Semantic Segmentation of Retinal Blood Vessels from Fundus Images by

using CNN and the Random Forest Algorithm

Ayoub Skouta

1 a

, Abdelali Elmouﬁdi

2 b

, Said Jai-Andaloussi

1 c

and Ouail Ouchetto

1 d

Computer and Systems Laboratory, Hassan II University, Casablanca, Morocco

Data4Earth Laboratory, Sultan Moulay Slimane University, Beni Mellal, Morocco

{ay.skouta, elmouﬁdi10, andaloussi.said, ouail.ouchetto}@gmail.com

Keywords:

Funds Images, Diabetic Retinopathy, CAD System, Semantic Segmentation, Blood Vessel Detection,

Artiﬁcial Intelligence, Deep Learning, Convolutional Neural Networks.

Abstract:

Abstract: In this paper, we present a new study to improve the automated segmentation of blood vessels in

diabetic retinopathy images. Pre-processing is necessary due to the contrast between the blood vessels and the

background, as well as the uneven illumination of the retinal images, in order to produce better quality data to

be used in further processing. We use data augmentation techniques to increase the amount of accessible data

in the dataset to overcome the data sparsity problem that deep learning requires. We then use the CNN VGG16

architecture to extract the feature from the preprocessed background images. The Random Forest method will

then use the extracted attributes as input parameters. We used part of the augmented dataset to train the model

(1764 images, representing the training set); the rest of the dataset will be used to test the model (196 images,

representing the test set). Regarding the model validation phase, we used the dedicated part for testing the

DRIVE dataset. Promising results compared to the state of the art were obtained. The method achieved an

accuracy of 98.7%, a sensitivity of 97.4% and speciﬁcity of 99.5%. A comparison with some recent previous

work in the literature has shown a signiﬁcant advancement in our proposal.

1 INTRODUCTION

The human eye is a visual organ, similar to a sphere

with a diameter of 2-3 cm, weighing about 8 grams. It

receives the light emitted by objects, focuses it, then

transmits the information to the brain for analysis.

The eye is surrounded by a tough membrane called

the sclera. The main structures of the eyeball are :

The cornea is completely transparent and is located

in front of the eye in front of the iris, but because it

is transparent, it cannot be seen. It can be observed

with a special microscope called a slit lamp. When

we look at an eye, we can see that the iris is colored

and shrinks depending on the light, and that the sclera

is white and opaque and covers the entire back sur-

face of the eye. The crystalline lens is a converging

lens-like organ that reﬂects the image and directs it

to the retina. The ﬁgure shows a human eye with all

its parts (Abr

amoff et al., 2010). The retina is the in-

https://orcid.org/0000-0002-6176-910X

https://orcid.org/0000-0002-8574-9584

https://orcid.org/0000-0002-6864-1141

https://orcid.org/0000-0001-8287-215X

ner membrane of the eyeball. It is a light sensitive

layer that captures light rays. The retina is made up

of three layers of nerve cells: ganglion cells, bipo-

lar cells and visual cells that are either cones or rods.

Light passes through the upper layers of the retina

to reach the cones and rods. After a series of pho-

tochemical reactions, the information is transmitted

to the bipolar cells, then to the ganglion cells and ﬁ-

nally to the optic nerves, which send information to

the brain in the form of electrical signals, which in

turn interpret the signals received as visual images.

The cones are grouped in the center of the retina at

the macula, where visual acuity is best. They are the

color vision cells that work only in well-lit environ-

ments. The rods appear around the macula, they are

the cells of black and white vision they react to very

weak lights. Progressive damage to the macula leads

to diseases such as macular degeneration or, in severe

cases, creates a macular hole, which causes blood ves-

sels in the macula to rupture (Jakobiec, 1982). Dia-

betes is a comprehensive metabolic disorder that can

lead to various vascular complications in the body.

There are two types of diabetes: type 1 diabetes and

type 2 diabetes. Both result in high blood sugar lev-

Skouta, A., Elmouﬁdi, A., Jai-Andaloussi, S. and Ouchetto, O.

Semantic Segmentation of Retinal Blood Vessels from Fundus Images by using CNN and the Random Forest Algorithm.

DOI: 10.5220/0010911800003118

In Proceedings of the 11th International Conference on Sensor Networks (SENSORNETS 2022), pages 163-170

ISBN: 978-989-758-551-7; ISSN: 2184-4380

163

Figure 1: This anatomy of the eye and the retina.

els related to a hormone called insulin produced by

the pancreas; this hormone is responsible for the reg-

ulation of blood sugar levels in the blood (Breiman,

2001a). Diabetic retinopathy is a microangiopathy

that has the manifestations of occlusion and leakage

of microvascular ﬂuid and blood in the retina, it is es-

sential to understand the signs of occlusion and leak-

age in the retina before understanding the pathogen-

esis and sign of diabetic retinopathy. In the presence

of diabetes and hyperglycemia, several things happen

in the blood vessels, the blood vessel walls that help

nourish the retina deteriorate, the blood cells become

distorted and the blood thickens and then ﬁnally a mi-

crovascular occlusion as there will be irregularities in

the blood ﬂow and a decrease in oxygen. This re-

sults in visual manifestations called lesions such as

microaneurysms, hemorrhages, hard exudates, cot-

tony spots (Agurto et al., 2010). This phenomenon

due to diabetes can progressively damage the struc-

ture of the eyeball, leading to severe vision loss or

even blindness (Kanski and Bowling, 2011). Several

studies have focused on the use of automated tech-

niques for the early diagnosis of diabetic retinopa-

thy. These technologies are becoming increasingly

advantageous in the healthcare sector, thanks to ar-

tiﬁcial intelligence (AI), which allows the creation

of reliable software for decision support in the diag-

nosis of medical images (Tham Chen2014g, ). The

use of traditional machine learning (ML) approaches

such as random forest algorithm, decision tree, near-

est neighbor method (KNN), K-means, support vector

method (SVM), etc. (Kotsiantis et al., 2007), (Liaw

et al., 2002) are used to analyze fundus images to de-

velop models that can predict outcomes without pro-

gramming, through operations distributed over mul-

tiple layers (Kapoor et al., 2019). Machine learn-

ing techniques are divided into two categories: su-

pervised and unsupervised learning. The supervised

technique requires labeled data to train the segmenta-

tion model, but the unsupervised method requires no

prior knowledge of the labeled data to train the seg-

mentation model. Since the introduction of convo-

lutional neural networks (CNNs), deep learning has

been remarkably successful, and it is now considered

the most successful AI model in all computer vision

applications. It has proven to be very effective and

useful in extracting features for use in various ma-

chine learning applications. The input layer, hidden

layers and output layer are the three types of layers

found in CNNs. These layers consist of linked nodes,

each with its own output layer sending a weighted

quantity to the activation function. Several strategies

have been established to improve the quality of vas-

cular segmentation, using classical machine learning

or deep learning methods, but more can be done. The

second section of this paper contains an overview of

related previous work, and then the proposed method

and materials used are described in the third section

of this study. The results of our experiment are de-

scribed in Section 4. A discussion is presented in Sec-

tion Five. Finally, the main conclusions are presented

in the last section.

2 RELATED PREVIOUS WORKS

In the medical domain, many computer-aided diag-

noses have used to help in diagnosing of many de-

ceases. Even more, many diseases are early-detected

using artiﬁcial intelligence. And one of the newest

examples is the creation of two models one to clas-

sify a COVID-19 test of a suspect patient as posi-

tive or negative, and the other to classify the hospital-

ization units of patients with COVID-19 (de Oliveira

et al., 2021). Others diseases are breast cancer (El-

mouﬁdi et al., 2018), (Elmouﬁdi et al., 2014), (Elm-

ouﬁdi et al., 2015), (Elmouﬁdi, 2019) and diabetic

retinopathy (Skouta et al., 2021), (El Hossi et al.,

2021), (Stabingis et al., 2018), (Balkys and Dzemyda,

2012). In this paper, the purpose of determining reti-

nal vascular segmentation is to predict abnormali-

ties that can occur in the retina, including diabetic

retinopathy and glaucoma. The advent of convolu-

tional neural networks has led to improvements in re-

ducing the burden on specialists with screening pro-

grams that have begun to generate high-accuracy seg-

mentation (Salazar-Gonzalez et al., 2014). Accord-

ing to the literature presented in this section, different

methods are applied to segment blood vessels in fun-

dus images. In this paper, we focus on the use of deep

CNNs, which have shown excellent performance in

their application. The remainder of this section will

be devoted to presenting related work on blood vessel

segmentation from retinal fundus images; for exam-

ple, the approach proposed by Khalaf et al (Khalaf

SENSORNETS 2022 - 11th International Conference on Sensor Networks

164

et al., 2016) used a CNN to segment vessels, using an

input patch derived from the original image as input.

Three convolutional layers and a fully connected layer

at the end of the network constitute the CNN. Their

technique classiﬁes each patch into three categories

based on the center pixel: background, large vessels,

and small vessels. Liskowski and Krawiec (Liskowski

and Krawiec, 2016) presented another patch-based

approach to determine the number of neurons in the ﬁ-

nal fully connected layer. Background and vessels are

the two categories predicted by CNN. Segmentation

is performed by connecting each pixel to a neuron, so

that the CNN output is a two-dimensional vector. Yu

et al (Yu et al., 2020) developed a CNN for hierar-

chical vessel division by ﬁrst extracting vessel trees

using a graph-based method and then classifying the

retinal vasculature using two algorithms (PLDA and

AHCA). The ﬁnal layers of the CNN are fully con-

nected layers. Shuangling Wang et al (Wang et al.,

2015) designed a retinal blood vessel segmentation

technique that combines a CNN as a feature extractor

from raw images and a Random Forest (RF) algorithm

as a classiﬁer. Zhou et al (Zhou et al., 2017) propose

a CNN solution to extract blood vessel features. The

extracted features are then provided to the dense CRF

to perform segmentation. The DeepVessel CNN pub-

lished by Fu et al. (Fu et al., 2016) combines the out-

put side layer and the CRF layer to model long-range

pixel interactions and nonlocal pixel correlations. Hu

et al (Hu et al., 2018) wrote a paper that performs

binary segmentation by describing an RCF-inspired

multi-scale CNN architecture that provides a com-

prehensive description of vascular features combined

with an improved cross-entropy loss function. The

emergence of the U-net architecture in 2015 (Ron-

neberger et al., 2015) has made a great contribution in

the ﬁeld of biomedical image segmentation. It has be-

come popular in blood vessel segmentation. Several

authors have customized this promising method to de-

velop modiﬁed versions; Zhang and Chung (Zhang

and Chung, 2018), Yan et al. (Yan et al., 2018), Yan

et al. (Yang et al., 2017), Wu et al. (Wu et al., 2018),

Wu et al. (Wu et al., 2020a), Wang et al. (Wang et al.,

2020), Wu et al. (Wu et al., 2019), Wang et al. (Wang

et al., 2019), Ma et al. (Ma et al., 2019), Mishra et al.

(Mishra et al., 2020).

2.1 Materials

In this section, we outline the hardware used, the

datasets used and the proposed methodology. We

tested our solution in a GPU environment using the

Keras library and TensorFlow.

2.2 Databases Used

The suggested approach has been trained, validated

and tested using the Drive, HRF datasets. These

datasets are freely accessible and are used in the

vast majority of vascular segmentation studies. The

DRIVE (Digital Retinal Image for Vessel Extraction)

dataset is a series of 20 images that includes a manual

segmentation of vessels for each image. There are 20

more photos in the test set. The images have a size of

768 x 584 pixels. the exception of seven photos that

showed slight symptomatic RD, all images DRIVE

were normal (Staal et al., 2004).

Figure 2: Left: Examples of fundus from the DERIVE

database. Right: Examples of ground truth data from the

DERIVE database.

Table 1: Presentation of the databases implemented in the

proposed approach.

Datasets Source Images Digitizer

Drive

Avail-

able:

Online

Staal

et al.

(Wu

et al.,

2019)

40 (33

healthy, 7

mild early

DR) Res-

olutions :

768x584

Canon: CR5

non-

mydriatic

3CCD

fundus

camera,

FOV 45

2.3 Image Preprocessing

The high similarity in the fundus image between the

vessels and the background can lead to erroneous seg-

mentation. The data is optimized in the preprocessing

step to provide a better image with a clear distinc-

tion between the vessels and the background, helping

CNN to interpret the input images and allows for cor-

rect segmentation. This step starts by cropping the

black border of the original images. After the orig-

inal color image is cropped, the color order is BGR

(blue, green, red) instead of RGB, so the ﬁrst step

is to convert the image from BGR to RGB. We then

extract the green channel from the multichannel im-

age after splitting it into several channels, which gives

more contrast between the blood vessels and the back-

Semantic Segmentation of Retinal Blood Vessels from Fundus Images by using CNN and the Random Forest Algorithm

165

ground than the red and blue channels. The green

channel is then smoothed and noise is reduced us-

ing a bilateral ﬁlter while vascular contours are pre-

served. Vascular information is further highlighted by

applying contrast-limited adaptive histogram equal-

ization (CLAHE) to the image generated by the bi-

lateral ﬁlter. We used the image sharpening method

to highlight each individual pixel, enhance the color

it emanates, and increase the pixel density, making it

more visible to CNN. Finally, a gamma correction, to

correct the retinal image in order to remove the un-

even light factor. The results of the pre-processing

are shown in Figure 3.

Figure 3: Result of preprocessing step.

2.4 Data Augmentation

The step of data augmentation is proposed of this

study; it consists in modifying the existing pre-

processed images in the original database as well as

its associated mask in order to increase the size of the

dataset that will be used for the training and testing

of the proposed method. Data augmentation strate-

gies reduce overﬁtting and give the proposed model

more power while also enhancing accuracy and re-

silience. Having a CNN with a huge dataset increases

the model’s capacity to generalize to new pictures

from other databases. To achieve this, we employ

a number of techniques to generate various kinds of

sample while preserving the attributes of the source

images: The original images and the images obtained

after the preprocessing stage, together with their ac-

companying ground-truth masks, are rotated from 0

to 360

◦

at a 30

◦

angle each time. Randomly, we add

noise, change the brightness, change the colorimetry,

vertical and horizontal ﬂips and horizontal and verti-

cal ﬂips to each image created by the loop that rotates

the image from 0 to 360 degrees. These are the trans-

formations used in the data update process. The total

data recovered from the original photos and the im-

ages acquired after the pre-processing step is equal

to 1960 images, which are divided into 1764 images,

representing the training set (90 percent of the gener-

ated images) and 196 images, representing the vali-

dation set (ten percent of the total generated images).

Figure 4: An example of images after the data augmentation

procedure.

3 PROPOSED METHODOLOGY

3.1 Feature Extraction by CNN

Segmentation is the classiﬁcation of pixels, where

each pixel, rather than the whole picture, is assigned

to a distinct category. The feature extraction approach

is critical for a successful segmentation process. Ex-

traction is a critical operation because the features

used to deﬁne the candidate areas have a direct impact

on the accuracy with which each pixel in the input

image is classiﬁed as a blood vessel or background.

The CNN has recently demonstrated the capacity to

recognize the most sophisticated, basic, and signif-

icant visual elements such as edges, corners, orien-

tated edges, and so on. This is crucial information

for the analysis and categorization of pixels. Follow-

ing that, subsequent layers integrate these character-

istics to capture higher order features (Wang et al.,

2019). In this technique, the CNN VGG16 architec-

ture is utilized as a feature extractor for pixels in terms

of quantiﬁable metrics that can be used in the clas-

siﬁcation stage to determine if the pixels correspond

to real blood from the vessel or not. The Random

SENSORNETS 2022 - 11th International Conference on Sensor Networks

166

Forest classiﬁer (Breiman, 2001b) takes the role of

the VGG16 model’s last fully connected layer. This

method is known as transfer learning, and it involves

merely training the top level of the network. The

VGG16 architecture is composed of 5 blocks with a

total of 16 layers. Essentially it is based on the use of

3x3 convolution with stride equal to 1 and a ”Same”

padding, i.e. the size of the input image matches

the size of the output image. After each convolution

layer, a max 2x2 pooling layer is used to reduce the

size.

3.1.1 Segmentation using the Random Forest

Algorithm

Our goal is to take pretrained VGG16 weights used as

feature extractors and then segment the retinal images

using a random drill. The images resulting from the

data augmentation operation are physically stored on

the hard disk of our computer, and then we have to

load all the images and their associated masks. Since

CNN is suitable for large images, we kept the origi-

nal size of the images in our training database, which

have a height of 584 and width of 565, and then they

are converted from RGB to BGR because the opencv

library simply reads the images as BGR. Next, we im-

ported the VGG16 model, so now we will also import

the weights corresponding to the weights in the im-

agenet database. The include top variable is set to

false, this basically means that we have not imported

the dense layers and the output layer. The size of the

input images of the VGG16 model is changed as the

model defaults to 244x244x3 at the size of our input

images. The next step is to apply the feature extrac-

tor to our training data and see what the features look

like. As shown in Figure 5, the features of the origi-

nal augmented images and the preprocessed and aug-

mented images are merged into a bag of features. The

Figure 5: Displays of extracted characteristics.

number of features generated is glaring, therefore we

have eliminated features such as the features that rep-

resent the background of the retinal image. These fea-

tures are useless to use in training, as they will require

additional training time. The pixels that represent the

blood vessels are kept as features. After the feature

reduction operation we feed the random forest algo-

rithm to segment the retinal blood vessels.

4 EXPERIMENTAL RESULT

4.1 Evaluation Metrics

We examine the performance of the segmentation

ﬁndings with relation to the expert’s manual segmen-

tation using the following performance measures: ac-

curacy, sensitivity, and speciﬁcity as indicators.

Table 2: Presentation of the databases implemented in the

proposed approach.

Metrics Formule Description

Sensitivity

T P

T P + FN

The ratio of correctly cat-

egorized vascular pixels

compared to genuine vas-

cular pixels is known as

the true positive rate.

Speciﬁcity

T N

T N + FP

As compared to real non-

vascular pixels, the frac-

tion of correctly diag-

nosed non-vascular pix-

els.

Accuracy

T N + T P

T P + FP + TN + FN

As a proportion of the to-

tal number of pixels in the

picture, the accuracy rep-

resents how well blood

• True positives (TP) refers to the number of accu-

rately segmented blood vessel pixels;

• The amount of accurately divided background

pixels is shown by True Negatives (TN);

• False positives (FP) are background pixels that

have been segmented incorrectly into blood ves-

sel pixels;

• False negatives (FN) are pixels in blood vessels

that have been mistakenly labeled background.

4.2 Results

We used the DRIVE dataset to train the suggested

models in 100 epochs. The results of our method’s

segmentation are more substantial. The accuracy is

98.7%, the sensitivity is 97.4%, and the speciﬁcity

is 99.5 percent. Training details are depicted in the

Semantic Segmentation of Retinal Blood Vessels from Fundus Images by using CNN and the Random Forest Algorithm

167

Figure 6: Model accuracy performance.

Figure 7: Model loss performance.

ﬁgures, which also show the accuracy and losses at-

tained.

We used a subset of the DRIVE database dedi-

cated to testing. Our primary aim is to verify the accu-

racy of our suggested technique. We examine the de-

gree of similarity between the segmentations obtained

using our proposed network and the ground truth.

Table 3: Comparison of segmentation performances for

DRIVE.

Works Year Acc Sen Spe

(Tamim

et al., 2020)

2020 0.9607 0.7542 0.9843

(Tian et al.,

2020)

2020 0.958 0.8639 0.9690

(Wu et al.,

2020b)

2020 0.9582 0.7996 0.9813

(Boudegga

et al., 2021)

2021 0.9819 0.8448 0.99

Proposed

method

2021 98.7 97.4 99.5

Acc: Accuracy Sen: Sensitivity Spe: Speciﬁcity

5 CONCLUSIONS

Early treatment of diabetic retinopathy with blood

vessel segmentation helps people with diabetes avoid

severe visual loss. Deep learning is one of the most

Figure 8: Comparisons of segmentation results using the

test subset of the DRIVE database.

sophisticated methods for segmentation challenges,

as it improves accuracy. The effective convolutional

neural network architecture used will help ophthal-

mologists to eradicate vision loss related to diabetic

retinopathy. In this research, we suggest the use of a

VGG16 model to extract characteristics and combine

it with the random forest technique to automate blood

vessel segmentation. Our technique was developed

from the DRIVE dataset, which has been shown to be

resilient.

REFERENCES

Michael D Abr

amoff, Mona K Garvin, and Milan Sonka.

Retinal imaging and image analysis. IEEE reviews in

biomedical engineering, 3:169–208, 2010.

Frederick A Jakobiec. Ocular anatomy, embryology, and

teratology. Harpercollins, 1982.

Leo Breiman. Random forest, vol. 45. Mach Learn, 1,

2001a.

Carla Agurto, Victor Murray, Eduardo Barriga, Sergio

Murillo, Marios Pattichis, Herbert Davis, Stephen

Russell, Michael Abr

amoff, and Peter Soliz. Multi-

scale am-fm methods for diabetic retinopathy lesion

detection. IEEE transactions on medical imaging, 29

(2):502–512, 2010.

Jack J Kanski and Brad Bowling. Clinical ophthalmol-

ogy: a systematic approach. Elsevier Health Sciences,

2011.

SENSORNETS 2022 - 11th International Conference on Sensor Networks

168

Sotiris B Kotsiantis, I Zaharakis, P Pintelas, et al. Su-

pervised machine learning: A review of classiﬁcation

techniques. Emerging artiﬁcial intelligence applica-

tions in computer engineering, 160(1):3–24, 2007.

Andy Liaw, Matthew Wiener, et al. Classiﬁcation and re-

gression by randomforest. R news, 2(3):18–22, 2002.

Rahul Kapoor, Stephen P Walters, and Lama A Al-Aswad.

The current state of artiﬁcial intelligence in ophthal-

mology. Survey of ophthalmology, 64(2):233–240,

2019.

Rodrigo Felipe Albuquerque Paiva de Oliveira, Carmelo

Jose Albanez Bastos Filho, Ana Clara AMVF

de Medeiros, Pedro Jose Buarque Lins dos Santos,

and Daniela Lopes Freire. Machine learning applied

in sars-cov-2 covid 19 screening using clinical analy-

sis parameters. IEEE Latin America Transactions, 19

(6):978–985, 2021.

Abdelali Elmouﬁdi, Khalid El Fahssi, Said Jai-Andaloussi,

Abderrahim Sekkaki, Quellec Gwenole, and Mathieu

Lamard. Anomaly classiﬁcation in digital mammog-

raphy based on multiple-instance learning. IET Image

Processing, 12(3):320–328, 2018.

Abdelali Elmouﬁdi, Khalid El Fahssi, Said Jai-Andaloussi,

Nabil Madrane, and Abderrahim Sekkaki. Detection

of regions of interest’s in mammograms by using lo-

cal binary pattern, dynamic k-means algorithm and

gray level co-occurrence matrix. In 2014 Interna-

tional Conference on Next Generation Networks and

Services (NGNS), pages 118–123. IEEE, 2014.

Abdelali Elmouﬁdi, Khalid El Fahssi, Said Jai-Andaloussi,

and Abderrahim Sekkaki. Automatically density

based breast segmentation for mammograms by us-

ing dynamic k-means algorithm and seed based re-

gion growing. In 2015 IEEE International Instru-

mentation and Measurement Technology Conference

(I2MTC) Proceedings, pages 533–538. IEEE, 2015.

Abdelali Elmouﬁdi. Pre-processing algorithms on digital x-

ray mammograms. In 2019 IEEE International Smart

Cities Conference (ISC2), pages 87–92. IEEE, 2019.

Ayoub Skouta, Abdelali Elmouﬁdi, Said Jai-Andaloussi,

and Ouail Ochetto. Automated binary classiﬁcation

of diabetic retinopathy by convolutional neural net-

works. In Advances on Smart and Soft Computing,

pages 177–187. Springer, 2021.

Amine El Hossi, Ayoub Skouta, Abdelali Elmouﬁdi, and

Mourad Nachaoui. Applied cnn for automatic diabetic

retinopathy assessment using fundus images. In Inter-

national Conference on Business Intelligence, pages

425–433. Springer, 2021.

Giedrius Stabingis, Jolita Bernatavi

cien

e, Gintautas Dze-

myda, Alvydas Paunksnis, Lijana Stabingien

e, Povi-

las Treigys, and Ramut

e Vai

caitien

e. Adaptive eye

fundus vessel classiﬁcation for automatic artery and

vein diameter ratio evaluation. Informatica, 29(4):

757–771, 2018.

Gediminas Balkys and Gintautas Dzemyda. Segmenting the

eye fundus images for identiﬁcation of blood vessels.

Mathematical Modelling and Analysis, 17(1):21–30,

2012.

Ana Salazar-Gonzalez, Djibril Kaba, Yongmin Li, and Xi-

aohui Liu. Segmentation of the blood vessels and op-

tic disk in retinal images. IEEE journal of biomedical

and health informatics, 18(6):1874–1886, 2014.

Aya F Khalaf, Inas A Yassine, and Ahmed S Fahmy. Con-

volutional neural networks for deep feature learning

in retinal vessel segmentation. In 2016 IEEE Interna-

tional Conference on Image Processing (ICIP), pages

385–388. IEEE, 2016.

Paweł Liskowski and Krzysztof Krawiec. Segmenting reti-

nal blood vessels with deep neural networks. IEEE

transactions on medical imaging, 35(11):2369–2380,

2016.

Linfang Yu, Zhen Qin, Tianming Zhuang, Yi Ding,

Zhiguang Qin, and Kim-Kwang Raymond Choo. A

framework for hierarchical division of retinal vascular

networks. Neurocomputing, 392:221–232, 2020.

Shuangling Wang, Yilong Yin, Guibao Cao, Benzheng Wei,

Yuanjie Zheng, and Gongping Yang. Hierarchical reti-

nal blood vessel segmentation based on feature and

ensemble learning. Neurocomputing, 149:708–717,

2015.

Lei Zhou, Qi Yu, Xun Xu, Yun Gu, and Jie Yang. Improv-

ing dense conditional random ﬁeld for retinal vessel

segmentation by discriminative feature learning and

thin-vessel enhancement. Computer methods and pro-

grams in biomedicine, 148:13–25, 2017.

Huazhu Fu, Yanwu Xu, Stephen Lin, Damon Wing Kee

Wong, and Jiang Liu. Deepvessel: Retinal vessel seg-

mentation via deep learning and conditional random

ﬁeld. In International conference on medical image

computing and computer-assisted intervention, pages

132–139. Springer, 2016.

Kai Hu, Zhenzhen Zhang, Xiaorui Niu, Yuan Zhang, Chun-

hong Cao, Fen Xiao, and Xieping Gao. Retinal vessel

segmentation of color fundus images using multiscale

convolutional neural network with an improved cross-

entropy loss function. Neurocomputing, 309:179–191,

2018.

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-

net: Convolutional networks for biomedical image

segmentation. In International Conference on Medi-

cal image computing and computer-assisted interven-

tion, pages 234–241. Springer, 2015.

Yishuo Zhang and Albert CS Chung. Deep supervision with

additional labels for retinal vessel segmentation task.

In International conference on medical image comput-

ing and computer-assisted intervention, pages 83–91.

Springer, 2018.

Zengqiang Yan, Xin Yang, and Kwang-Ting Cheng. A

three-stage deep learning model for accurate retinal

vessel segmentation. IEEE journal of Biomedical and

Health Informatics, 23(4):1427–1436, 2018.

Yehui Yang, Tao Li, Wensi Li, Haishan Wu, Wei Fan, and

Wensheng Zhang. Lesion detection and grading of di-

abetic retinopathy via two-stages deep convolutional

neural networks. In International Conference on Med-

ical Image Computing and Computer-Assisted Inter-

vention, pages 533–540. Springer, 2017.

Yicheng Wu, Yong Xia, Yang Song, Yanning Zhang, and

Weidong Cai. Multiscale network followed network

model for retinal vessel segmentation. In Inter-

national Conference on Medical Image Computing

Semantic Segmentation of Retinal Blood Vessels from Fundus Images by using CNN and the Random Forest Algorithm

169

and Computer-Assisted Intervention, pages 119–126.

Springer, 2018.

Yicheng Wu, Yong Xia, Yang Song, Yanning Zhang, and

Weidong Cai. Nfn+: A novel network followed net-

work for retinal vessel segmentation. Neural Net-

works, 126:153–162, 2020a.

Kun Wang, Xiaohong Zhang, Sheng Huang, Qiuli Wang,

and Feiyu Chen. Ctf-net: Retinal vessel segmenta-

tion via deep coarse-to-ﬁne supervision network. In

2020 IEEE 17th International Symposium on Biomed-

ical Imaging (ISBI), pages 1237–1241. IEEE, 2020.

Yicheng Wu, Yong Xia, Yang Song, Donghao Zhang,

Dongnan Liu, Chaoyi Zhang, and Weidong Cai.

Vessel-net: retinal vessel segmentation under multi-

path supervision. In International Conference on

Medical Image Computing and Computer-Assisted In-

tervention, pages 264–272. Springer, 2019.

Bo Wang, Shuang Qiu, and Huiguang He. Dual encod-

ing u-net for retinal vessel segmentation. In In-

ternational Conference on Medical Image Comput-

ing and Computer-Assisted Intervention, pages 84–

92. Springer, 2019.

Wenao Ma, Shuang Yu, Kai Ma, Jiexiang Wang, Xinghao

Ding, and Yefeng Zheng. Multi-task neural networks

with spatial activation for retinal vessel segmentation

and artery/vein classiﬁcation. In International Con-

ference on Medical Image Computing and Computer-

Assisted Intervention, pages 769–778. Springer, 2019.

Suraj Mishra, Danny Z Chen, and X Sharon Hu. A data-

aware deep supervised method for retinal vessel seg-

mentation. In 2020 IEEE 17th International Sym-

posium on Biomedical Imaging (ISBI), pages 1254–

1257. IEEE, 2020.

Joes Staal, Michael D Abr

amoff, Meindert Niemeijer,

Max A Viergever, and Bram Van Ginneken. Ridge-

based vessel segmentation in color images of the

retina. IEEE transactions on medical imaging, 23(4):

501–509, 2004.

Leo Breiman. Random forest, vol. 45. Mach Learn, 1,

2001b.

Nasser Tamim, Mohamed Elshrkawey, Gamil Abdel Azim,

and Hamed Nassar. Retinal blood vessel segmenta-

tion using hybrid features and multi-layer perceptron

neural networks. Symmetry, 12(6):894, 2020.

Chun Tian, Tao Fang, Yingle Fan, and Wei Wu. Multi-path

convolutional neural network in fundus segmentation

of blood vessels. Biocybernetics and Biomedical En-

gineering, 40(2):583–595, 2020.

Yicheng Wu, Yong Xia, Yang Song, Yanning Zhang, and

Weidong Cai. Nfn+: A novel network followed net-

work for retinal vessel segmentation. Neural Net-

works, 126:153–162, 2020b.

Henda Boudegga, Yaroub Elloumi, Mohamed Akil,

Mohamed Hedi Bedoui, Rostom Kachouri, and

Asma Ben Abdallah. Fast and efﬁcient retinal blood

vessel segmentation method based on deep learning

network. Computerized Medical Imaging and Graph-

ics, 90:101902, 2021.

SENSORNETS 2022 - 11th International Conference on Sensor Networks

170