameras which produce a non-mydriatic retinal image: (Left) Canon CR6-45NM and (Right) Topcon TRC-NW6S

Fig. 3 Kinds of digital images for retinal: (from left to right) digital fluorescing angiography, color fundus photography and red-free photogra-phy

Fig. 4 (Left) CCD a camera for retina and (Right) sample of a retinal image captured from the eye



1 3

20 of 32

capture the images. However, this method is also accompa-nied by adverse effects and intrusiveness which should be considered depending on the tool or application. Due to the less intrusive nature of obtaining the red-free images and digital color fundus images, this method is selected for use in this study.

Due to several transmutation and measurement appara-tus that are designed to work on grayscale, color images must be subjected to transformation. The pixel intensities of color images are defined through three line methods: blue, red and green (RGB) so that the synoptic composite can be produced. For the additional color information of RGB images to be conveyed, a 3-D array is required. Mostly, the green channel is used in the analysis of fundus image since the difference between the vessel features and background is larger than that of other channels. Numerous imperative eye diseases and infections as well as systemic maladies show themselves within the retina. Whereas a number of other anatomical structures contribute to the method of vision, this study focused on image analysis, retinal imaging, and image investigation. Through this study, phases of image acquisition, image analysis, and clinical relevance are aimed together considered their commonly inserted relationships. Medical imaging is a really imperative premise for the iden-tification and patient treatment. In specific in ophthalmol-ogy photographs of the eye foundation are utilized by the specialists to analyze, diagnose and report diseases such as diabetic retinopathy or glaucoma. Additionally, in the medi-cal images are regularly encourage assessed by automated computer program to supporting the diagnostics process.

4 Retinal vessel segmentation techniques

The aim of identifying and localizing the retinal vessels is to distinguish the retina’s diverse vasculature structure tissues (which could be tight or wide) from the background of the fundus image as well as other retinal anatomical structures like abnormal lesions, macula, and optic disc. The attention of researchers is continuously focused on the area of retinal vessels identification, because of the presence of non-inva-sively fundus imaging techniques and the important details that are obtainable from the vasculature structure for the recognition and diagnosis task of a wide range of retinal pathology such as age-related macular degeneration (AMD hypertension, diabetic retinopathy (DR) and glaucoma. Recently, there has been an advancement in the development of innovative computer-aided techniques for the segmenta-tion of retinal vessels, and in recent times, the application of these techniques is justified in clinical applications. Oph-thalmologists obtain relevant information from the retinal vasculature structure (RVS), which helps them to detect and diagnose of a wide range of retinal pathology like diabetic

retinopathy, glaucoma, age-related macular degeneration and retinopathy of prematurity. Ophthalmologists also use such vital information to diagnose diseases associated with the heart or brain, which are linked to non-standard differences in retinal vascular shape. Thus, the variations that occur in venules morphology and the retina’s arterioles are of iden-tification assessment as they are key pointers of certain abnormalities.

Generally, vessel segmentation is one of the major areas in medical image segmentation (Lesage 2009; Kirbas 2004) and the retinal vessel segmentation falls under this category. Within the context of retina vessels segmentation, there are many methodologies and algorithms that are improved and applied for the automated localization method, segmenta-tion technique and RVS for feature extraction (Fraz 2012; Aparna and Rajan 2017; Arunkumar 2018). In the current study, a review of recent and early literature has been made, covering the techniques and methodologies that have been proposed for detecting and segmenting retinal vasculature shapes in 2-D retinal fundus cases. In the previous studies, the theoretical underpinning of each category of segmenta-tion is presented, as well as the benefits and limitations of each category. In general, there is a variety of algorithms and techniques available for segmentation, because situations and cases vary. However, all the techniques of retinal seg-mentation share the same stages, which are pre-processing, processing and post-processing tasks. The papers reviewed in our study are classified according to the technique or algo-rithm utilized in the processing phase, resulting in six main groups, which include the following: (1) machine learning (2) kernel-based techniques; (3) the multi-scale methods; (4) the model-based methods; (5) the adaptive local threshold-ing techniques; and (6) the mathematical morphology-based techniques. These six classifications are further clustered into two main categories (machine learning techniques or rule-based techniques) as presented in Fig. 5 and character-ized in Table 1.

There are particular rules that must be followed in an algorithm outlined in the group of rule-based techniques. On the other hand, in the machine learning category, the pre-segmentation retinal case (gold standard or ground truth) is used to create a labelled dataset which is utilized dur-ing the process of training. Nevertheless, when the prob-lem of image analysis is faced by a non-image processing specialist, he/she quickly understands that an independent image processing technique or a one image transformation technique is often not the way out. Thus, this is represented in Fig. 5, where the hybrid nature of these techniques is denoted by nested lobes. A majority of the problems associ-ated with image analysis are complex, especially, medical ones. However, these problems can be solved by combining a variety of basic techniques and transformations to achieve high-performance hybrid. It is for these reasons that the use



1 3

of 32 20

of hybrid techniques in solving problems associated with retinal vessels localization becomes essential as suggested in (Yang et al. 2008).

There are several metrics available for the assessment of retinal segmentation algorithm in terms of its ability to effi-ciently extract the RVS. In this regard, the commonly used metrics include average precision, average accuracy, average true-positive rate (TPR), average sensitivity (recall, TPR), and average false-positive rate (FPR). Of all the aforemen-tioned metrics, the most commonly used ones in the area of medical research are specificity and sensitivity. When higher values of specificity and sensitivity are achieved, better diag-nosis can also be achieved. Sensitivity is indicative of the algorithm’s ability to detect vessels’ pixels, while specificity indicates an algorithm’s ability to detect non-vessel pixels. These two metrics are features of a method, and they are related to the accuracy factor in several medical imaging areas, which include retinal vessels segmentation (Kauppi et al. 2007). The evaluation metrics are used based on the equations presented below (Walter et al. 2002):

in which TP is known as true positives, FP is known as false positives, FN is known as false negatives, and TN is known

(1)Accuracy = (TP + TN)∕(TP + FN + FP + TN),

(2)Sensitivity (Recall) = TP∕(TP + FN),

(3)Specificity = TN∕(TN + FP),

(4)Pr ecision = TP∕(TP + FP),

as true negatives. In contrast, other researchers evaluated their works using the receiver operating characteristic (ROC) curve (Abd Ghani 2018; Mohammed 2017b), a particular approach that strongly determine certain variables through-out the segmentation process. For FPR and TPR values, the ROC curve becomes a non-linear function. For the best per-formance, the best area under the ROC can be 1.

A majority of the available algorithms and techniques for the segmentation of retinal vessels employ the use of the most popular datasets in the field, which include (1) structur-ing analysis of the retina (STARE) (Staal et al. 2004; Nie-meijer et al. 2011) and (2) digital retinal image for vessel extraction (DRIVE) (Hoover et al. 2000). These databases are widely accepted and used in the area of retinal vessel segmentation tasks and classification process, such that the evaluation of all studies of segmentation process is made using these databases. Furthermore, the high resolution of the retinal fundus images that can be obtained using these datasets makes them popular. Their popularity is also based on the accessibility of physically labeled ground truth cases arranged through two experts in the field. There are 40 reti-nal images contained in the DRIVE dataset, and these 40 images are divided into two equal parts; the first part is for training, while the second part is for testing. On the other hand, there are 20 images contained in the STARE data-set, of which 10 are normal images of the retina, while the remaining 10 are abnormal images. However, a good number of researchers prefer to use other datasets that are not so popular for validation and evaluation of the performance of their proposed techniques or algorithms. Some of such datasets include High-Resolution Fundus (HRF) (Bankhead

Fig. 5 Retinal vessels segmentation techniques



1 3

20 of 32

Tabl

e 1

Crit

ical

revi

ew o

f app

roac

hes f

or th

e cl

assi

ficat

ion

of re

tinal

ves

sel s

egm

enta

tion

tech

niqu

es d

epen

ding

on

the

med

ical

imag

es

Aut

hor(

s)/y

ear

Imag

e pr

oces

sing

met

hods

/tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n

Cha

udhu

ri et

al.

(198

9)O

ptic

al a

nd sp

atia

l pro

perti

es–

–K

erne

l-bas

edC

hanw

imal

uang

(200

3)Lo

cal e

ntro

py d

epen

ding

on

thre

sh-

oldi

ng, M

atch

ed fi

lterin

g, v

ascu

lar

inte

rsec

tion

dete

ctio

n, a

nd le

ngth

fil

terin

g

–ST

AR

E–

Al-R

awi e

t al.

(200

7)M

odifi

ed p

aram

eter

s with

Gau

ssia

n m

atch

ed fi

lter

ROC

cur

ve m

etric

DR

IVE

–

Vill

alob

os-C

asta

ldi e

t al.

(201

0)Th

e G

LCM

and

seco

nd lo

cal e

ntro

pyA

ccur

acy,

spec

ifici

ty;,

sens

itivi

tyD

RIV

E–

Zhan

g (2

010)

Firs

t-ord

er D

eriv

ativ

e G

auss

ian

with

Th

e M

atch

ed F

ilter

Acc

urac

y, fa

lse

posi

tive

rate

(FPR

)D

RIV

E, S

TAR

E

Zhu

(201

1)C

hara

cter

ize

the

inte

nsity

dist

ribu-

tions

by

Piec

e-w

ise

Gau

ssia

n sc

aled

sc

hem

e

––

–

Kum

ar e

t al.

(201

6)G

abor

filte

r and

thre

shol

ding

with

G

LCM

ROC

cur

ve m

etric

DR

IVE,

STA

RE

–

Ods

trcili

k (2

013)

Enha

nced

mat

ched

filte

r and

min

i-m

um e

rror

thre

shol

ding

met

hod

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityD

RIV

E, S

TAR

E

Zolfa

ghar

nasa

b (2

014)

Mat

ched

filte

r with

Cau

chy

distr

ibu-

tion

and

kern

el fu

nctio

nA

ccur

acy ,

fals

e po

sitiv

e ra

te (f

pr)

DR

IVE

Sing

h an

d Sr

ivas

tava

(201

6)Lo

cal e

ntro

py th

resh

oldi

ng w

ith

Mod

ified

Gau

ssia

n m

atch

ed fi

lter

Acc

urac

y, sp

ecifi

city

;, se

nsiti

vity

DR

IVE

Kum

ar e

t al.

(201

6)G

auss

ian

kern

el w

ith C

ontra

st Li

mite

d A

dapt

ive

Hist

ogra

m E

qual

izat

ion

Acc

urac

y, sp

ecifi

city

;, se

nsiti

vity

DR

IVE,

STA

RE

Sing

h an

d Sr

ivas

tava

(201

6)M

atch

ed fi

lter w

ith G

umbe

l pro

babi

l-ity

dist

ribut

ion

Acc

urac

y, re

ceiv

er o

pera

ting

char

ac-

teris

tics

DR

IVE,

STA

RE

Chu

tata

pe e

t al.

(199

8)G

auss

ian

and

Kal

man

filte

rs w

ith

mat

ched

filte

r–

–Ve

ssel

trac

king

Sofk

a (2

006)

Mat

ched

filte

r with

ves

sel c

onfid

ence

, bo

unda

ries m

easu

rem

ents

for V

esse

l tra

ckin

g

The

reca

ll an

d pr

ecis

ion

DR

IVE,

STA

RE

Ade

l et a

l. (2

009)

The

SMF

7 Si

mul

ated

for B

ayes

ian

vess

el tr

acki

ngSe

gmen

tatio

n m

atch

ing

fact

or m

etric

20 fu

ndus

imag

es h

as b

een

colle

cted

Wu

et a

l. (2

007)

Hes

sian

mat

rix w

ith m

atch

ed fi

lters

fo

r Ves

sel t

rack

ing

Sens

itivi

ty, f

alse

pos

itive

rate

(FPR

)D

RIV

E, S

TAR

E

Yedi

dya

(200

8)K

alm

an fi

lter f

or V

esse

l tra

ckin

gFa

lse

nega

tive

rate

(FN

R).

True

neg

a-tiv

e ra

te (T

PR)

DR

IVE

Yin

et a

l. (2

010)

Vess

el tr

acin

g by

Sta

tistic

al m

etho

dFa

lse

posi

tive

rate

(FPR

)Tr

ue n

egat

ive

rate

(TPR

)D

RIV

E

Li e

t al.

(201

3)B

ayes

ian

theo

ry fo

r Ves

sel t

rack

ing

––

–D

e et

al.

(201

5)M

athe

mat

ical

gra

ph th

eory

with

Ves

-se

l tra

ckin

gG

eom

etric

fals

e po

sitiv

e ra

te (G

FPR

)D

RIV

E, S

TAR

E



1 3

of 32 20

Tabl

e 1

(con

tinue

d)

Aut

hor(

s)/y

ear

Imag

e pr

oces

sing

met

hods

/tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n

Bud

ai e

t al.

(201

0)M

ulti-

scal

ing

with

Gau

ssia

n py

ram

idA

ccur

acy,

spec

ifici

ty;,

sens

itivi

tyD

RIV

E, S

TAR

EM

ulti-

scal

eM

oghi

mira

d et

al.

(201

0)W

eigh

ted

med

ialn

ess f

unct

ion

with

M

ulti-

scal

eRe

ceiv

er o

pera

ting

char

acte

ristic

s, ac

cura

cyD

RIV

E, S

TAR

E

Abd

alla

h et

al.

(201

1)A

niso

tropi

c di

ffusi

on w

ith M

ulti-

scal

eRe

ceiv

er o

pera

ting

char

acte

ristic

sST

AR

ER

atta

than

apad

et a

l. (2

012)

Line

prim

itive

s for

Mul

ti-sc

ale

Fals

e po

sitiv

e ra

te (F

PR)

DR

IVE

Kun

du (2

012)

Scal

e-sp

ace

base

d on

Mor

phol

ogic

al

Ang

ular

Mea

ns sq

uare

err

orD

RIV

EM

orph

olog

ical

bas

ed m

etho

d

Fruc

ci (2

014)

Dire

ctio

nal M

aps,

Wat

ersh

ed tr

ans-

form

and

Con

trast

Acc

urac

y, p

reci

sion

DR

IVE

Jiang

(201

7)M

orph

olog

ical

ope

ratio

ns w

ith G

loba

l th

resh

oldi

ngA

ccur

acy,

exe

cutio

n tim

eD

RIV

E, S

TAR

E

Diz

daro

et a

l. (2

012)

Edge

det

ectio

n an

d Le

vel s

et in

itial

i-za

tion

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityD

RIV

E Pr

opos

ed D

atas

etD

efor

mab

le m

odel

Zhan

g et

al.

(201

5)Sn

akes

con

tour

sA

ccur

acy,

spec

ifici

ty; s

ensi

tivity

DR

IVE

Zhao

(201

5)H

ybrid

regi

on in

tere

st an

d ac

tive

cont

our

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityD

RIV

E, S

TAR

E

Gon

gt e

t al.

(201

5)Lo

cal r

egio

n in

tere

st w

ith L

evel

set

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityD

RIV

EJia

ng (2

003)

Ada

ptiv

e th

resh

oldi

ng fo

r Kno

wle

dge-

guid

ed lo

cal

Fals

e po

sitiv

e ra

te, f

alse

pos

itive

rate

, fil

ter r

espo

nse

anal

ysis

STA

RE

Ada

ptiv

e lo

cal t

hres

hold

ing

Akr

am e

t al.

(200

9)A

dapt

ive

thre

shol

ding

for S

tatis

tical

m

etho

dsA

ccur

acy,

rece

iver

ope

ratin

g ch

arac

-te

ristic

s (ro

c)D

RIV

E

Chr

istod

oulid

is (2

016)

Mul

ti-sc

ale

base

d on

Loc

al a

dapt

ive

thre

shol

ding

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityEr

lang

en D

atas

et

Nek

ovei

(199

5)B

ack

prop

agat

ion

neur

al n

etw

ork

Sens

itivi

tyM

achi

ne le

arni

ngSa

lem

and

Nan

di (2

006)

K-n

eare

st ne

ighb

ors

Spec

ifici

ty; s

ensi

tivity

STA

RE

Xie

(201

3)Fu

zzy

c-m

eans

and

Gen

etic

Alg

o-rit

hm–

DR

IVE

Akh

avan

(201

4)Fu

zzy

c-m

eans

for t

he V

esse

l Tra

ck-

ing

Acc

urac

yD

RIV

E, S

TAR

E

Emar

y et

al.

(201

4)C

ucko

o se

arch

met

hod

and

fuzz

y c-

mea

ns a

ppro

ach

Acc

urac

y, sp

ecifi

city

;, se

nsiti

vity

DR

IVE,

STA

RE

Maj

i et a

l. (2

015)

Ran

dom

fore

st te

chni

que

with

Dee

p ne

ural

net

wor

k ap

proa

chA

ccur

acy,

rece

iver

ope

ratin

g ch

arac

-te

ristic

s (RO

C)

DR

IVE

Gu

(201

5)D

ecis

ion

tree

Acc

urac

yD

RIV

E, S

TAR

ESh

arm

a (2

015)

Fuzz

y Lo

gic

met

hod

Acc

urac

yD

RIV

ERo

y (2

015)

The

auto

-enc

oder

neu

ral n

etw

ork

Rece

iver

ope

ratin

g ch

arac

teris

tics

(roc)

DR

IVE,

STA

RE

Lahi

ri et

al.

(201

6)Th

e au

to-e

ncod

er n

eura

l net

wor

kA

ccur

acy

DR

IVE

Maj

i et a

l. (2

016)

Ense

mbl

e of

12

CN

NA

ccur

acy

DR

IVE



1 3

20 of 32

Tabl

e 1

(con

tinue

d)

Aut

hor(

s)/y

ear

Imag

e pr

oces

sing

met

hods

/tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n

Lisk

owsk

i (20

16)

CN

NRe

ceiv

er o

pera

ting

char

acte

ristic

sD

RIV

E, S

TAR

ELi

skow

ski (

2016

)D

CN

NA

ccur

acy,

rece

iver

ope

ratin

g ch

arac

-te

ristic

s (RO

C)

DR

IVE,

STA

RE

Fraz

et a

l. (2

012)

Dec

isio

n tre

eA

ccur

acy

DR

IVE,

STA

RE

and

CH

ASE

dat

aset

Ense

mbl

e cl

assi

ficat

ion

-bas

ed m

etho

dD

asgu

pta

2017

)C

NN

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ity;

rece

iver

ope

ratin

g ch

arac

teris

tics

(RO

C)

DR

IVE

Fu (2

016)

Con

ditio

nal r

ando

m fi

eld

and

conv

o-lu

tiona

l neu

ral n

etw

ork

base

d on

m

ulti-

scal

e

Acc

urac

y, sp

ecifi

city

; sen

sitiv

ityST

AR

E,D

RIV

E, a

nd C

HA

SE1

Mac

hine

lear

ning

Gao

et a

l. (2

017)

U-n

et a

nd G

auss

ian

mat

ched

filte

rPr

ecis

ion,

spec

ifici

ty se

nsiti

vity

, and

ac

cura

cyD

RIV

EM

achi

ne le

arni

ng

Zhu

(201

7)39

-D d

iscr

imin

ativ

e fe

atur

e ve

ctor

s an

d Ex

trem

e Le

arni

ng M

achi

neA

ccur

acy,

sens

itivi

ty, a

nd sp

ecifi

city

DR

IVE

Mac

hine

lear

ning

Bra

ncat

i (20

18)

CN

N in

clud

ing

dire

ctio

nal fi

lters

Acc

urac

y, se

nsiti

vity

, and

spec

ifici

tyD

RIV

EM

achi

ne le

arni

ngH

u (2

018)

The

CN

N w

ith e

nhan

ced

cros

s-en

tropy

loss

func

tion

Acc

urac

yST

AR

E an

d D

RIV

EM

achi

ne le

arni

ng

Zhan

g (2

018)

U-n

et w

ith re

sidu

al c

onne

ctio

n an

d ed

ge-a

war

e m

echa

nism

Acc

urac

y, a

nd A

UC

ST

AR

E, D

RIV

E, a

nd C

HA

SE1

Mac

hine

lear

ning

Soom

ro (2

018)

Inde

pend

ent c

ompo

nent

ana

lysi

sA

ccur

acy,

sens

itivi

tyST

AR

E an

d D

RIV

EH

ajab

dolla

hi e

t al.

(201

8)C

NN

s with

com

bina

tion

of p

runi

ng

and

quan

tizat

ion

Acc

urac

yST

AR

EM

achi

ne le

arni

ng

Guo

(201

8)Re

info

rcem

ent s

ampl

e le

arni

ng w

ith

CN

Ns

Acc

urac

y, a

nd a

ucST

AR

E, a

nd D

RIV

E,M

achi

ne le

arni

ng

Wu

(201

8)M

ultis

cale

net

wor

k fo

llow

ed n

etw

ork

Acc

urac

yD

RIV

E, a

nd C

HA

SE1

Mac

hine

lear

ning

Soom

ro (2

019)

CN

N w

ith P

rinci

pal C

ompo

nent

A

naly

sis a

nd T

he m

orph

olog

ical

m

appi

ngs

Acc

urac

y, se

nsiti

vity

STA

RE,

DR

IVE,

and

CH

ASE

1M

achi

ne le

arni

ng

Dha

rmaw

an (2

019)

New

dire

ctio

nally

sens

itive

and

CN

NM

athe

ws c

orre

latio

n co

effici

ent,

F1-s

core

, sen

sitiv

ity a

nd g

-mea

nST

AR

E, a

nd D

RIV

EM

achi

ne le

arni

ng

Jiang

et a

l. (2

019)

Dila

ted

Mul

ti-Sc

ale

CN

NA

ccur

acy,

sens

itivi

ty, s

peci

ficity

, and

ro

cST

AR

E, D

RIV

E, a

nd C

HA

SE1

Mac

hine

lear

ning

Xiu

qin

et a

l. (2

019)

Dee

p le

arni

ng U

-Net

mod

elA

ccur

acy,

sens

itivi

ty, s

peci

ficity

DR

IVE

Mac

hine

lear

ning

Jin (2

019)

DU

Net

with

U-s

hape

arc

hite

ctur

eA

ccur

acy,

and

AU

C

STA

RE,

DR

IVE,

CH

ASE

1, W

IDE

and

SYN

THE

Mac

hine

lear

ning

+ de

form

able

mod

el

Bin

h (2

019)

Sobe

l ope

rato

r con

ditio

n an

d sa

lient

re

gion

com

bine

dA

ccur

acy

STA

RE

Vess

el tr

acki

ng

Hat

amiz

adeh

et a

l. (2

019)

CN

N a

nd e

ncod

er-d

ecod

er a

rchi

tec-

ture

Acc

urac

yD

RIV

E an

d C

HA

SE1

Mac

hine

lear

ning



1 3

of 32 20

Tabl

e 1

(con

tinue

d)

Aut

hor(

s)/y

ear

Imag

e pr

oces

sing

met

hods

/tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n

Lv e

t al.

(202

0)A

trous

con

volu

tion(

AA

-UN

et) a

nd

guid

ed U

-Net

Acc

urac

y, a

nd A

UC

ST

AR

E, D

RIV

E, a

nd C

HA

SE1

Mac

hine

lear

ning

Zou

et a

l. (2

020)

Loca

l Reg

ress

ion

for M

ulti-

labe

l Cla

s-si

ficat

ion

mod

elA

ccur

acy

STA

RE,

DR

IVE

Mac

hine

lear

ning

Wu

(202

0)N

etw

ork

follo

wed

net

wor

kA

ccur

acy,

and

AU

C

STA

RE,

DR

IVE,

and

CH

ASE

1M

achi

ne le

arni

ngM

ou (2

021)

U-N

et w

ith fu

nctio

nal b

lock

sA

ccur

acy

DR

IVE

Mac

hine

lear

ning

Isav

and

Rah

man

i et a

l. (2

020)

Mor

phol

ogic

al re

cons

truct

ion

and

Gab

or fi

lter

Acc

urac

yST

AR

E, D

RIV

E

Zhou

et a

l. (2

020)

Hid

den

Mar

kov

mod

el a

nd w

eigh

ted

line

dete

ctor

Acc

urac

y, se

nsiti

vity

, spe

cific

ity,

feat

ure

sim

ilarit

y in

dex,

dic

e co

effi-

cien

t and

stru

ctur

al si

mila

rity

inde

x

STA

RE,

DR

IVE

Mac

hine

lear

ning

+ de

form

able

mod

el

Fran

cia

et a

l. (2

020)

Resi

dual

U-N

et W

ith a

U-N

etRe

call

and

f1-s

core

STA

RE,

DR

IVE

Mac

hine

lear

ning

Roch

a (2

020)

Mor

phol

ogic

al o

pera

tions

, 2D

Gab

or

wav

elet

, and

CLA

HE

Acc

urac

yST

AR

E, D

RIV

E, a

nd H

RF

Saro

j et a

l. (2

020)

Fréc

het P

DF

base

d m

atch

ed fi

lter

appr

oach

Spec

ifici

ty, s

ensi

tivity

and

aver

age

accu

racy

STA

RE,

DR

IVE,

and

RM

SDM

achi

ne le

arni

ng

Mou

(202

1)N

ew c

urvi

linea

r stru

ctur

e se

gmen

ta-

tion

netw

ork

CS2

-Ne

The

inte

r-cla

ss d

iscr

imin

atio

n an

d in

tra-c

lass

resp

onsi

vene

ssST

AR

E, D

RIV

E, IO

STA

R, C

OR

N-1

, O

CTA

, OC

T R

PED

eep

lear

ning

Tian

yu M

a (2

021)

Ense

mbl

ing

Low

Pre

cisi

on M

odel

sD

ice

scor

e, re

call,

pre

cisi

onH

VSM

R, M

ICCA

I, M

RI

Dee

p le

arni

ngH

uang

et a

l. (2

021)

Dee

p ne

ural

net

wor

ks-b

ased

(DN

N-

base

d) m

odul

eRe

solu

tion

VIC

AVR

, VA

RIA

, STA

RE,

CH

ASE

-D

BI,

ROD

REP

,HR

FD

eep

lear

ning

Dan

, et a

l. (2

021)

Mul

ti-sc

ale

conv

olut

ion

Ker

nel U

-Nrt

mod

elSp

ecifi

city

, sen

sitiv

ity a

nd av

erag

e ac

cura

cyST

AR

E, D

RIV

EM

achi

ne le

arni

ng

Kha

ing

et a

l. (2

021)

AD

I-G

VF

Segm

enta

tion

with

EM

in

itial

izat

ion

Sens

itivi

ty a

nd av

erag

e ac

cura

cyM

ESSI

DO

R, D

RIO

NS-

DB

and

D

IAR

ET-D

B1

Mac

hine

lear

ning

Kam

ran

et a

l. (2

101)

Mul

ti-sc

ale

gene

rativ

e ad

vers

aria

l ne

twor

ksA

UC

D

RIV

E, C

HA

SE-D

B1,

and

STA

RE

Mac

hine

lear

ning

Saha

Tch

inda

(202

1)C

lass

ical

edg

e de

tect

ion

filte

rs a

nd th

e ne

ural

net

wor

kA

ccur

acy,

sens

itivi

ty, s

peci

ficity

e D

RIV

E, C

HA

SE a

nd S

TAR

EM

achi

ne le

arni

ng

Ash

win

(202

1)C

onvo

lutio

nal n

eura

l net

wor

k ba

sed

on th

e ite

rnet

arc

hite

ctur

eA

CC

D

RIV

E an

d SB

VPI

, REI

DA

-R,

REI

DA

-EE

Mac

hine

lear

ning

Ham

ad (2

020)

Fuzz

y C

-mea

ns c

luste

ring

Sens

itivi

ty, s

peci

ficity

, and

acc

urac

y(D

IAR

ETD

B0/

1, ID

RID

, and

e-o

ptha

Mac

hine

lear

ning



1 3

20 of 32

2012), (DIARETDB) dataset (Decencière 2014), (Messidor) dataset (Decencière 2014), and Automated Retinal Image Analyzer (ARIA) dataset (Odstrcilik 2013).

The larger part procedures that are presented within this review are assessed on the dataset STARE as well as the dataset DRIVE. These datasets have a small run of fundus images as training cases. The execution measures are com-puted especially on some morphological attributes cases. Some cases interior STARE and DRIVE do not offer the features related to the retinal picture, for example, the spo-radic foundation of gray-level values, irregularities in lumi-nance of inter/intra image, and differentiate float. The fun-dus captured totally different natural conditions and imaging rebellious for retinal vessel division is an open inquire about for creating the segmentation procedures. The supervised algorithms are much way better than the unsupervised algo-rithms with regard to their execution for vessel segmented methods. There is still room for improving an effective seg-mentation method to detect and diagnose the retinal mala-dies that are based on vasculature networking of the retinal images. In this study, several retinal vessel segmentation techniques are presented and discussed such as Ensemble classifiers, artificial neural network, Statistical learning, and clustering methods, while few techniques are discovered in the retinal vessels segmentation by blind signal separation, support vector machine, and hidden Markov model.

4.1 Kernel‑based methods

The kernel-based techniques depend on vessel pixels with scattered intensities to create a filter kernel that is capable of detecting the boundaries of RVS. The kernel-dependent techniques are usually employed at the image pre-processing phase of other retinal segmentation methodologies because

of their ability to improve the mapping of vessel boundaries. The use of one of several frameworks are suggested and applied in retinal vessels outlining for profile-based kernels that are constructed depending on the indication that the features of the retinal vessels can be described by the inten-sity distribution of retinal vessels. These features are used to create a map for the detection of vessels. The basic principle of a kernel-based technique is to compare the variations in the pixel’s intensity together with the retinal vessel cross-segment outline with a pre-defined pattern that serves the kernel. Thus, in the classic matched filter-dependent meth-ods, the matched filter kernel is applied on the grey retinal images and using the thresholding step to enable the detec-tion of retinal vessels.

There is a wide range of areas in vascular width meas-urement (Li et al. 2005) and vessel type classification (Saha Tchinda 2021) on which the retinal vessel profiling is applied. For the detection and extraction of vessels, it is used for the creation of a detection map, which in turn enables the extraction of vessels by means of filter-based approaches or region growing. There are two main catego-ries of retinal vascular, which are Gaussian-based shaped and non-Gaussian-based shaped (Ma 2015). The first study in this area is made by Chaudhuri et al. (1989), who found that with a Gaussian function the differences of the cross-segment outline of the retina are high, similar to that shown in Fig. 6.

In Chaudhuri et al. (1989), it is stated that the cross-segment outline of RVS has an estimated Gaussian shape. They also observe the emergence of coordinated filters with Gaussian kernels, which is discussed in earlier studies of retinal vessel tree recognition. Various shape of Gaussian fil-ters (diverse in Gaussian factors standards σ and µ) are used and employed in matching different vessel sizes in a simple

Fig. 6 Cross-segment intensity outline



1 3

of 32 20

and efficient way because these filters allow the modelling of the retinal vessels cross-segment as a Gaussian task. Nev-ertheless, the response of matched filters to both non-vessels and vessels structures is strong, thereby causing the perfor-mance to degrade in terms of false recognition ratio. There are three relevant factors that must be considered when a matched filter kernel is designed: (1) limited curvature of RVS, where bell-shaped piecewise linear segments are used to approximate the curvature of vessel segments; (2) vessel’s width, which gradually reduces as it traverses between the fovea area and optical disc as presented in Fig. 6; and (3) precise pixel intensity cross-segment outline scattering for the retinal blood vessel part (Singh et al. 2015).

In a study by Singh and Srivastava (2016), they use the same idea by Chaudhuri et al. (1989) to execute and apply using the DRIVE database. The segmentation outcomes, which are regenerated, show that the precision, specific-ity, and sensitivity of 0.9387, 0.9647 and 0.6721, respec-tively, are achieved. In another study, a profile kernel-based algorithm is proposed by (Zhu 2011) for the extraction of RVS based on outlining the cross-segmenting of retinal ves-sel. This is achieved through the use of part-wise Gauss-ian scaled scheme. Upon completion of the modelling, the application of a stage arrangement function is made based on the data obtained by oriented log-Gabor wavelet. Sub-sequent to the mapping of the boundary region of the RVS, the extraction of the cross-sections is made using a method used by the author in this field (Zhu 2010). In a study by Villalobos-Castaldi et al. (2010), a good performance of vessel extraction is achieved by combining matched filter with entropy and an adaptive thresholding technique. In this study, they use the DRIVE dataset, in which a matched filter is employed. Subsequently, the co-occurrence matrix tex-ture feature (Lenskiy 2010) that accounts for the quantity of changes among all sets of grey retinal levels is taken, depicting the changes of grey levels. Then a second-entropy thresholding is used to exploit the entropy of the fundus image grey level scattering so that the background pixels can be segmented from the vessels at the foreground. The vas-cular structures are obtained within approximately 3 s, and high detection accuracy is achieved with a value of 0.9759, 0.9648 and 0.9480 for the precision, sensitivity and specific-ity, respectively. The procedure used by Villalobos-Castaldi et al. (2010) is used by Chanwimaluang and Fan (2003) for the extraction of both the optic disc and retinal vessel. They used the STARE dataset in their study. Nevertheless, the time expended is approximately 2.5 min per retina case, and the larger part of this time is used during the processes of matched filtering and local entropy thresholding technique. In addition, the post-processing step, which involves lengthy filtering for the elimination of isolated pixel is required in their study, but not required in Villalobos-Castaldi et al. (2010). Subsequently, the retinal vascular intersection/

crossovers are identified through the application of the mor-phological thinning stage. Meanwhile, the identification of optic disc is made in two stages including (1) the detection of maximum local variance to facilitate the identification of optic disc center, and (2) the use of snake active contour for the identification of optic disc boundary. Despite the exact methodological steps followed by Odstrcilik (2013) Chan-wimaluang (2003) and Villalobos-Castaldi et al. (2010), the performance achieved is totally different.

In a study by Singh et al. (2015), the significant influ-ence of Gaussian kernel factors on the consequent stages of fundus image techniques is highlighted. The procedure proposed by Chanwimaluang (2003) and Villalobos-Castaldi et al. (2010) is used by Singh et al. (2015), who modify the Gaussian function so that the overall perfor-mance is enhanced. They report an accuracy value of 0.9459, a specificity of 0.9721 and a sensitivity of 0.6735 respectively, utilizing the DRIVE dataset in compari-son with the ratio of ROC metric of 0.9352 achieved by Al-Rawi et al. (2007) who also use the DRIVE dataset. However, in Al-Rawi et al. (2007), a group of changes is used for Gaussian-kernel factors. In another study, Kaur and Sinha (Kumar et al. 2016), use the procedure used by Chanwimaluang et al. (2010) and Singh et al. (2015), but rather than using the Gaussian filter, they use the Gabor filter at the initial step of vessel extraction. They obtain the enhanced vessels through sets of 12 various oriented Gabor filters ranging from 0 to 170 degrees. The perfor-mance of the Gabor-filter-based approach is better than that of Gaussian in specificity and ROC metric. However, the overall sensitivity achieved is more than that achieved by Chanwimaluang (2003) and Singh and Srivastava (2016). Using the DRIVE dataset, Singh and Srivastava (2016) and Singh et al. (2015) evaluate the performance of their work, while in (Frucci 2014), both the STARE with challenges of pathologies and DRIVE dataset are employed in the performance evaluation. Their results show that a high rate of specificity (96%) is achieved in the presence of abnormal parts in retinal fundus images with good detection. Zhang et al. (2010) apply two matched filters to a retinal image: the Gaussian cross-sectional symmetric profile of the retinal vessels, and the asym-metric cross-sectional profile of non-vessels. One of the matched filters results from the (zero-mean) kernel of the symmetric Gaussian construction, whereas the other one results from the Gaussian (FDOG) kernel construction of first-order derivative. The vessels are then detected using the existing matched filter response while the dynamic threshold is used to establish and adjust using the local mean of the response of first-order derivative of Gauss-ian kernel. The dynamic threshold is then utilized in the threshold stage and then applied in the filter stage. In the proposed method, the variance among the FDOG kernel



1 3

20 of 32

response for the vessels and non-vessels areas (for exam-ple lesions, bright blobs, and optic disc) is exploited with the aim of determining the level of thresholding based on local mean signal. The results of their experiments for STARE and DRIVE datasets reveal that false detection is significantly minimized through the application of hybrid matched filtering kernels. As compared to Gaussian ker-nel, the false detection is reduced to less than 0.05 using the hybrid matched filtering kernels, even for vessels that are thin, with an average of 0.9510 accuracy in the ordi-nary samples of retinal fundus cases, with the ratio 0.9439 for pathological cases.

Based on the discussion of techniques provided above, the majority of traditional matched filter-based approaches do not directly improve the rendering of matched filter-based approach; rather, they improve the rendering of the threshold techniques, which in turn evaluate the matched filter kernel. In the work by (Zolfagharnasab 2014), the CPDF Function is used in place of the Gaussian kernel of matched filter, and their results show that a total precision of 0.9170 is achieved with 3.5% FP rate through the use of the DRIVE dataset. In the algorithm proposed by (Kumar et al. 2016), the integral zero-crossing property of LoG filter is used, whereby the application of two filters for matched case with LoG ker-nel function is made for fundus images in the retinal case especially to facilitate the detection of RVS. The Contrast Limited Adaptive Histogram Equalization (CLAHE) tech-nique is used to improve the RVS. Their experimental results reveal that an accuracy of 0.9626 is achieved, whereas the sensitivity is 0.7006 and the specificity is 0.9871 for the DRIVE dataset. On the other hand, for the STARE data-set, the accuracy achieved is 0.9637, while the sensitivity is 0.7675 and the specificity is 0.9799. Their result is compared with that of (Odstrcilik 2013) who achieve a precision of 0.9340, a sensitivity of 0.7060 and a specificity of 0.9693 for the DRIVE dataset. However, for the STARE dataset the accuracy achieved is 0.9341, the sensitivity is 0.7847 and the specificity is 0.9512 by utilizing enhanced two-dimensional Gaussian kernel and two-dimensional matched filter. Singh and Strivastava (2016) introduce a new matched filter kernel, by suggesting the Gumbel PDF as the kernel function. They notice the trivial skewness of vessel-cross-segment outline is greatest by Gumble PDF as for Cauchy PDF and Gaussian kernels suggested in (Chaudhuri et al. 1989; Al-Rawi et al. 2007). The authors use the entropy-based optimal threshold-ing at the threshold level, while the use of length filtering is employed during the post-processing steps to enable the elimination of isolated pixels. An improved rendering by the proposed methodology achieve a recognition precision of 0.9522 for the DRIVE dataset and an accuracy ratio of 0.9270 for the STARE dataset. The ROC metric value is 0.9287 for the DRIVE dataset and 0.9140 for the STARE dataset.

4.2 Vessel tracing/tracking methods

These techniques are aimed at tracing the ridges of retina fundus medical images based on a group of facts. The pre-liminary stage of the tracking algorithm involves the selec-tion of seeds that are defined manually or automatically. To detect the vessels’ ridges, zero-crossing of the gradient and curvature are inspected. Nevertheless, a pre-processing step is required in “clean-limbed” ridges detection, in which complex steps are required for the enhancement of sizes and orientations of vessels. To this end, the high level of depend-ence on pre-processing steps limits the vessel tracking. In the tracing methods, it is unnecessary for the seed point process, which is the opening point of the tracking method at the middle of RVS. For example, in a study by Chutatape et al. (1998), the seed points are extracted from the circumference of the optic disc, then, an extended Kalman filter is used to trace the centers of the vessels. As an examining area for the starting point of vascular outline, a semi-ellipse is found around the optical disc. This idea is also used in a research by (Li et al. 2013). The location of the candidate pixels are chosen on the semi-ellipse, as such, the vessel is tracked based on the Bayesian theory method.

In a research by (Wu et al. 2007), a vessel tracking approach is proposed for the extraction of a retinal vascula-ture. According to the proposed methodology, the matched filters are combined with a knowledge that allows the edge details at the vessel’s matching borders to be exploited; an argument primarily used by (Sofka 2006) for the extraction RVS. Upon the enhancement of the contrast amid the vessels and other tissues of the retina and when the data of the ori-entations and sizes of the improved RVS are presented, the vessels are traced using the ridges through the centerlines beside the ridge seeds, which are automatically chosen. The researchers used the DRIVE dataset to test the performance of their methodology in terms of tracking, and their results show the successful detection of retinal vasculature skeleton with a FP ratio of 19.3%, and the preponderance of false tracked vessels are few vessels. Unlike Wu et al. (2007), the use of Kalman filters are employed in the study conducted by Yedidya and Hartley (Yedidya 2008), in which a pursuit approach is proposed for tracing the centers of the retinal vessels. With their proposed methodology, both thin and wide vessels are detected, even when the images of the retina contain noise; this is achieved by proposing a linear model. Their proposed model is made up of four stages. The first stage involves finding a set-seed exploring point through the image by coiling together the entire image of the retina with a group of matched filters at diverse orientations and scales. The aim of this is to find at least a single seed point at each vessel, thereby eliminating the need to go through all branches. The second stage involves the tracing of blood vessels’ centers using Kalman filter; the tracing begins from



1 3

of 32 20

the seed points that are found at the first stage. In the third stage, the tracking is stopped the moment the possibility of vessel pursuit is small for various consecutive moves. The fourth stage involves the pursuit of segmentation results when tracking fails in less than the minimum number of steps. With their proposed pursuit approach, they are able to achieve true positive rate of 85.9% for the detection of retinal vessels, and false positivity of 8.1% via the DRIVE datasets.

A new technique is designed by (De et al. 2015) for the extraction of filamentary retina structure. In their work, the mathematical graph theory is used, and the building of the technique is made by establishing a relation between the tracing issue and the digraph matrix-forest theorem in the algebraic theory subject. The purpose behind doing so is to pinpoint the vessel cross-over. There are two central stages involved in the proposed technique, which is the segmenta-tion stage and tracing stage. The segmentation stage involves the extraction of the main skeleton of the RVS, whereas, in the second stage, which is the tracing stage, the digraph representation is constructed, thereby enabling the trac-ing process to perform as digraph depending on the label propagation by means of the Matrix-forest theorem method. With the use of their implemented method, both neural and retinal tracing are performed, and the results of their empiri-cal evaluation show that the performance of the proposed technique is high in both cases. An enhanced version of the statistical-based tracking method suggested by (Adel et al. 2009), is presented by (Yin et al. 2010). This method allows the iterative detection of edge points based on a Bayesian approach through the use of continuity properties of blood vessels and local grey levels statistics. To improve the accu-racy and tracking efficiency of the method, the geometric properties of the vessel are combined with the grey level. The results of the experiments which are performed on both real and synthetic images of the retina (the DRIVE data-base) produced favorable outcomes, in which the TP rate is 0.73 and the FP rate is 0.039. Nonetheless, a more thorough evaluation of retinal images may be required, because of the comparatively low detection rate (TPR) achieved. This is to enable a broader usage of the proposed method for the detection of vessels.

4.3 Mathematical morphology‑based methods (MMBM)

The methods based on mathematical morphology are a group of theory that is derived from mathematics. It is viewed as the use of the lattice theory on spatial construc-tions. This theory is associated with the shapes found within the frame of the image, rather than the intensities of pixels. In other words, it does not focus on the information about the content of an image, where the intensities of pixels are regarded as topographical highs. Traditionally, the use of

mathematical morphology is employed in paired images, and later, it becomes a general processing framework. To this end, it is also applied to grey and colored images through morphological operators. In a typical scenario, the application of morphological procedures can be made to the fundus-paired images and expanded to grey images in the TVS. The MMBM processes are classified into open and close processes involving dilation and erosion. Erosion operation is employed in reducing objects present within the fundus, while dilation is utilized to increase the objects’ size. The MMBM openings are employed when the elimination of structures in the image is required, which involves the application of erosion tracked with the dilation operation. MMBM closing involves filling some of the outlines in the fundus image or merging them through the application of dilation process tracked with the erosion operation.

The Morphological Angular Scale-Space (MASS) approach is carried out by (Kundu 2012) for the purpose of segmenting retinal images. The central idea of this technique is to rotate the elements of different linear structuring lengths (multi-scale) at various angles to determine the components which are connected. The technique also helps to ensure that the vessels are connected, from which the creation of the scale-space occurs by means of the varying lengths of linear structuring elements.. The non-vessel elements, such as those obtained from processed retinal image, are lessened by gradual evolution to higher scales, which are built using the details extracted from lower scales. A certain scale case specified from an experimental outcome and utilizing the vesselness measurement is used by (Salem and Nandi 2006), having the lowest MSR value of 0.0363 as reported by the authors and this value is achieved over 50 fundus cases of retinal images obtained from the DRIVE database. Regard-less of the morphological procedures, the use of morpho-logical mechanisms is employed in the task associated with segmentation of retinal vessels: top-hat transform, geodesic distance, gradient, watershed transform, and distance func-tion. Watershed transmission is improved within the model of MMBM by (Salem and Nandi 2006). The underpinning ideology of this approach emerges from the geographical phenomena of flood water on earth. The presence of water-shed is seen in the form of separating lines created by rain falling over an entire area. Frucci et al. (2014) segment RVS through the use a watershed-based segmentation algorithm. In the proposed algorithm, watershed from both directional information and contrast obtained from the image of the retina are combined. First of all, the image is segmented into multi-regions using watershed transform. Then, each region is allocated a unique grey-level value. In every area, the variation in grey-level in terms of its adjacent areas is calculated to enable the computation of a contrast value. An indicator map consisting of 16 ways is obtained through the application of a 9 × 9 window. The pixels’ grey levels



1 3

20 of 32

with the standard deviation are aligned along these ways. Therefore, pixels that are situated in the same premises are allocated the same direction depending on the occurrences of ways within the watershed area. Then, the indicator map and the contrast of every watershed area are obtained and a precursory localization of the RVS is captured, where the areas possessing the maximum contrast (TP variance) are most possibly defined as non-vessel areas. Then, they are denoted as the vessel part. The DRIVE set of information is used to assess the work of the proposed algorithm, and the results show that a recognition accuracy of 77% and preci-sion of 95% are achieved.

With the aim of extracting the RVS, a new argument is proposed by (Jiang 2017), who used global threshold on the basis of morphological procedures. The DRIVE and STARE datasets are employed in the evaluation of the proposed system. The authors report that the system demonstrated a precision of 95.27% for the cross-dataset assessment and 95.88% for 1 dataset assessment. Considering computational complexity and time, the system is designed to minimize the computing complication and processing of multiple inde-pendent operations, which are paralleled, thereby achiev-ing an implementation time of 1.677 s for every case of the retinal images on the CPU platform to obtain the best execution time.

4.4 Multi‑scale‑based methods

The central notion of the manifold scale (multi-scale) method is to identify the retinal image at stages of multiple scales by integrating a datum that include a specific image with a set of one-factor set at multiple scales of derived images (Lindeberg 2011). The representation is valid since the outlines at stages of the coarse scales are the identical structures of simplified transmutations at fine scales with consideration to the small kernels. The individual con-ceivable small kernels that fit the spatial shift invariance and the linearity are Gaussian and its derivatives kernels, which increase with their widths (scales σ) (Babaud et al. 1986). Initially, the scale-space is the model of the multi-scale fundus case as an image demonstration. There are two kinds of multi-scale demonstration that are commonly used, which are the pyramid, and Quad-tree (Zhu 2010). The conception of a majority of retinal vessel segmentation methodologies is based on the pyramid multi-scale kind, in which the representation of the grey level details is made in a manner that allows cases processes to be combined with sequential smoothing phases. Using Gaussian kernels with varying scales leads to a reaction that is presented by (Zhu 2010). The vessel-likeness is determined by eigen values of a Hessian matrix provided by (Tankyevych et al. 2008), thereby resulting in the enhancement of RVS. The purpose of processing the Hessian matrix by means of analyzing

eigen values is to acquire the main tracks of the vessels; the local second-order outlines in the image of the retina can be decomposed to obtain the track of the lowest curvature along the RVS (Frangi et al. 1998). A rapid reduction occurs in the size of the retinal image with scale levels. This in turn leads to an increase in the number of computation needed. Nevertheless, it is limited to extracting heterogeneous struc-tures like retinal lesions and fixed-size structures like optic disc. Therefore, multi-scale approaches can be more appro-priate for structures possessing different length and width within the same image. Budai et al. (2010) propose a classic multi-scale based method for the retinal vessel segmentation. There are three main phases in the proposed technique: (i) Gaussian pyramid generation; (ii) analysis of neighborhood; and (iii) method of fusion for fundus images. Subsequent to the extraction of the green channel of raw images of the retina, a Gaussian pyramid of purpose hierarchy is produced. Its clear that there are three levels in the hierarchy, includ-ing, level 0, 1, and 2. The green channel that represents the original retinal image at level 0 for the highest resolution must be obtained, The height and width of the image start to decrease as the next level is approached.

The second phase involves the analysis of a 3 × 3 neigh-borhood window investigation for every level. To analyze each pixel, the Hessian matrix is calculated by computing a couple of eigenvalues λl and λh of the Hessian matrix which is indicative of the biggest one λh and the scale of smallest curvature λl of the target pixel of the neighborhood window. Then a vessel likeness measure Pvessel = 1 − λl is calculated using the ratio of these values; the value of Pvessel deter-mines whether the target pixel belongs to vessel tree or not. If Pvessel value is close to one, it means that it is most likely a vessel pixel since λl and λh are similar to each other. At every scale level, this analysis is carried out for each pixel. The final stage involves the use of two hysteresis thresholds to binarize the results of segmentation obtained from differ-ent levels, and then the use of pixel-wise OR operation is employed in combining the results, thereby creating the final segmented images as the outcomes. The proposed approach achieved an accuracy of 93.8%, a specificity of 97.5%, and a sensitivity of 65.1% for the STARE database. In addition, for the DRIVE database, the accuracy is 94.9%, the specificity is 96.8%, and the sensitivity achieved is 75.9%. In (Abdallah et al. 2011), a t multi-scale retinal vessel segmentation based on a two-step method is used. In the primary stage, denois-ing of the grey layer from corrupted retinal image is carried out beside the improved Gaussian noise. This is achieved through the application of a flux-dependent anisotropic dif-fusion method; the multi-scale response of multi-level reso-lution of the fundus retinal images is calculated. The second stage involves establishing a model for the analysis of Hes-sian matrix including the eigenvectors of every scale. The findings of the multi-level study are representative of the



1 3

of 32 20

pixel-wise maximum of the outcomes acquired from all the scales. The ROC curve metric achieved was 0.94514 for the STARE database as reported by the proposed method. Algo-rithm for the blood vessel segmentation in the retina images is used by (Rattathanapad et al. 2012). The segmentation in this algorithm is based on multilevel line detection and connection of line primitives. With the use of multilevel line detection, the images of the retina can be extracted as mul-tiple ratios of Gaussian smoothing factors. Then, extraction of the line primitives at various scales are integrated into the vessel feature extraction outcome. The DRIVE dataset is used for validating the proposed algorithm, which dem-onstrates the ability to identify most of the main measure of vessel skeleton with the FP ratios. Moghimirad et al. (2010) propose a novel approach for segmenting retinal vessel based on a multi-scale method. In the proposed approach, func-tion of weighted medialness integrated with the eigenval-ues of the Hessian matrix is utilized. There are two main phases in the proposed approach. The first phase involved the extraction of the medial axis of retinal vessels through the use of two-dimensional medialness function multiplied with smoothed eigenvalues. For the second stage, vessels are reconstructed, involving the extraction of vessels’ center-line and simultaneous estimation of vessels’ radii to acquire the outcomes of segmented image. The validation of the proposed technique is made using the DRIVE and STARE datasets, and the results reveal that the assessment of the method obtained a good accuracy of 0.9659 and an ROC curve metric of 0.9580 for the DRIVE database. For the STARE database, the accuracy achieved is 0.9756 and the ROC curve metric is 0.9678.

4.5 Methods based on model

The idea of deformable framework is utilized to define a lot of image processing algorithms and techniques, through which the difference of a given class of objects in an image are abstractly modeled. The main kinds of these methods are concerned with the modelling of shape variations, involv-ing the representation of the shape as a surface of a flex-ible curve, and then the shape is deformed so that it can be matched to a particular object class instance. The process of deformation is not random, because it functions based on two dominant theories, which are Curve Evolution and Energy Minimization, which originated from geometry, physics and approximation theories. There are two major categories of deformable models, which include geometric and parametric models (McInerney and Terzopoulos 1996).

4.5.1 Models of parametric deformable (PDM)

The PDM modeling is also referred to as snake or active contours that are defined as parametrized curves, which

naturally rely on certain factors to be produced. The PDM modeling like the active contour modelling is basically aimed at segmenting objects in fundus retinal images modalities by assigning curves on objects’ borders in the retinal images. This process can be referred to as dynamic contour modelling, because it is modified within a location in the region of the target object. In this way, dynamic evolution of the model can occur to fit the outline of object through an iterative revision. The reference to snakes arises from the use of parametric curve to represent a curve, but because it relies on a factor to effect the meas-ure of curves through the procedure of fitting, it is topo-logically rigid. This means that it is not flexible enough to define the objects that consist of an adaptable quantity of independent measures. In addition, another limitation of snake-based segmentation technique is that it is unable to join to precise vessel borders, when there is a high level of noise or if the contrast levels of the vessels are relatively low, or if the vessels are “empty” (McInerney and Terzo-poulos 1996).

Based on snake contours, Jin (2019) developed a new segmentation algorithm, which consists of three key stages: (1) first, the parameters initialization method based on the Hessian feature edges, in which the use of the feature is employed to extract all darker linear outlines in the retina fundus image. Depending on the seeds of extracted linear structure, the images used are divided into (N) localization areas (R); (2) the second stage involves representing each localization area R as retina fundus image using pixel’s intensity impact, and then, the construction of the function of snake energy is made on the image representation so that the snake’s location is identifed between the neighborhood of vessel edges and real ones; (3) in the last stage, as all model-based methods close, the region growing method utilizes R to obtain the outcome of retina fundus image area, and then post-processing is applied on the grown area through context feature. The validation of the proposed methodology is made using the DRIVE dataset, and the results show that the performance of the proposed method-ology is remarkable, achieving an accuracy of up to 95.21%, a sensitivity of 75.08%, and a specificity of 96.56%. Zhao (2015) propose an efficient and effective infinite perimeter active contour model with hybrid region for vessel segmen-tation. Fusion of region data of the retinal sample like the grouping of intensity data and local stage improvement map are used. The use of the local stage based on improvement map is employed to preserve the edges of vessels because of its superior nature, while a correct feature segmentation is indicated by the given details of the image intensity. The performance of the proposed method is evaluated for the DRIVE and STARE datasets, and the results show that accu-racy of 0.956, specificity of 0.978, and sensitivity of 0.780 are achieved for STARE, while for DRIVE, a sensitivity of



1 3

20 of 32

0.7420, a specificity of 0.9820, and an accuracy of 0.9820, are achieved.

4.6 Models of geometric deformable (GDM)

The Level-set and Fast Marching techniques are numerical methods specially tailored to track the propagation of sur-faces. The starting point of GDM originates from the inves-tigation of curves and surfaces that are regarded as interfaces and is first created and implemented by (McInerney and Terzopoulos 1996). Subsequently, it is suggested by (McIn-erney and Terzopoulos 1996) that the curve should be rep-resented as a level set based on Euclidean distance instead of parameter-dependency, meaning that zero-level set of an auxiliary distance function ∅ represents the contour. The basis of representing contours in a flexible way is the level sets theory; here reparameterization is not required for the joining or breaking of contours. A new level-set method that does not require an initialization for level-set function and a local region area descriptor are used by (Gongt et al. 2015). There are two main steps in the proposed technique which are described as follows: (1) a contour C is found, with which the entire image of the retina is divided into many sub-regions depending on whether the pixels are within the contoured area or not; and (2) a clustering algo-rithm is applied on the sub-regions obtained from the first step, thereby producing a local cluster value, which is a new region information that is used in redefining an energy function in the first step, until the algorithm converges. The major contribution of the paper lies in the second step, which eliminates the heterogeneity of pixel intensities of the retinal image. In addition, the second step helps in provid-ing information regarding the local intensity at image pixel level, thereby eliminating the need for level-set function re-initialization, which is regarded as the main limitation of level-set-based techniques. The DRIVE dataset is used to evaluate the proposed technique and the results show that the technique achieves values of 0.9360, 0.7078, and 0.9699 for accuracy, sensitivity, and specificity, respectively. Dizdaro et al. (2012) enhance the level-set-based segmentation of retinal vessels in terms of edge detection and initialization stages, which are required as a pre-requirement for level-set methods. Initialization steps involve the determination of the points of seed by sampling vessels’ centerlines that rely on identification of ridges via a ridge detection technique. Then a phase map is built to enable the determination of accurate boundaries of the retinal vessel tree. To test the pro-posed algorithm, the authors created their dataset, which is used together with the DRIVE dataset. Results show that the method achieves different ratios, with an accuracy of 0.9412, a sensitivity of 0.9743, and a specificity of 0.7181 for the DRIVE dataset, and an accuracy of 0.9453, a sensitivity of 0.9640 and a specificity of 0.6130 for the STARE dataset.

Conclusively, both geometric and parametric deformable models have the same problem, which is the requirement of seed points set that are automatically or manually detected.

4.7 Adaptive local thresholding techniques

A threshold technique is regarded as an image segmentation methodology that is popular and straightforward. In natural scene images, objects are arranged in a manner that makes them undistinguishable, but the objects such as tissues and organs in medical images are more discernible. Thus, extensive use of the thresholding techniques is employed in research that focus on the segmentation of medical image, in which different grey levels are used in representing the different organs and tissues. Basically, the thresholding techniques is employed in searching for a global value that enables the optimal maximization of the separation between varying classes in the image. At a global level, the effective-ness of thresholding is indicated by clearly defined areas. Another way to determine the effectiveness of thresholding is if the grey levels clustered around values that have the lowest interference. The global thresholding is regarded as a key source of faulty segmentation due to inferior qual-ity of source material, unequal illumination, distortion, and anatomical objects possessing multiple classes and hybrid characteristics. Furthermore, since the gradual transition between different grey levels can be noticed in the image of the retina, noise distortions, uneven illumination and other major segmentation errors become observable because of the pixel-wise approach employed in global thresholding. One of the key challenges of global thresholding is identi-fying the pixel segmentation that includes similar levels of grey (pixel intensity) for the same anatomical object. Thus, as a solution, it is suggested that the use of region-wise thresholding methodologies be employed in the identifica-tion of retinal vessels. It is also suggested that the develop-ment and implementation of such methodologies be made through different approaches that are categorized into three major groups including, fuzzy-based adaptive thresholding, statistical, as well as knowledge based.

In a work by (Christodoulidis 2016), the MTVF method is employed in the segmentation of small thin vessels, based on the fact that 10% of the entire surface of vas-cular network is made up of small retina vessels fundus (Niemeijer et al. 2011), which, in turn, is representative of the numerical model based on the adaptive threshold-ing method. There are four main stages in the proposed technique, which include the pre-processing; multi-scale line identifying vessel improvement; the adapted thresh-olding technique; and Multi-Scale Tensor Voting process-ing known as MTVF; and the post-processing step. The pre-processing stage involves the extraction of the raw image that includes the green channel of the retina. Then,



1 3

of 32 20

an application that corrects the contrast (Foracchia et al. 2005) is made to the image to enhance its contrast. The removal of noise is achieved through the application of Dual-tree complex wavelet transform. Subsequent to the pre-processing steps, the enhancement of retinal vessel fundus image is made using the multi-scale line recogni-tion method used in (Kingsbury 1998). The vessels are iso-lated by feeding the outcome of the multi-scale line indi-cator to an adaptive thresholding technique. The authors obtain different ranks of the improved thresholding by fit-ting the histogram of MSLD reaction to the modulus of the Gaussian function. Furthermore, they modify the optimum global threshold by varying the distance of Gaussian func-tion from the mean via the following equation:

where µGaussian and σGaussian denote the corresponding mean and standard deviation of the fitted Gaussian function. Then, different thresholds are obtained by experimentally varying the α variable, which is regarded as the nucleus of adaptive thresholding. Consequent to the completion of the adaptive thresholding stage, the separation of several smaller vessels occurs. Therefore, the smaller vessels are rejoined using a multi-scale tensor voting method introduced by (McInerney and Terzopoulos 1996). In conclusion, adaptive threshold-ing is employed in the extraction of large and medium-sized retinal vessels, while the smaller vessels are extracted using MTVF. Finally, a post-processing step is applied using morphological cleaning for the removal of the remnants of non-vasculature components left after the process of adap-tive thresholding. The authors test their method using the recently available Erlangen dataset (Odstrcilik 2013), and the results show that they achieve an average accuracy of 94.79%, a sensitivity of 85.06%, and a specificity of 95.82%.

In the study by (Akram et al. 2009), retinal vessels are automatically located and extracted using an adaptive thresholding technique. Through the selection of points to differentiate vessels from the remaining part of the image, a binary vascular mask is created using the statis-tical-based adaptive thresholding. There are two phases in the proposed method, and they are adaptive threshold-ing and pre-processing phases. The pre-processing phase involves feeding the Gabor wavelet filter with the mono-chromatic RGB retinal image to enhance the vasculature pattern, especially for vessels that are thin and less visible, based on an image examination method that is applied in (Christodoulidis 2016). The enhanced retinal image obtained holds maximum grey values for the background, while the pixels belonging to vessels holds intensity values that are greater than that of the background. The DRIVE dataset is used to test the proposed technique, and the results show that the accuracy achieved is 0.9469 and the

(5)T = |�Gaussian| + �|�Gaussian|,

ROC under the curve metric is 0.963. Jiang et al. (2019) propose an information discovery with full guide for local adaptive thresholding method, which uses a certification process based on multi-thresholding analytical system. In its most fundamental structure, assume a binary image, I, as the binary outcomes from a thresholding procedure at threshold levels T, from which the recognition process is utilized to choose if any of the areas in I is a well-defined object. The process is executed on a set of various thresh-olding methods. The last stage of image localization is achieved through a combination of various outcomes of diverse thresholding methods. In conclusion, binariza-tion is the process through which objects hypotheses in an image are generated through a number of hypothetic thresholds, and then the object is accepted or rejected based on a specific classification. The main contribution of the proposed algorithm is the classification procedure, in which any details like color intensity, shape and contrast related to the object under consideration is incorporated. Hence, the name knowledge-guided technique due to the variation in thresholding levels during the process of seg-mentation, which is based on the information available about a desired object.

4.8 Machine learning (ML) approaches

The utilization of ML is increasing rapidly in the field of pattern recognition, including computer-aided diagnosis and medical image analysis (Mohammed 2018). Since the 1960s, pattern recognition has gained popularity in the area of research (Mostafa 2019) and has made remarkable advance-ment over the years, resulting in a wide range of methods and paradigms that are relevant in several fields like retinal vascular structure. Typically, there are three main classifi-cation of machine learning algorithm, which include, rein-forcement learning, unsupervised and supervised learning. This classification is typically based on the reactions y for input data x. The supervised learning method usually occurs in situations when each input x, have an associated observ-able output y, while, in the case of the former two classes, the lack of data makes this correspondence absent. Unsuper-vised learning involves the exploration of relevant patterns in the input data without the need for explicit supervision (Arunkumar et al. 2018). The assumption in reinforcement learning is that there is a specific class of model that is fol-lowed by the dynamics of the system at hand (Arunkumar 2018). As a means of identifying the blood vasculature structure in angiogram images, Nekovei (1995) designed a back propagation ANN vessel technique. The technique uses raw grey-intensity values of pixels in place of feature extraction. To process each pixel in the image, a window that covers the number of pixels is created around the image, while the neural network is fed with the raw grey intensities



1 3

20 of 32

as input. The coverage of the whole window is achieved by sliding window process (pixel by pixel). The training data-set consists of angiogram patch samples that are manually selected, which shows equality between the vessel distri-bution and background pixels so that the biasing of neural network towards pixels classification is prevented. Using this technique, the difficulty associated with feature extrac-tion is avoided. The performance of this approach for vessel detection is 92% on angiograms. The investigation in domain adaptation and transfer learning is conducted by (Weiss et al. 2016) in the domain of retinal vessels fundus images segmentation task, in which a denoised stacked autoen-coder ANN is trained with sufficiently labeled smaller than expected patches of retinal vessels fundus images obtained from the DRIVE database. The DRIVE database is consid-ered as the main area, where the autoencoder is adjusted for arrangement on the STARE database, which represents the main field. There are two layers of encoding in the stacked auto-encoder with 100–400 nodes for every layer, respec-tively, and is followed by a layer of SoftMax regression. As a show of potent knowledge transfer, an accelerated learning performance is demonstrated by the proposed technique with area under ROC curve equal to 0.92. A hybrid framework of deep and ensemble learning is designed by (Maji et al. 2016) to enable a thorough detection of fine retinal vessels. In this framework, an advanced ML, Deep Neural Network (DNN), is employed for detecting vessel-ness based on unsupervised learning via a denoising autoencoder using inadequately pre-pared retinal vascular patches of fundus images. The learn-ing outline of retinal vasculature patches for fundus image is utilized as weights in a Deep Belief Network (DBN) and the reaction of the DBN is utilized in supervised learning task using the random forest technique with the end goal of vasculature tissues detection. A comprehensive exploitation of the denoising autoencoder is carried out in terms of its high capability. The DRIVE dataset is employed in training and testing the method. Regardless of an average accuracy of 0.9327 achieved by the proposed method, it is consid-ered as relatively weak compared with other state-of-the-art approaches. The uniqueness of the framework lies in its abil-ity to detect both fine and coarse retinal vascular structure, as well as its performance consistency. Based on the work of Maji et al. (2015) and Roy and Dasa 2015), Lahiri et al. (2016) demonstrate two parallel levels of stacked ensemble based on denoised autoencoder networks. A specific vessel orientation is distinguished by each kernel. In the first level of the ensemble, (n) denoising a parallel stacked via autoen-coders for the same architecture are trained, while the sec-ond level involves the implementation two stacked denoised autoencoders for parallel training, and then the fine-tuning of the last architecture is carried out until a satisfactory accu-racy is accomplished. The SoftMax classifier is employed in combining the decisions of individual members of the

ensemble. The method demonstrate consistency, reliability, and average accuracy of up to 0.953.

In the work by (Maji et al. 2015), vessel pixels are dis-tinguished from non-vessel pixels using DBN method with ensemble of 12 Convolutional Neural Network (CNN). There are three layers contained by each CNN and the training for each of these layers is made separately using a random selection of 60,000 patches including 31 × 31 × 31-sized obtained from 20 raw color retinal images from the DRIVE dataset. During the extraction of these patches, each convolutional network produce the probabilities of vessel-ness separately. Then the final vessel-ness probability of each pixel is formed by averaging the individual responses. Despite the inability to achieve the highest accuracy of detection (0.9470) by the method, its superiority is demon-strated through the presentation of learning vessel from data, because the strength and accuracy of numerous specialties represented by the ensemble of CNN is greater than that of a single neural network. In the study conducted by (Gu 2015), they focus on the detection and restoration of small foreground retinal filamentary structure through the use of an iterative two-steps method that is created based on the Latent Classification Tree (LCT) framework. Subsequent to the construction of the confidence map, an adequately high threshold is located on it, resulting in a partial image seg-mentation comprises of two major vessels; long and thick. The authors obtain the rest of the low confidence map (fila-mentary filaments) through the use of latent classification tree. The reconnection of the filamentary structure to the major filaments (large vessels) is made through a novel mat-ting and accomplishment domain method. These stages are carried out repeatedly until the entire surface is scanned. The method is tested on the DRIVE and the STARE data-bases, and the result shows a good accuracy in the detection stage of 97.32% for the DRIVE database and an accuracy of 97.72% for the STARE database, which is expected because of the obvious reduction in the FP formed via false identity of the right retinal RVS. The performance of the proposed technique is promising, and aside extraction of retinal ves-sels, the method can be applied in other areas like 3D MRI images and neural tracing procedure. An automated method, which rapidly and accurately segments retinal and optical disc outlines is used in (Liskowski 2016), in which CNNs are used for supervised segmentation. Just like for all NNs, the training of the CNN layers is carried out in a special-ized way so that both retinal and optic disc localization are addressed. The DRIVE and STARE datasets are employed in the validation of the proposed method for retinal ves-sel segmenting stage. Based on the results, the ROC curve metric achieved is 0.822 for the DRIVE database and 0.831 for the STARE database. Within the context of CNN-based methods, astounding execution has been accomplished by (Liskowski 2016) with an ROC metric of 0.99 and an



1 3

of 32 20

accuracy of 95.33%, whereas (Dasgupta 2017) achieved an AUC of 0.974 for automated retinal vessel segmentation. In the work of (Salem and Nandi 2006), a new enhancement is made to the typical KNN clustering technique for reti-nal vessel segmentation task. The feature vectors consist of the local maxima of the largest eigen value, green channel intensity, and the local maxima of the gradient magnitude that represents each pixel. Based on these features, a modi-fied version of KNN is used in clustering the image pixels without the use of a training set. The evaluation of the seg-mentation method is carried out for the STARE database, which shows a sensitivity of 77% and a specificity of 90% in the three feature vectors. The sensitivity is 76% and the specificity is 93%, while utilizing only the maximum eigen-value feature. Sharma and Wasson (Sharma 2015) propose the fuzzy method in the retinal vessel segmentation task uti-lizing a variety of high-pass and low-pass filter versions of retinal fundus images as the input. There is a wide range of fuzzy rules contained in the fuzzy logic and each fuzzy rule is built based on different thresholding values. The pixel values are selected and discarded using the thresholding values according to the fuzzy rules, leading to the extrac-tion of a vessel. An average accuracy of 95% is achieved by the methodology for the DRIVE database. A two-process combination between a fuzzy logic technique and a vessel tracking procedure is attempted by (Akhavan 2014). The first process involves the detection of the centerlines of improved retinal images, whereas, the second stage involves the use of Fuzzy C-Means (FCM) clustering to fill the retinal vessels. The centerline images are combined with fuzzy segmented images to obtain the final segmented result. Centerlines are utilized as first points for the FCM with region growing method. The assessment results of the techniques show an accuracy of 72.52% for the DRIVE database and 77.66% for the STARE database. A new methodology based on fuzzy c-means clustering and genetic algorithm is proposed in (Xie 2013). The pre-processing step involves the extraction of the green channel of raw retinal images, which is also improved using histogram equalization method. Subsequently, the reti-nal images are divided into two main levels, which are the texture and smooth level. The input of the processing stage is the texture layer because of the amount of information contained in it. The genetic algorithm is used together with the fuzzy c-means approach to cluster the features involv-ing data acquired in the first step. First, the near optimal solution of the global solution is obtained using the genetic algorithm and secondly, the estimated result is utilized as first values of the fuzzy c-means approach. The work of fuzzy c-means is eased and enhanced by genetic algorithm in terms of producing the best outcome without experienc-ing the challenge of local recognition of best results. Emary et al. (2014) attempt to address the issue associated with the objective function of a typical fuzzy c-means using the

possibility kind of fuzzy c-means technique enhanced by Cuckoo search approach. In their work, the novel clustering approaches, possibility c-means proposed by (Roy and Dasa 2015) and possibility fuzzy c-means proposed by (Krinidis 2010) are employed in establishing the best kind of fuzzy c-means that is consequently utilized for the segmentation of retinal vessels fundus images. The heuristic search algo-rithm of the Cuckoo Search is used in investigating the best of the proposed algorithm. The assessment for the STARE database demonstrates an accuracy of 0.94478, a specificity of 0.987 and a sensitivity of 0.586, whereas for the DRIVE database, an accuracy of 0.938, a specificity of 0.984, and a sensitivity of 0.628 are achieved.

5 Summary of findings of retinal vessel segmentation methods modern and approaches

The previous section provides the categorization and description of the extant retinal blood vessel segmentation methodologies. The results of the performances of the meth-ods are compared and the authors evaluate the performance of their methodologies using publicly available datasets. The different methodologies for the segmentation of reti-nal vessels follow similar procedures. The first step in each methodology is the pre-processing step, which involves the extraction of the green or grey layer (from the image of raw color retinal, followed by the enhancement of the image’s contrast. The next step is the processing step which is the nucleus of the algorithm, involving the use of the various techniques that are categorized in the previous section. In the last step of post-processing, the initial segmented image is submitted to be processed for smoothing and edge pre-serving and improvements. With regard to the categories of retinal segmentation provided in Fig. 5, there is no best algorithm, mathematical scheme, or technique that meets all the performance metrics for achieving high segmenta-tion. However, there are a number of factors that help in determining the most appropriate methodology, and some of them include (1) achieved accuracy that in turn, lays on the resulting sensitivity and specificity. With this metric, segmentation is regarded as the best if it is able to attain the biggest potential value of sensitivity, or shows small false detection for other retinal structures, while the specificity is maintained at optimal level. However, when the performance of the method is high in terms of the detection of pathologi-cal retinal images, then the method’s optimality increases; (2) time and computational complexity: with a higher level of accuracy, the computational time and power needed by the techniques tend to decrease; (3) robustness: a technique is considered to be superior than other methods if it dem-onstrates robustness with respect to parameters variation.



1 3

20 of 32

The backbone of different automated computer-aided systems for screening and diagnosis of cardiovascular and ophthalmologic diseases is the system’s ability to accu-rately detect and segment the retinal vascular structure. Despite the development and implementation of several promising methodologies, more research is needed so that the blood vessel methodologies can be further improved, especially for images with pathological retina and noisy behind the few number of retinal images for which datasets are available for public use. The role of expert diagnosis in practical applications cannot be replaced by retinal vessels segmentation systems, instead, they are meant to facili-tate accuracy in diagnosis, while reducing the workload of ophthalmologists. This way, experts are able to diagnose a large number of patients with high level of accuracy and comparable time.

Based on the algorithmic perspective, the reliability, per-formance and accuracy of the various methods of extract-ing vessels in the literature are determined by the special features possessed by retinal images. A rough outline of the features is as follows:

• There is a wide range of vessel widths, from less than a pixel up to 12 pixels wide.

• Vessels are often of low contrast, especially narrow ves-sels.

• Stronger responses are produced by different types of structures, like the optic disc, retina boundary, and pathologies at their boundaries.

• A bright strip, which is referred to as the central reflex runs down the center of some vessels, thereby leading to a complicated intensity cross-section, makes it harder to distinguish from two side-by-side vessels.

The results that are obtained from the application of the previously mentioned techniques are influenced by these features. In fact, of all the techniques presented in (Lesage 2009; Kirbas 2004; Fraz 2012), the most flexible tools for the extraction of vessel is demonstrated by active contours and neural networks. On one hand, the ability of neural networks to learn a suitable training set, including all possible objects or features makes it more appropriate for use on medical images. Nevertheless, each time a new feature is introduced to the network, a new training is required. Another limitation is that it is challenging to debug the network’s performance. Furthermore, active contours have been found to be a flex-ible and suitable tool for this task, because they are able to exploit the mixed-control; both bottom-up (image data) and top-down (prior to approximate knowledge about the loca-tion, shape and dimension of the structures). Object bounda-ries that are ambiguous or noisy can be managed by active contours. Such boundaries are commonly found in medical images such as MRI, ultrasound images, angiographies.

6 Retinal vessel classification techniques

The area of retinal vessel segmentation has undergone extensive research (Weiss et al. 2016), yet, less attention is given to the area that deals with the automatic classi-fication of segmented vessels. There are some challenges associated with the classification of vessels in images of the retinal fundus, and some of those challenges include the low contrast accompanying the fundus image, and the inhomogeneity of the background lighting. The inhomoge-neity occurs as a result of the process of imaging, whereas the low contrast is caused by the variation between the background and the contrast of the various blood vessels. This means that the contrast of thicker vessels is higher than those that are thinner. Another challenge is related to the color changes that occur in the retina for different sub-jects which are rooted in biological features. Most of the techniques used for the classification of the retinal vessel are based on geometric and visual characteristics that set the veins apart from the arteries. Basically, there are four areas in which the veins differ from the arteries: the shade of veins is darker than that of arteries, veins are thicker than arteries which are lighter, and the arteries are easily recognized by the central flex. In addition, there is often an alternation between veins and arteries close to the optic disc and before branching off. Nevertheless, in many situ-ations, these variations do not adequately distinguish the arteries from the veins. For instance, in situations when the quality of images is low, the central flex in the exter-nal areas is eliminated. A very dark shade is contained by the external regions of the image due to the effect of shading that arises from the inhomogeneity of the image’s lighting. Such situations create resemblance between the arteries and the veins, resulting in the misclassification of some vessels. In addition, since thickness varies from the highest value close to the optic disc to the smallest value in the external areas, it cannot be relied upon as a suit-able classification feature. The branching off of key veins and arteries into the optic disc, increases the possibility of having two arteries or veins that are adjacently positioned outside the optic disc.

There are five major steps involved in the approaches used in many studies that have focused on the classifi-cation of vessels in fundus images. These steps include the following: (1) segmentation of vessel, (2) selection of ROI for the classification of its vessels, (3) extraction of features from various areas of the vessel, (4) feature vec-tors’ classification, and (5) merging of results so that the vessel’s final label can be determined. In typical situations, the extraction of features involves three main procedures including, segment-based, profile-based, and pixel based. Profile is described as a piece of vessel possessing one



1 3

of 32 20

pixel thickness that is perpendicular to the orientation of the vessel. There are two major divisions of approaches that are proposed for the classification of vessels; semi-automatic and automatic. The semi-automatic approaches require the labeling of major vessels by an expert and the subsequent propagation of these labels is effected over the vessel network by means of vessel tracking approaches that employ the use of connectivity details, as well as the characteristics of the structure of a vascular tree. Unlike the semi-automatic, the automatic approaches initially involve the extraction of the centreline pixels of the ves-sels that constitute the vessel skeleton from the image of the retinal fundus. This is followed by the extraction of a wide range of features for each centreline pixel. In the last step of the automatic approach, the classifier labels each pixel as a vein or an artery.

6.1 Retinal vessel semi‑automatic techniques

The work in Martinez-Perez et al. (2002) proposes a semi-automatic method to enable the analysis of retinal vessel. Their approach involves the individual analysis of venous and arterial trees. After a branch has been identified by an expert as either an artery or a vein, the topological and geo-metric features are calculated for each of the branch’s seg-ment using an automatic procedure. An approach, in which the conditional optimization is employed, as proposed by (Rothaus et al. 2007, 2009), in their work. This approach, which is an extended version of the work of Martinez- Perez, is based on the anatomical properties of veins and arter-ies. In their method, the propagation of labels is carried out during the creation of vessel graph through the use of some starting segments that have been manually labelled. In the work by Estrada et al. (Estrada 2015), a semi-automatic-approach is presented, in which a combination of domain-specific knowledge and graph-theoretic methods are used. With the approach, which is dependent on the estimation of the vascular topology, the whole vasculature is analyzed. The method is regarded as an extended version of the tree topology estimation framework that is proposed by (Estrada et al. 2014), because it is dependent on the estimation of vascular tree topology, and combines expert domain-spe-cific features for the construction of likelihood model. In the next step of their proposed framework, the space of prob-able solutions that correspond with the projected vessels is searched repeatedly so that the model can be maximized. Four datasets [WIDE (Estrada et al. 2014), AV-DRIVE (Qureshi et al. 2013), CT-DRIVE (Qureshi et al. 2013) and AV-INSPIRE (Niemeijer et al. 2011)] are used to measure the performance of the presented method. The test results show that classification accuracy rates of 91.0, 93.5, 91.7, and 90.9% are achieved for WIDE, AV-DRIVE, CT-DRIVE and AV-INSPIRE, respectively. Reviews of extant literature

show that fully automated approaches that are employed in the clinical context are deployed in most of the studies conducted in the area of retinal vessel classification. The subsequent section presents an extensive review of such fully automated classifiers.

6.2 Retinal vessel automatic techniques

The complexity of artery and vein classification in retinal fundus images arises from the resemblance between the descriptive characteristics of these two structures, as well as the difference in contrast and lighting of fundus images. The two main problems associated with retinal images are the inhomogeneity in contrast and the lighting as a result of changes that occur in intra- and inter-images. However, these changes must be removed so that useful color information is obtained. It is for this reason that authors like (Grisan 2003) perform an analysis of the background image to enable the correction of these changes by statistically estimating their characteristics. The work by (Grisan 2003) in 2003 was one of the earliest works in the area of automatic methods, which were designed for the automatic classification of retinal ves-sels. One of the main steps involved in the analysis of fundus image is the extraction of vessel network, which requires the implementation of a vessel tracking process as well as a set of vessel segments. In their work, the vessel network is extracted automatically using the sparse tracking method. Vessel network symmetry and local features are exploited by dividing the retinal image into different portions having the same number of veins and arteries. The assumption is that the local features of the two types of vessels within these zones are quite different. This approach involves the division of an area around the optic disc (within 0.5–2 of the optic disc diameter from its centre) into four zones, with each zone containing one of the main arches. Despite the pres-ence of several features, the mean of the hue channel, and the difference of the red channels in each vessel segment are regarded as the most unique and suitable features that can be used for classification. Given two adjacent vessels, within the clinical context, the darker vessel is regarded as the vein, and in the absence of a significant difference in the red val-ues, the vessel which possesses more color homogeneity is regarded as a vein. Subsequent to the extraction of features, the fuzzy clustering algorithm is used in the classification of vessels. The criterion for classification is the Euclidean dis-tance between each pixel and the mean value of features in each class. Lastly, the labels of pixels found in each segment are merged based on major voting, and then the classifica-tion of the entire is carried out. Subsequent to the classifi-cation of vessel in the given zone, this classification can be extended from this zone by means of vessel tracking, which is made in a situation when the texture and color provides only little information for distinguishing between arteries



1 3

20 of 32

and veins. In this study, the analysis of 35 fundus images is performed and the algorithm is developed using 11 of the images and validated using the remaining 24. Their results show that for the 24 images, an overall error rate of 12.4% is achieved. This approach is considered as a more appropriate approach for optic disc-centered images, because the process of classification is implemented around the optic disc, and the underpinning assumption is the number of veins and arteries possessed by all four quadrants is the same.

In 2007, another method was developed by (Ruggeri et al. 2007) through AVR assessment in the area from 0.5 to 1 disc diameter from the optic disc margin. The authors provide a correlation with a manual reference standard on 14 images with variation ranging between 0.73 and 0.83, subject to the protocol used for the calculation of AVR. Subsequently, a modified version of this method is presented by (Tramon-tan et al. 2008), who improve the vessel tracking algorithm, resulting in an increase in the correlation with a reference standard of up to 0.88 for 20 images that are obtained from a DCCT study. In another study, the authors (Diabetes Control and Complications Trial Research Group 1987) employ the red contrast parameter, which is defined by the ratio of the peak of the central line intensity value to the largest intensity value of two edges of a vessel. The classification of veins and arteries in this study is made based on the average value of the red contrast along the vessel that is responsible for determining the probability of a vein. The classification of veins and arteries are made based on the average value of red contrasts along the vessel that is responsible for determin-ing the probability of falling under the category of veins. In a study conducted by (Li et al. 2003), the central flex in the green channel distinguishes the arteries from the veins. In their work, they identify the type of vessel through the application of the minimum Mahalanobis distance classifier. Using 505 vessel segments that are obtained from a wide range of fundus images, their experiments are performed and the results show that for arteries, a positive rate of 82.46% is achieved by the classifier, whereas the positive rate for veins is 89.03%.

To be able to separate veins from arteries, a wide range of features and classifiers are tested by (Jelinek et al. 2005). These tests are carried out using eight features, some of which are the standard deviation and the mean of red, blue, green and hue channels. The classifiers which used in their work are 13 classifiers obtained from Weka toolbox (Azuaje 2006; Efron 1983; Hall 1999). Their results show that the hue’s standard deviation, and the mean of the green are the best features, whereas for the best results of classification, a mean accuracy of 70% is achieved for eight images by the Naïve-Bayes classifier. (Narasimha-Iyer et al. 2007) carry out an investigation of four classifiers (SVM, 5-NN, NN, and Fisher linear discriminant) with the aim of classifying retinal vessels. Their results show that the Support Vector Machine

achieves the best results. The separation of the arteries from the veins is carried out using functional and structural fea-tures. Here, the central reflex is used as a structural indica-tor, as well as the ratio of the vessel optical densities from images at oxygen-sensitive and oxygen-insensitive as a func-tional feature. The evaluation of the classifier is performed by applying the classifier to a set of 251 segemented ves-sels that are acquired from 25 dual wavelength images. The results obtained show that the true positive which is achieved by their classifier is 97% for arteries, while that of veins is 90%. In the work by (Kondermann 2007) in 2007, ROI-based and two profile-base methods of feature extraction were investigated. In addition to that, the authors use two methods of classification based on NN and SVM for the sep-aration of veins and arteries in the fundus images. The RGB color space values are used as the profile-based features, whereby the mean values of the features are subtracted. The subtracted mean values are previously determined for each centreline pixel, as well as each pixel belonging to its profile. The ROI-based features are obtained from a square location found around each centreline pixel, and this centreline pixel is rotated in a manner that allows the alignment of the hori-zontal axis with the vessel’s main axis. In addition, through the use of multiclass principle component analysis, the size of the feature vector is reduced prior to the application of the classifier.

An ensemble classifier of boot strapped decision trees is proposed by (Fraz et al. 2014), so that vessels in fundus images can be classified as veins or arteries. The extraction of features from HIS and RGB color spaces is performed in three different ways, including, profile-based, pixel-based and segment-based. There are two ways through which the segment-based can be calculated: first, the calculation of the intensities’ variance and the mean is made for the whole segment, whereas the second method involves the division of segments that are comparatively large into smaller pieces having a length of about 50 pixels, and the calculations of mean and variance of the intensities are carried out in the pieces. Thus, a set feature is extracted, containing 51 color features, of which 16 are selected through the use of out-of-bag feature importance index. Their approach is used for the classification of all vessels in the whole image, and it is eval-uated using 3149 segmented vessels that are obtained from 40 macula-centered EPIC Norfolk fundus images (Forouhi 2012). Based on the reported results, their proposed method achieves a classification rate of 83%. A local binary pat-tern (LBP) is presented by (Hatami 1605), who use the pro-posed method to extract features of robustness in fundus images of low quality and low contrast. The LBP extracts features that are capable of providing sufficient details from the shape and texture. In their work, they investigate mul-tiple classifiers like bagging, AdaBoost, majority voting random subspace, random committee and rotation forest,



1 3

of 32 20

as well as, single classifiers like Cart, Bayesian Network, MLP, Random Tree, LibSVM and Naive Bayes. These clas-sifiers are investigated for the purpose of solving the clas-sification problem associated with retinal vessels. The meth-ods studied by the authors, are evaluated using 20 images that are acquired from the STARE dataset (Guo 2017). For their experiments, nine methods of feature extraction are employed, while 13 methods of classifications are used. The experimental results obtained by these authors show that there is significant improvement in performance of the clas-sifiers when the LBP features are used. In comparison with single classifiers, the performance of the multiple classi-fiers is better with the use of LBP features. These groups of researchers use the MS-R-LBP together with random com-mittee classifier, which in turn produces the best results with a classification rate of 90.7% and an AUC value of 0.97.

7 Summary findings on the automatic retinal vessel classification techniques

Due to the existing challenges associated with retinal ves-sels’ classification, it is difficult to achieve high accuracy of classification. In existing methods, arteries are distinguished from veins by means of four characteristics which include, color variation, central reflex, variation in thickness, and branching information of veins and arteries surrounding the optic disc. In a wide range of classification methods, the use of features, which provide the descriptions of color as well as color differences of the vessels, is employed. One of the main problems associated with classification is the varia-tion in the absolute color of blood vessels between images, and even in the same image. This difference is attributed to some factors like flash spectrum, saturation of hemoglobin oxygen, aging and cataract development, flash artifacts, cam-era nonlinear optical distortions, and focus (Niemeijer et al. 2011). In addition to the aforementioned factors, another factor that should be considered is the images’ resolution. Noise is introduced to images of high resolution; it reduces the color information. In studies in which a fundus camera has a variety of resolutions, it is important to normalize the resolution of images, because several methods of classifica-tion depend on color information. However, a vessel’s thick-ness cannot be relied upon for feature classification because it varies along the vessel, and it is significantly influenced by the vessel’s segmentation. Due to the high volumes of oxy-gen carried by the arteries, the internal parts of the arteries are brighter than the outer parts, meaning that, in arteries, the central reflex is more noticeable. This feature is high-lighted by some researchers as a very critical feature that is used to distinguish between the vessels. However, it is only in the thicker vessels that the central flex is noticeable.

With all these factors, the computational task of auto-matically classifying veins and arteries becomes very chal-lenging, which affects the system’s accuracy. In attempts to address this problem, some researchers employ the use of clustering rather than classification. Other researchers attempt to make the problems simpler by selecting just the main vessels surrounding the optic disc during the process of classification. However, this limits the analysis to just the main vessels, while avoiding the ambiguous information that emerges from smaller veins and arteries. For applica-tions like AVR calculation, the classification of the main vessels is adequate. Apart from AVR, there is a wide range of parameters that are used to evaluate the different available approaches. Classification error is one of the criteria that should be seriously considered due to its great relevance, because a method’s reliability is higher when the number of errors is less. The error rates are used to determine if manual labeling can be replaced with automatic methods. A quantitative comparison of the different proposed methods of the classification of vessel in fundus images cannot be justified due to the different datasets with varying resolu-tion, intensities and imaging conditions that are used for the evaluation of these methods based on a wide range of parameters including AUC, accuracy of classification, AVR and correlation coefficient. Based on the stated limitations, efforts are made in this study to assess the proposed methods based on a different point of view, which is the image model-ling approach. In a study by (Zahra Amini 2016), the differ-ent models that are employed in the processing of medical image are categorized in a detailed manner. Based on the literature review, there are two main categories of model-ling methods, which are transform and spatial domains. However, there are smaller categories in all of the models, including, stochastic, deterministic, partial differential equa-tion and geometric. In this study also, automatic methods of retinal vessels’ classification are categorized based on the related modelling category. These methods of automatic classification of retinal vessels are summarized in Table 2, which is based on, features, datasets, classifiers, methods of evaluation and results, and most especially, the model-ling approaches. The best results obtained by each of the approaches are presented.

In this paper, the proposed methods of classifying arteries and veins in fundus images are extensively reviewed. These approaches are categorized into automatic and semi-auto-matic categories. The semi-automatic approaches involve experts initiating the classification to distinguish between veins and arteries, after which, an automatic propagation occurs throughout the vascular network of the fundus image (Sarah Husham et al. 2020; O.I.O. et al. 2018; Elhoseny et al. 2021). For this reason, vessel-tracking algorithms are employed and in these algorithms, connectivity details and structural characteristics are critical. On the other hand,



1 3

20 of 32

Tabl

e 2

The

sum

mar

y fin

ding

s on

the

met

hods

of a

utom

atic

cla

ssifi

catio

n of

retin

al v

esse

ls w

hich

is b

ased

on,

feat

ures

, dat

aset

s, cl

assi

fiers

, met

hods

of e

valu

atio

n an

d re

sults

, and

mos

t esp

e-ci

ally

, the

mod

ellin

g ap

proa

ches

Aut

hor(

s)/y

ear

Feat

ures

ext

ract

ion

tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n ca

tego

ry

Gris

an (2

003)

Mea

n H

—Va

rianc

e R

Acc

35 fu

ndus

imag

es h

ave

been

col

lect

edVe

ssel

trac

king

and

Fuz

zy c

luste

ring

Li e

t al.

(200

3)Th

e C

entra

l refl

ex in

GTP

505

fund

us im

ages

hav

e be

en c

olle

cted

Min

imum

Mah

alan

obis

dist

ance

Jelin

ek e

t al.

(200

5)std

of R

GB

, H a

nd M

ean

Acc

20 fu

ndus

imag

es h

ave

been

col

lect

edN

aive

-Bay

esRu

gger

i et a

l. (2

007)

Mea

n H

—Va

rianc

e R

Cor

rela

tion

and

AVR

14 fu

ndus

imag

es h

ave

been

col

lect

edVe

ssel

trac

king

and

Fuz

zy c

luste

ring

Nar

asim

ha-I

yer e

t al.

(200

7)Th

e ve

ssel

opt

ical

den

sitie

s rat

ioTP

25 fu

ndus

imag

es h

ave

been

col

lect

edSu

ppor

t vec

tor m

achi

neK

onde

rman

n (2

007)

RGB

Acc

4 fu

ndus

imag

es h

ave

been

col

lect

edN

N m

ultil

ayer

per

cept

ron

Tram

onta

n et

al.

(200

8)R

con

trast

Cor

rela

tion

and

AVR

20 fu

ndus

imag

es h

ave

been

col

lect

edVe

ssel

trac

king

-bas

ed T

hres

hold

ing

Nie

mei

jer e

t al.

(201

1)St

eere

d G

auss

ian

deriv

ativ

es,

HSI

+ w

idth

, and

con

trast,

AU

C c

urve

DR

IVE

k-N

eare

st ne

ighb

ors

Kar

ssem

eije

r et a

l. (2

010)

RGB

Acc

DR

IVE

Line

ar d

iscr

imin

ant a

naly

sis

Mur

amat

su e

t al.

(201

0)G

con

trast,

RG

B +

R,

Acc

DR

IVE

Line

ar d

iscr

imin

ant a

naly

sis

Nie

mei

jer e

t al.

(201

1)RG

B, a

nd H

SIM

ean

AVR

and

AU

C

65 fu

ndus

imag

es h

as b

een

colle

cted

fro

m Io

wa

Uni

vers

ityLi

near

dis

crim

inan

t ana

lysi

s

Roth

aus (

2011

)Fi

ve m

odel

Fea

ture

s and

HIS

Cla

ssifi

catio

n er

ror

MA

RS

data

set

k-M

eans

clu

sterin

gZa

mpe

rini e

t al.

(201

2)Po

sitio

nal f

eatu

res a

nd 1

6 co

lors

Acc

42 fu

ndus

imag

es h

as b

een

colle

cted

fro

m N

inew

ells

Hos

pita

lLi

near

nor

mal

Bay

es

Vázq

uez

(201

2)M

edia

n G

Acc

ICAV

R-2

data

set

Vess

el tr

acki

ng b

ased

k-m

eans

clu

sterin

gM

irsha

rif e

t al.

(201

3)8

(R, G

, LA

B) c

olor

feat

ures

Acc

13 fu

ndus

imag

es h

as b

een

colle

cted

fro

m E

ye H

ospi

tal +

DR

IVE

Line

ar d

iscr

imin

ant a

naly

sis

Rela

n et

al.

(201

3)Va

rianc

e of

R a

nd M

ean

of (R

, G, H

)A

cc35

fund

us im

ages

has

bee

n co

llect

edG

auss

ian

mix

ture

mod

el e

xpec

tatio

n m

axim

izat

ion

Rela

n et

al.

(201

4)Va

rianc

e of

R a

nd M

ean

of (R

, G, H

)A

cc35

fund

us im

ages

has

bee

n co

llect

ed

from

ORC

AD

ESLe

ast s

quar

e su

ppor

t vec

tor m

achi

ne

Josh

i et a

l. (2

014)

G, H

var

ianc

e an

d M

ean

Acc

EYEC

HEC

K d

atas

etFu

zzy

C-m

eans

clu

sterin

gD

asht

bozo

rg e

t al.

(201

3)19

col

or fe

atur

esAV

R, A

ccV

ICAV

RD

RIV

EIN

SPIR

ELi

near

dis

crim

inan

t ana

lysi

s and

gra

ph

labe

ling

resu

ltsFr

az e

t al.

(201

4)H

SI c

olor

+ fe

atur

es R

GB

Acc

40 fu

ndus

imag

es h

as b

een

colle

cted

Dec

isio

n tre

esH

atam

i (16

05)

9 fe

atur

e (L

BP,

PCA

, RG

B, e

tc.)

AU

C a

nd A

ccST

AR

EM

ultip

le c

lass

ifier

syste

ms

Rela

n (2

016)

Four

col

or fe

atur

es a

re m

ean

of h

ue

(MH

),mea

n of

red

(MR

), va

rianc

e of

re

d (V

R),

and

mea

n of

gre

en (M

G)

Acc

DR

IVE,

INSP

IRE-

AVR

Squa

red-

loss

mut

ual i

nfor

mat

ion

clus

ter-

ing

(SM

IC)

Das

gupt

a (2

017)

Gab

or, a

t Rid

ge fe

atur

es v

ario

us sc

ales

an

d de

gree

sA

cc a

nd A

UC

D

RIV

EC

NN

Kho

waj

a et

al.

(201

8)Si

x di

sting

uish

ing

feat

ure

extra

ctio

n te

chni

ques

: LB

P, lo

cal i

nten

sitie

s, H

OG

, hig

h-or

der l

ocal

aut

ocor

rela

-tio

ns, d

iver

genc

e of

vec

tor fi

eld,

and

m

orph

olog

ical

tran

sfor

mat

ion

Acc

CH

ASE

DB

1, S

TAR

E, a

nd D

RIV

EH

iera

rchi

cal c

lass

ifica

tion



1 3

of 32 20

the automatic approaches involve the automatic classifica-tion of vessels, without human intervention. The automatic approaches are mostly preferred because they are more suit-able for clinical use. An unbiased comparison of the per-formance of the different automatic approaches is impos-sible, due to a wide range of datasets and parameters that are used to evaluate the performance of these approaches (Subathra et al. 2020; Al-Dhief 2020). Thus, in this study, the methods are evaluated from a modelling point of view based on the categorization which (Zahra Amini 2016) pre-sented. It is observed from Table 2 that the geometric models are employed in a majority of the extant approaches for the step involving feature extraction, whereas, the determin-istic or stochastic strategies are used in the step involving classification.

8 Conclusion

In this study, an overview of the problem domain is deliber-ated in three parts. In the first part, the anatomy and physiol-ogy of the eye are discussed in relation to the Retinal Quan-tification Measures along with lesions and associated risks. Furthermore, a review of image modalities for a retina is made. In addition, a discussion on the Medical Images of Retinal Fundus is presented in this study. The second part of the study provides a detailed discussion on the techniques used for the segmentation of retinal vessels based on the current literature. In addition, the different techniques used in the methodologies proposed by different authors are dis-cussed, and a comparison of the results of the methods is made. Publicly available datasets are employed in the evalu-ation of these methodologies, which follow a similar proce-dure: the first step is the pre-processing step, which involves the extraction of the green or grey layer from an image of the raw color retina followed by the enhancement of the image’s contrast. The next step is the processing step, which is the nucleus of algorithms. Finally, in the post-processing step, the initial segmented image is deployed for edge preserving, smoothing and improvements. Based on the reviews made in this study, there are two main categories of modelling methods, which are transform and spatial domains. However, there are smaller categories in all of the models, including, stochastic, deterministic, partial differential equation and geometric. In the third part of this study, automatic meth-ods of retinal vessels classification are categorized based on the related modelling category. These methods of automatic classification of retinal vessels are summarized in Table 2, which is based on features, datasets, classifiers, methods of evaluation and results, and modelling approaches. The best results obtained by each of the approaches are presented.

Tabl

e 2

(con

tinue

d)

Aut

hor(

s)/y

ear

Feat

ures

ext

ract

ion

tech

niqu

esPe

rform

ance

eva

luat

ion

met

rics

Valid

atio

n da

tase

t use

dM

etho

d/te

chni

que

clas

sific

atio

n ca

tego

ry

Jeba

seel

i et a

l. (2

019)

Uni

form

dist

ribut

ion

of g

ray

valu

esA

cc, S

ens,

and

Spec

DR

IVE

Kirs

ch’s

tem

plat

e an

d Fu

zzy

C-M

eans

Hes

linga

et a

l. (2

020)

12 fe

atur

es p

er in

divi

dual

Acc

, and

AU

C

5222

fund

us im

ages

has

bee

n co

llect

edD

eep

neur

al n

etw

orks

Usm

an (2

020)

Tran

sfer

Lea

rnin

gA

cc26

80 fu

ndus

imag

es h

as b

een

colle

cted

Dee

p N

eura

l Net

wor

kW

.W. e

t al.

(202

0)En

hanc

emen

t bas

ed o

n th

e RG

B (r

ed,

gree

n, b

lue)

col

or sp

ace

Acc

HD

R im

ages

Low

-ligh

t im

age

enha

ncem

ent

Cao

et a

l. (2

020)

Rem

ovin

g su

ch lo

w fr

eque

ncy

in th

e ro

ot

dom

ain

Acc

200

retin

al fu

ndus

imag

esG

amm

a m

ap a

nd C

LAH

E

Nav

eed

et a

l. (2

021)

Intro

duci

ng a

noi

se re

mov

al m

etho

dol-

ogy

Acc

, Sen

s, an

d Sp

ecD

RIV

E, S

TAR

E A

ND

CH

ASE

Ense

mbl

e B

lock

Mat

chin

g 3D

Filt

er

(S-B

M3D

)Je

na e

t al.

(202

1)Re

duce

s the

gap

bet

wee

n su

perv

ised

and

un

supe

rvis

ed m

etho

dsA

UC

, Acc

DR

IVE,

STA

RE

Flow

-Bas

ed C

onsi

stenc

ies

Jiang

(202

0)Ex

tract

the

info

rmat

ion

of d

iffer

ent t

hick

-ne

sses

of b

lood

ves

sels

accu

racy

, sen

sitiv

ity,

spec

ifici

ty,F

-mea

sure

, and

A

UC

DR

IVE,

CH

ASE

, and

STA

RE

Mul

ti-Sc

ale

Resi

dual

Atte

ntio

n N

etw

ork

Lian

toni

(202

1)Th

e ap

plic

atio

n of

the

Ada

ptiv

e A

CO

met

hod

has s

ucce

ssfu

lly o

ptim

ized

the

resu

lt of

retin

al v

esse

l edg

e de

tect

ion

Acc

RGB

imag

eG

radi

ent b

ased

ant

spre

ad m

odifi

catio

n on

an

t col

ony

optim

izat

ion

met

hod,

Pea

k Si

gnal

-to-N

oise

Rat

io (P

SNR

)Re

lan

and

Rela

n (2

020)

Gre

at c

apab

ility

to e

nhan

ce c

ompu

ter

assi

sted

diag

nosi

s aA

ccIN

SPIR

E-AV

R, V

ICAV

R, a

nd M

ESSI

-D

OR

Con

siste

nt G

auss

ian

mix

ture

s



1 3

20 of 32

Funding This study was not funded.

Declarations

Conflict of interest There is no conflict of interests.

Ethical approval This article does not contain any studies with human participants or animals performed by any of the authors.

References

Abd Ghani MK et al (2018) Decision-level fusion scheme for naso-pharyngeal carcinoma identification using machine learning techniques. Neural Comput Appl 32(3):625–638

Abdallah MB, Malek J, Krissian K and Tourki R (2011) An auto-mated vessel segmentation of retinal images using multiscale vesselness. . In eighth international multi-conference on systems, signals and devices IEEE (1–6)

Abdulhay E et al (2018) Computer aided solution for automatic seg-menting and measurements of blood leucocytes using static microscope images. J Med Syst 42(4):58

Adel M, Rasigni M, Gaidon T, Fossati C and Bourennane S (2009) Statistical-based linear vessel structure detection in medical images. In 2009 16th IEEE international conference on image processing (ICIP) (649–652)

Akhavan RAFK (2014) A novel retinal blood vessel segmentation algo-rithm using fuzzy segmentation. Int J Elect Comp Eng 4(4):561

Akram MU, Tariq A and Khan SA (2009) Retinal image blood vessel segmentation. In 2009 international conference on information and communication technologies IEEE (181–192)

Al-Dhief FT, Latiff NMA, Malik NNNA, Sabri N, Baki MM, Albadr MAA, Abbas AF, Hussein YM, Mohammed MA (2020) Voice pathology detection using machine learning technique. In: 2020 IEEE 5th international symposium on telecommunication tech-nologies (ISTT). IEEE, pp 99–104

Al-Fahdawi S et al (2016) A fully automatic nerve segmentation and morphometric parameter quantification system for early diagno-sis of diabetic neuropathy in corneal images. Comput Methods Programs Biomed 135:151–166

Al-Fahdawi S et al (2018) A fully automated cell segmentation and morphometric parameter system for quantifying corneal endothe-lial cell morphology. Comput Methods Programs Biomed 160:11–23

Al-Rawi M, Qutaishat M, Arrar M (2007) An improved matched filter for blood vessel detection of digital retinal images. Comput Biol Med 37(2):262–267

Aparna CLSP, Rajan J (2017) Recent advancements in retinal vessel segmentation. J Med Syst 41(4):70

Arunkumar N et al (2018) K-Means clustering and neural network for object detecting and identifying abnormality of brain tumor. Soft Comput 23(19):9083–9096

Arunkumar N et al (2018) Fully automatic model-based segmentation and classification approach for MRI brain tumor using artificial neural networks. Concurr Comput Pract Exp 32:1

Azuaje F, Witten IH, Frank E (2006) Data mining: practical machine learning tools and techniques 2nd edition. BioMed Eng OnLine 5:51. https:// doi. org/ 10. 1186/ 1475- 925X-5- 51

Babaud J, Witkin AP, Baudin M, Duda RO (1986) Uniqueness of the Gaussian kernel for scale-space filtering. IEEE Trans Pattern Anal Mach Intell 1:26–33

Bankhead P et al (2012) Fast retinal vessel detection and measure-ment using wavelets and edge location refinement. PLoS ONE 7(3):e32435

Binh NT, Tuyet VTH, Hien NM, Thuy NT (2019) Retinal vessels segmentation by improving salient region combined with Sobel operator condition. In: International conference on future data and security engineering. Springer, Cham, pp 608–617

Brancati N, Frucci M, Gragnaniello D, Riccio D (2018) Retinal ves-sels segmentation based on a convolutional neural network. In: Iberoamerican congress on pattern recognition. Springer, Cham, pp 119–126

Budai A, Michelson G and Hornegger J (2010) Multiscale blood ves-sel segmentation in retinal fundus images. In Bildverarbeitung für die Medizin 2010 - Algorithmen, Systeme, Anwendungen, pp 261–265

Cao L, Li H, Zhang Y (2020) Retinal image enhancement using low-pass filtering and α-rooting. Signal Processing 170:107445

Chanwimaluang TAF (2003) G, <an-efficient-algorithm-for-extrac-tion-of-anatomical-structures-i.pdf>. IEEE. (In proceedings 2003 international conference on image processing): (Cat. No. 03CH37429. (1, I-1093)

Chapman N, Dell’Omo G, Sartini MS, Witt N, Hughes A, Thom S, Pedrinelli R (2002) Peripheral vascular disease is associated with abnormal arteriolar diameter relationships at bifurcations in the human retina. Clin Sci 103(2):111–116

Chaudhuri S, Chatterjee S, Katz N, Nelson M, Goldbaum M (1989) Detection of blood vessels in retinal images using two-dimen-sional matched filters. IEEE Trans Med Imaging 8(3):263–269

Christodoulidis A et al (2016) A multi-scale tensor voting approach for small retinal vessel segmentation in high resolution fundus images. Comput Med Imaging Graph 52:28–43

Chutatape O, Zheng L and Krishnan SM (1998) Retinal blood vessel detection and tracking by matched Gaussian and Kalman filters. . In proceedings of the 20th annual international conference of the IEEE engineering in medicine and biology society, biomedi-cal engineering towards the year 2000 and Beyond (Cat. No. 98CH36286) (6 3144–3149)

Couper DJ, Klein R, Hubbard LD, Wong TY, Sorlie PD, Cooper LS, Brothers RJ, Nieto FJ (2002) Reliability of retinal photography in the assessment of retinal microvascular characteristics: the atherosclerosis risk in communities study. Am J Ophthalmol 133(1):78–88

da Rocha DA et al (2020) An unsupervised approach to improve con-trast and segmentation of blood vessels in retinal images using CLAHE, 2D Gabor wavelet, and morphological operations. Res Biomed Eng 36(1):67–75

Dan Y et al (2021) Retinal blood vessel segmentation method based on multi scale convolution kernel U net model. J Northeast Univ (Nat Sci) 42:1

Dasgupta AASS (2017) A fully convolutional neural network based structured prediction approach towards the retinal vessel segmen-tation. In 2017 IEEE 14th international symposium on biomedi-cal imaging (ISBI 2017), (248–251)

Dashtbozorg B, Mendonça AM, Campilho A (2013) An automatic graph-based approach for artery/vein classification in retinal images. IEEE Trans Image Process 23(3):1073–1083

De J, Cheng L, Zhang X, Lin F, Li H, Ong KH, Yu W, Yu Y, Ahmed S (2015) A graph-theoretical approach for tracing filamentary structures in neuronal and retinal images. IEEE Trans Med Imag-ing 35(1):257–272

Decencière E et al (2014) Feedback on a publicly distributed image database: the messidor database. Image Anal Stereol 33(3):231

De Silva A, Perera MV, Wijethilake N, Jayasinghe S, Nanayakkara ND, De Silva A (2021) A thickness sensitive vessel extraction frame-work for retinal and conjunctival vascular tortuosity analysis. arXiv preprint arXiv: 2101. 00435



1 3

of 32 20

Dharmawan DA et al (2019) A new hybrid algorithm for retinal vessels segmentation on fundus images. IEEE Access 7:41885–41896

Diabetes Control and Complications Trial Research Group (1987) Color photography vs. fluorescein angiography in the detection of diabetic retinopathy in the Diabetes Control and Complica-tions Trial. Arch Ophthalmol 105:1344–1351

Dizdaro B, Ataer-Cansizoglu E, Kalpathy-Cramer J, Keck K, Chiang MF and Erdogmus D (2012) Level sets for retinal vasculature segmentation using seeds from ridges and edges from phase maps. In 2012 IEEE international workshop on machine learn-ing for signal processing. IEEE (1–6)

Efron B (1983) Estimating the error rate of a prediction rule: improve-ment on cross-validation. J Am Stat Assoc 78(382):316–331

Elhoseny M, Mohammed MA, Mostafa SA, Abdulkareem KH, Maashi MS et al (2021) A new multi-agent feature wrapper machine learning approach for heart disease diagnosis. Comput Mater Contin 67(1):51–71

Emary E, Zawbaa HM, Hassanien AE, Schaefer G and Azar AT (2014) Retinal vessel segmentation based on possibilistic fuzzy c-means clustering optimised with cuckoo search. In 2014 international joint conference on neural networks (IJCNN) IEEE (1792–1796)

Estrada R, Tomasi C, Schmidler SC, Farsiu S (2014) Tree topology esti-mation. IEEE Trans Pattern Anal Mach Intell 37(8):1688–1701

Estrada R et al (2015) Retinal Artery-Vein Classification via Topology Estimation. IEEE Trans Med Imaging 34(12):2518–2534

Foracchia M, Grisan E, Ruggeri A (2005) Luminosity and contrast normalization in retinal images. Med Image Anal 9(3):179–190

Forouhi NG et al (2012) Circulating 25-hydroxyvitamin D concentra-tion and the risk of type 2 diabetes: results from the European Prospective Investigation into Cancer (EPIC)-Norfolk cohort and updated meta-analysis of prospective studies. Diabetologia 55(8):2173–2182

Francia GA, Pedraza C, Aceves M, Tovar-Arriaga S (2020) Chaining a U-net with a residual U-net for retinal blood vessels segmenta-tion. IEEE Access 8:38493–38500

Frangi AF, Niessen WJ, Vincken KL, Viergever MA (1998) Multi-scale vessel enhancement filtering. In International conference on medical image computing and computer-assisted intervention. Springer 1998:130–137

Fraz MM et al (2012) Blood vessel segmentation methodologies in retinal images–a survey. Comput Methods Programs Biomed 108(1):407–433

Fraz MM, Remagnino P, Hoppe A, Uyyanonvara B, Rudnicka AR, Owen CG, Barman SA (2012) An ensemble classification-based approach applied to retinal blood vessel segmentation. IEEE Trans Biomed Eng 59(9):2538–2548

Fraz MM, Rudnicka AR, Owen CG, Strachan DP and Barman SA (2014) Automated arteriole and venule recognition in retinal images using ensemble classification. International conference on computer vision theory and applications (VISAPP). IEEE (3: 194–202): 194–202

Frucci M, Riccio D, Di Baja GS, Serino L (2014) Using contrast and directional information for retinal vessels segmentation. In: 2014 tenth international conference on signal-image technology and internet-based systems. IEEE, pp 592–597

Fu H, Xu,Y, Lin S, Wong DWK, Liu J (2016) Retinal vessel segmenta-tion via deep learning network and fully-connected conditional random fields. In: International conference on medical image computing and computer-assisted intervention. Springer, Cham, pp 132–139

Gao X, Cai Y, Qiu C and Cui Y (2017) Retinal blood vessel segmenta-tion based on the Gaussian matched filter and U-net. In 2017 10th international congress on image and signal processing, Biomedi-cal engineering and informatics (CISP-BMEI) IEEE (1–5)

Gongt H, Li Y, Liu G, Wu W and Chen G (2015) A level set method for retina image vessel segmentation based on the local cluster

value via bias correction. In 2015 8th international congress on image and signal processing (CISP) IEEE (413–417)

Grisan EARA (2003) A divide et impera strategy for automatic classifi-cation of retinal vessels into arteries and veins. In Proceedings of the 25th annual international conference of the IEEE engineering in medicine and biology society (IEEE Cat. No. 03CH37439). (1: 890–893). IEEE

Gu L (2015) and L. Learning to boost filamentary structure segmenta-tion, Cheng, pp 639–647

Guo Y et al (2017) A retinal vessel detection approach based on shear-let transform and Indeterminacy Filtering on Fundus Images. Symmetry 9(10):235

Guo Y et al (2018) A retinal vessel detection approach using convolu-tion neural network with reinforcement sample learning strategy. Measurement 125:586–591

Hajabdollahi M, Esfandiarpoor R, Najarian K, Karimi N, Samavi S and Reza-Soroushmeh SM (2018) Low complexity convolutional neural network for vessel segmentation in portable retinal diag-nostic devices. In 2018 25th IEEE international conference on image processing (ICIP). (2785–2789)

Hall MA (1999) Correlation-based feature selection for machine learn-ing, Doctor of Philosophy at The University of Waikato, Ham-ilton, New Zealand

Hamad H et al (2020) Exudates as Landmarks Identified through FCM Clustering in Retinal Images. Appl Sci 11(1):142

Hatami NAGM (2016) Automatic identification of retinal arteries and veins in fundus images using local binary patterns. arXiv preprint arXiv: 1605.00763

Hatamizadeh A, Hosseini H, Liu Z, Schwartz SD and Terzopoulos D (2019) Deep dilated convolutional nets for the automatic seg-mentation of retinal vessels. . arXiv preprint arXiv:, 1905.12120

Heslinga Pluim FG, Houben JPAJHM, Schram MT, Henry RM, Ste-houwer CD, Van Greevenbroek MJ, Veta BTTM (2020) Direct classification of type 2 diabetes from retinal fundus images in a population-based sample from The Maastricht Study In Medical Imaging 2020: computer-Aided Diagnosis. Int Soc Opt Photon 11314:113141N

Hoover AD, Kouznetsova V, Goldbaum M (2000) Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Trans Med Imaging 19(3):203–210

Hu K et al (2018) Retinal vessel segmentation of color fundus images using multiscale convolutional neural network with an improved cross-entropy loss function. Neurocomputing 309:179–191

Huang J-H, Huck Yang C-H, Liu F, Tian M, Liu Y-C, Wu T-W, Lin I (2021) DeepOpht: medical report generation for retinal images via deep models and visual explanation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp 2442–2452

Isavand Rahmani A, Akbari H, Esmaili S (2020) Retinal blood vessel segmentation using gabor filter and morphological reconstruc-tion. Sign Process Renew Energy 4(1):77–88

Jebaseeli TJ, Durai CAD, Peter JD (2019) Extraction of retinal blood vessels on fundus images by kirsch’s template and Fuzzy C-Means. J Med Phys 44(1):21

Jelinek HF, Depardieu C, Lucas C, Cornforth DJ, Huang W, Cree MJ (2005) Towards vessel characterization in the vicinity of the optic disc in digital retinal images. Image Vis Comput Conf 2:7

Jena R, Singla S, Batmanghelich K (2021) Self-supervised vessel enhancement using flow-based consistencies. arXiv preprint arXiv: 2101. 05145.

Jiang XAMD (2003) Adaptive local thresholding by verification-based multithreshold probing with application to vessel detec-tion in retinal images. IEEE Transactions Patt Anal Mach Intell 25(1):131–137

Jiang Z et al (2017) Fast, accurate and robust retinal vessel segmenta-tion system. Biocybern Biomed Eng 37(3):412–421



1 3

20 of 32

Jiang Y, Tan N, Peng T, Zhang H (2019) Retinal vessels segmentation based on dilated multi-scale convolutional neural network. IEEE Access 7:76342–76352

Jiang Y et al (2020) A multi-scale residual attention network for retinal vessel segmentation. Symmetry 13(1):24

Jin Q et al (2019) DUNet: a deformable network for retinal vessel segmentation. Knowl-Based Syst 178:149–162

Joshi VS, Reinhardt JM, Garvin MK, Abramoff MD (2014) Automated method for identification and artery-venous classification of ves-sel trees in retinal vessel networks. PLoS ONE 9:2

Kamran SA, Hossain KF, Tavakkoli A (2021) RV-GAN: retinal vessel segmentation from fundus images using multi-scale generative adversarial networks. arXiv: 2101.00535v1

Karssemeijer N et al (2010) Automated detection and classification of major retinal vessels for determination of diameter ratio of arteries and veins 7624:76240J

Kauppi T, Kalesnykiene V, Kamarainen J-K, Lensu L, Sorri I, Raninen A, Voutilainen R, Uusitalo H, Kälviäinen H, Pietilä J (2007) The diaretdb1 diabetic retinopathy database and evaluation protocol. In: BMVC, vol 1, pp 1–10

Khaing TT et al (2021) Glaucoma detection in mobile phone retinalim-ages based on adi-gvf segmentation withem initialization. ECTI Transactions on Computer Inform Technol 15:1

Khowaja SA, Khuwaja P, Ismaili IA (2018) A framework for retinal vessel segmentation from fundus images using hybrid feature set and hierarchical classification. SIViP 13(2):379–387

Kingsbury N (1998) The dual-tree complex wavelet transform: a new efficient tool for image restoration and enhancement. In 9th Euro-pean signal processing conference (EUSIPCO 1998) (1–4). IEEE

Kirbas CAQF (2004) A review of vessel extraction techniques and algorithms. ACM Comput Surv (CSUR) 36(2):81–121

Kondermann C, Kondermann D, Yan M (2007) Blood vessel clas-sification into arteries and veins in retinal images. In: Medical imaging 2007: image processing, vol 6512. International Society for Optics and Photonics, p 651247

Krinidis SACV (2010) A robust fuzzy local information C-means clustering algorithm. IEEE Transactions Image Process 19(5):1328–1337

Kumar D, Pramanik A, Kar SS and Maity SP (2016) Retinal blood vessel segmentation using matched filter and laplacian of gauss-ian. In 2016 international conference on signal processing and communications (SPCOM) IEEE June. (1–5)

Kundu AACRK (2012) Retinal vessel segmentation using morpho-logical angular scale-space. In 2012 third international confer-ence on emerging applications of information technology. IEEE (316–319)

Lahiri A, Roy AG, Sheet D and Biswas PK (2016) Deep neural ensem-ble for retinal vessel segmentation in fundus images towards achieving label-free angiography. In 2016 38th annual interna-tional conference of the IEEE engineering in medicine and biol-ogy society (EMBC) (1340–1343)

Lenskiy AA, Lee J (2010) Rugged terrain segmentation based on salient features. ICCAS 2010, Gyeonggi-do, Korea (South), pp 1737–1740. https:// doi. org/ 10. 1109/ ICCAS. 2010. 56697 87

Lesage D et al (2009) A review of 3D vessel lumen segmentation tech-niques: models, features and extraction schemes. Med Image Anal 13(6):819–845

Li H, Hsu W, Lee ML and Wang H (2003) A piecewise Gaussian model for profiling and differentiating retinal vessels. In Proceedings 2003 international conference on image processing (Cat. No. 03CH37429). IEEE (1: I-1069)

Li H, Hsu W, Lee ML, Wong TY (2005) Automatic grading of retinal vessel caliber. IEEE Trans Biomed Eng 52(7):1352–1355

Li H, Zhang J, Nie Q and Cheng L (2013) A retinal vessel tracking method based on bayesian theory. In 2013 IEEE 8th conference on industrial electronics and applications (ICIEA) (232–235)

Liantoni F et al (2021) Gradient based ant spread modification on ant colony optimization method for retinal blood vessel edge detec-tion. IOP Conf Ser Mater Sci Eng 1010:012021

Lindeberg T (2011) Scale-space theory: a basic tool for analyzing structures at different scales. J Appl Stat 21(1–2):225–270

Liskowski PAKK (2016) Segmenting retinal blood vessels with deep neural networks. IEEE Transactions Med Imaging 35(11):2369–2380 (9901)

Lv Y, Ma H, Li J, Liu S (2020) Attention guided U-net with atrous con-volution for accurate retinal vessels segmentation. IEEE Access 8:32826–32839

Ma ZALH (2015) Retinal vessel profiling based on four piecewise Gaussian model. IEEE international conference on digital signal processing DSP 1094–1097

Maji D, Santara A, Ghosh S, Sheet D and Mitra P (2015) Deep neu-ral network and random forest hybrid architecture for learning to detect retinal vessels in fundus images. In 2015 37th annual international conference of the IEEE engineering in medicine and biology society (EMBC) (3029–3032)

Maji D, Santara A, Mitra P and Sheet D (2016) Ensemble of deep convolutional neural networks for learning to detect retinal ves-sels in fundus images. arXiv preprint arXiv:, 2016. 1603.04833

Martinez-Perez ME, Highes AD, Stanton AV, Thorn SA, Chapman N, Bharath AA, Parker KH (2002) Retinal vascular tree morphol-ogy: a semi-automatic quantification. IEEE Trans Biomed Eng 49(8):912–917

McInerney T, Terzopoulos D (1996) Deformable models in medical image analysis: a survey. Med Image Anal 1(2):91–108

Mirsharif Q, Tajeripour F, Pourreza H (2013) Automated characteri-zation of blood vessels as arteries and veins in retinal images. Comput Med Imaging Graph 37(7–8):607–617

Moghimirad E, Rezatofighi SH and Soltanian-Zadeh H (2010) Multi-scale approach for retinal vessel segmentation using medialness function. In 2010 IEEE international symposium on biomedical imaging: from nano to macro (29–32)

Mohammed MA et al (2017a) Analysis of an electronic methods for nasopharyngeal carcinoma: prevalence, diagnosis, challenges and technologies. J Comput Sci 21:241–254

Mohammed MA et al (2017b) Artificial neural networks for automatic segmentation and identification of nasopharyngeal carcinoma. J Comput Sci 21:263–274

Mohammed MA et al (2018) A real time computer aided object detec-tion of nasopharyngeal carcinoma using genetic algorithm and artificial neural network based on Haar feature fear. Futur Gener Comput Syst 89:539–547

Mostafa SA et al (2019) Examining multiple feature evaluation and classification methods for improving the diagnosis of Parkinson’s disease. Cogn Syst Res 54:90–99

Mou L et al (2021) CS(2)-Net: deep learning segmentation of curvilin-ear structures in medical imaging. Med Image Anal 67:101874

Muramatsu C, Hatanaka Y, Iwase T, Hara T, Fujita H (2010) Auto-mated detection and classification of major retinal vessels for determination of diameter ratio of arteries and veins. In: Medical imaging 2010: computer-aided diagnosis, vol 7624. International Society for Optics and Photonics, p 76240J

Narasimha-Iyer H, Beach JM, Khoobehi B, Roysam B (2007) Auto-matic identification of retinal arteries and veins from dual-wave-length images using structural and functional features. IEEE Trans Biomed Eng 54(8):1427–1435

Naveed K et al (2021) Towards automated eye diagnosis: an improved retinal vessel segmentation framework using ensemble block matching 3D filter. Diagnostics (Basel) 11:1

Nayak J et al (2008) Automated identification of diabetic retinopathy stages using digital fundus images. J Med Syst 32(2):107–115



1 3

of 32 20

Nekovei RASY (1995) Back-propagation network and its configura-tion for blood vessel detection in angiograms. IEEE Transactions Neural Netw 6(1):64–72

Niemeijer M, Xu X, Dumitrescu AV, Gupta P, Van Ginneken B, Folk JC, Abramoff MD (2011) Automated measurement of the arte-riolar-to-venular width ratio in digital color fundus photographs. IEEE Trans Med Imaging 30(11):1941–1950

Obaid OI, Mohammed MA, Ghani MKA, Mostafa A, Taha F (2018) Evaluating the performance of machine learning techniques in the classification of Wisconsin breast cancer. Int J Eng Technol 7(4.36):160–166

Odstrcilik J et al (2013) Retinal vessel segmentation by improved matched filtering: evaluation on a new high-resolution fundus image database. IET Image Proc 7(4):373–383

Patton N et al (2006) Retinal image analysis: concepts, applications and potential. Am J Ophthalmol 141(3):603

Poplin R et al (2018) Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2(3):158–164

Qureshi TA, Habib M, Hunter A and Al-Diri B (2013) A manually-labeled, artery/vein classified benchmark for the DRIVE dataset. In proceedings of the 26th IEEE international symposium on computer-based medical systems (485–488)

Rattathanapad S, Mittrapiyanuruk P, Kaewtrakulpong P, Uyyanon-vara B and Sinthanayothin C (2012) Vessel extraction in retinal images using multilevel line detection. . In proceedings of 2012 IEEE-EMBS international conference on biomedical and health informatics (345–349)

Relan D, Ballerini L, Trucco E, MacGillivray T (2016) Retinal ves-sel classification based on maximization of squared-loss mutual information. In: Machine intelligence and signal processing. Springer, New Delhi, pp 77–84

Relan D, Relan R (2020) Unsupervised sorting of retinal vessels using locally consistent Gaussian mixtures. Comput Methods Programs Biomed 199:105894

Relan D, MacGillivray T, Ballerini L and Trucco E (2013) Retinal ves-sel classification: sorting arteries and veins. In 2013 35th annual international conference of the IEEE engineering in medicine and biology society (EMBC) (7396–7399)

Relan D, MacGillivray T, Ballerini L and Trucco E (2014) Automatic retinal vessel classification using a least square-support vector machine in VAMPIRE. In 2014 36th annual international confer-ence of the IEEE engineering in medicine and biology society (142–145)

Rothaus K, Jiang X (2011) Classification of arteries and veins in retinal images using vessel profile features. In: AIP conference proceed-ings, vol 1371, no 1. American Institute of Physics, pp 9–18

Rothaus K, Rhiem P and Jiang X (2007) Separation of the retinal vas-cular graph in arteries and veins. In International Workshop on Graph-Based Representations in Pattern Recognition. Springer, Berlin, Heidelberg 251–262

Rothaus K, Jiang X, Rhiem P (2009) Separation of the retinal vascular graph in arteries and veins based upon structural knowledge. Image Vis Comput 27(7):864–875

Roy AGASD (2015) Dasa: domain adaptation in stacked autoencoders using systematic dropout. In 2015 3rd IAPR Asian conference on pattern recognition (ACPR). IEEE (735–739)

Ruggeri A, Grisan E and De Luca M (2007) An automatic system for the estimation of generalized arteriolar narrowing in retinal images. In 2007 29th annual international conference of the IEEE engineering in medicine and biology society (6463–6466). IEEE

Saha Tchinda B et al (2021) Retinal blood vessels segmentation using classical edge detection filters and the neural network. Inform Med Unlock 23:100521

Salem SAS, NMAK Nandi (2006) Segmentation of retinal blood ves-sels using a novel clustering algorithm. In proceedings of the

2006 14th European signal processing conference, Florence, Italy, 1–5

Sarah Husham AM, Mostafa SA, Al-Obaidi MK, Mohammed MA (2020) comparative analysis between active contour and otsuthresholding segmentation algorithms in segmenting brain-tumor magnetic resonance imaging. J Inform Technol Manag. https:// doi. org/ 10. 22059/ jitm. 2020. 78889

Saroj SK, Kumar R, Singh NP (2020) Frechet PDF based matched filter approach for retinal blood vessels segmentation. Comput Methods Programs Biomed 194:105490

Sharma SAWEV (2015) Retinal blood vessel segmentation using fuzzy logic. J Netw Commun Emerg Technol 4:3

Singh NP, Srivastava R (2016) Retinal blood vessels segmentation by using Gumbel probability distribution function based matched filter. Comput Methods Programs Biomed 129:40–50

Singh NP, Kumar R and Srivastava R (2015) Local entropy threshold-ing based fast retinal vessels segmentation by modifying matched filter. In international conference on computing, communication and automation IEEE (1166–1170)

Sofka MASCV (2006) Retinal vessel centerline extraction using mul-tiscale matched filters, confidence and edge measures. IEEE Transactions Med Imaging 25(12):1531–1546

Soomro TA et al (2018) Impact of ICA-based image enhancement technique on retinal blood vessels segmentation. IEEE Access 6:3524–3538

Soomro TA et al (2019) Strided fully convolutional neural network for boosting the sensitivity of retinal blood vessels segmentation. Expert Syst Appl 134:36–52

Staal J, Abràmoff MD, Niemeijer M, Viergever MA, Van Ginneken B (2004) Ridge-based vessel segmentation in color images of the retina. IEEE Trans Med Imaging 23(4):501–509

Subathra MSP, Mohammed MA, Maashi MS, Garcia-Zapirain B, Sairamya NJ, George ST (2020) Detection of focal and non-focal electroencephalogram signals using fast Walsh-Hadamard trans-form and artificial neural network. Sensors 20(17):4952

Tankyevych O, Talbot H, Dokladal P (2008) Curvilinear morpho-Hessian filter. IEEE Int Symp Biomed Imaging 2008:1011–1014

Tianyu Ma HZ (2021) Hanley Ong. Ensembling Low Precision Models for Binary Biomedical Image Segmentation, IEEE Access

Tramontan L, Grisan E and Ruggeri A (2008) An improved system for the automatic estimation of the Arteriolar-to-Venular diameter Ratio (AVR) in retinal images. In 2008 30th annual international conference of the IEEE engineering in medicine and biology society (3550–3553). IEEE

Usman A, Muhammad A, Martinez-Enriquez AM, Muhammad A (2020) Classification of diabetic retinopathy and retinal vein occlusion in human eye fundus images by transfer learning. In: Future of information and communication conference. Springer, Cham, pp 642–653

Vázquez SG et al (2012) Improving retinal artery and vein classifi-cation by means of a minimal path approach. Mach Vis Appl 24(5):919–930

Villalobos-Castaldi FM, Felipe-Riverón EM, Sánchez-Fernández LP (2010) A fast, efficient and automated method to extract vessels from fundus images. J Visual 13(3):263–270

Walter T, Klein JC, Massin P, Erginay A (2002) A contribution of image processing to the diagnosis of diabetic retinopathy-detec-tion of exudates in color fundus images of the human retina. IEEE Trans Med Imaging 21(10):1236–1243

Weiss K, Khoshgoftaar TM, Wang D (2016) A survey of transfer learn-ing. J Big Data 3:1

Wu Y, Xia Y, Song Y, Zhang Y, Cai W (2018) Multiscale network fol-lowed network model for retinal vessel segmentation. In: Inter-national conference on medical image computing and computer-assisted intervention. Springer, Cham, pp 119–126



1 3

20 of 32

Wu CH, Agam G and Stanchev P (2007) A general framework for ves-sel segmentation in retinal images. In 2007 international sympo-sium on computational intelligence in robotics and automation. IEEE (37–42)

Wu Y et al (2020) NFN: a novel network followed network for retinal vessel segmentation. Neural Netw 126:153–162

Wang W, Wu X, Yuan X, Gao Z (2020) An experiment-based review of low-light image enhancement methods. IEEE Access 8:87884–87917

Xie S (2013) and H. Retinal Vascular Image Segmentation Using Genetic Algorithm Plus FCM Clustering, Nie, pp 1225–1228

Xiuqin P, Zhang Q, Zhang H, Li S (2019) A fundus retinal vessels segmentation scheme based on the improved deep learning U-Net model. IEEE Access 7:122634–122643

Yang Y, Huang S, Rao N (2008) An automatic hybrid method for retinal blood vessel extraction. Int J Appl Math Comput Sci 18(3):399–407

Yedidya T (2008) and R. Tracking of Blood Vessels in Retinal Images Using Kalman Filter, Hartley, pp 52–58

Yin Y, Adel M, Guillaume M and Bourennane S (2010) A probabilis-tic based method for tracking vessels in retinal images. In 2010 IEEE international conference on image processing (4081–4084)

Zahra Amini HR (2016) Classification of medical image modeling methods: a review. IEEE 12(2):130–148

Zamperini A, Giachetti A, Trucco E and Chin KS (2012) Effective features for artery-vein classification in digital fundus images. In 2012 25th IEEE international symposium on computer-based medical systems (CBMS) (1–6)

Zhang Y, CACS (2018) Deep supervision with additional labels for retinal vessel segmentation task. In: Frangi A, Schnabel J, Davatzikos C, Alberola-López C, Fichtinger G (eds) Medical image computing and computer assisted intervention: MIC-CAI 2018. MICCAI 2018. Lecture Notes in computer science Springer, Cham vol 11071.

Zhang Z, Yin FS, Liu J, Wong WK, Tan NM, Lee BH, Cheng J and Wong TY (2010) Origa-light: an online retinal fundus image database for glaucoma analysis and research. In 2010 annual

international conference of the IEEE engineering in medicine and biology (3065–3068). IEEE

Zhang B et al (2010) Retinal vessel extraction by matched filter with first-order derivative of Gaussian. Comput Biol Med 40(4):438–445

Zhang J, Tang Z, Gui W and Liu J (2015) Retinal vessel image seg-mentation based on correlational open active contours model. . In 2015 Chinese automation congress (CAC). IEEE (993–998)

Zhao Y et al (2015) Automated vessel segmentation using infinite perimeter active contour model with hybrid region information with application to retinal images. IEEE Trans Med Imaging 34(9):1797–1807

Zhou C, Zhang X, Chen H (2020) A new robust method for blood ves-sel segmentation in retinal fundus images based on weighted line detector and hidden Markov model. Comput Methods Programs Biomed 187:105231

Zhu T (2010) Fourier cross-sectional profile for vessel detection on retinal images. Comput Med Imaging Graph 34(3):203–212

Zhu TASG (2011) Retinal vessel extraction using a piecewise Gauss-ian scaled model. In 2011 annual international conference of the IEEE engineering in medicine and biology society, 2011, August. (pp. 5008–5011)

Zhu C et al (2017) Retinal vessel segmentation in colour fundus images using extreme learning machine. Comput Med Imaging Graph 55:68–77

Zolfagharnasab HANNAR (2014) Cauchy based matched filter for retinal vessels detection. J Med Sign Sens 4(1):1

Zou B et al (2020) Multi-label classification scheme based on local regression for retinal vessel segmentation. In: IEEE/ACM trans-actions on computational biology and bioinformatics. https:// doi. org/ 10. 1109/ TCBB. 2020. 29802 33

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.


1.

2.

3.

4.

5.

6.

Terms and Conditions Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”). Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. Byaccessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For thesepurposes, Springer Nature considers academic use (by researchers and students) to be non-commercial. These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personalsubscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used willapply. We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally withinResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will nototherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission asdetailed in the Privacy Policy. While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users maynot:

use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access

control;

use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is

otherwise unlawful;

falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in

writing;

use bots or other automated methods to access the content or redirect messages

override any security feature or exclusionary protocol; or

share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal

content. In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journalcontent cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or anyother, institutional repository. These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information orcontent on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Naturemay revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved. To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or impliedwith respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,including merchantability or fitness for any particular purpose. Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensedfrom third parties. If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner notexpressly permitted by these Terms, please contact Springer Nature at

[email protected]

mailto:[email protected]

Comprehensive review of retinal blood vessel segmentation ...

Documents

Transcript of Comprehensive review of retinal blood vessel segmentation ...