Deep Learning Residual-like Convolutional Neural Networks for Optic

Disc Segmentation in Medical Retinal Images

Amir Hossein Panahi

1 a

, Reza Askari Moghadam

1 b

and Kurosh Madani

Faculty of New Sciences and Technologies, University of Tehran, Tehran, Iran

LISSI Lab, Senart-FB Institute of Technology, University Paris Est-Creteil (UPEC), Lieusaint, France

Keywords:

Deep Learning, Residual-like CNN, Computer Vision, Image Segmentation, Glaucoma Detection, Eye

Fundus, Optic Disc Segmentation, Medical Application.

Abstract:

Eye diseases such as glaucoma, if undiagnosed in time, can have irreversible detrimental effects, which can

lead to blindness. Early detection of this disease by screening programs and subsequent treatment can prevent

blindness. Deep learning architectures have many applications in medicine, especially in medical image pro-

cessing, that provides intelligent tools for the prevention and treatment of diseases. Optic disk segmentation is

one of the ways to diagnose eye disease. This paper presents a new approach based on deep learning, which is

accurate and fast in optic disc segmentation. By Comparison proposed method with the best-known methods

on publicly available databases DRIONS-DB, RIM-ONE v.3, the proposed algorithm is much faster, which

can segment the optic disc in 0.008 second with outstanding performance concerning IOU and DICE scores.

Therefore, this method can be used in ophthalmology clinics to segment the optic disc in retina images and

videos as online medical assistive tool.

1 INTRODUCTION

Digital retinal fundus images are used for the primary

exploration of ophthalmic. Glaucoma is amongst the

main retinal illness, which is the cause of vision loss

and blindness in the world (Federation, 2013). Early

detection of this disease by screening programs and

subsequent treatment can prevent blindness. Com-

puter systems are beneﬁcial for diagnostic retinal im-

age analysis and can be the ﬁrst phase in automated

screening (Fraz et al., 2015). Glaucoma is the sec-

ond most important reason of blinding in recent years.

Based on research about 80 million persons to be dis-

turbed with glaucoma by the year 2020 (Gao et al.,

2019; Quigley and Broman, 2006). The optic nerve

ﬁbers damaged by glaucoma cannot be recovered. So

the most effective way is early detection to avoid in-

jury of retina vessels and nerve ﬁbers. The reason

of glaucoma is commonly dependent on the increase

of Intraocular Pressure (IOP) in the eye, which re-

sults from obstruction of intraocular ﬂuid (Xu et al.,

2012). The correct reason of this obstruction in most

of the time is unknown, but the other factors like old

age, steroid medication will affect the disease (Jack-

https://orcid.org/0000-0002-4624-1219

https://orcid.org/0000-0001-8394-7256

son and Radhakrishnan, 2014). The optic nerve car-

ries the data from the eye to the brain. By increasing

the IOP, the optic nerve damaged. Glaucoma does not

represent any signs until it has developed to advanced

steps (Bajwa et al., 2019). Nevertheless, if glaucoma

is recognized early, it is possible to reduce the disor-

der. World Health Organization (WHO) announces

glaucoma as the second biggest cause of blindness

in the world whose effects lead to irreversible vision

(Bourne et al., 2017). Glaucoma is usually deter-

mined by taking the medical history of a sick person

and assessment manually of Optic Disc (OD) using

ophthalmology to evaluate the conﬁguration and col-

oration of the optic nerve (Chen et al., 2015). Op-

tic Disc is the region of the optic nerve connecting

to the retina of each eye. In the case of glaucoma,

the intraocular pressure damages the nerve ﬁbers, and

the optic disc begins to deform, and color changes

to pale (Xu et al., 2012). In Figure 1 a healthy op-

tic disc with three various steps of glaucoma shown

sequentially. Cup-to-Disc Ratio (CDR), Optic disc

size, Ratio of Neuroretinal Rim, etc., are some of the

signiﬁcant architectonic signs of glaucoma in retinal

fundus images (Abbas, 2017). These signs are usu-

ally around the optic disc, which is Region Of Interest

(ROI). Thus, segmentation of this region, which is de-

Panahi, A., Moghadam, R. and Madani, K.

Deep Learning Residual-like Convolutional Neural Networks for Optic Disc Segmentation in Medical Retinal Images.

DOI: 10.5220/0009799100230029

In Proceedings of the 1st International Conference on Deep Learning Theory and Applications (DeLTA 2020), pages 23-29

ISBN: 978-989-758-441-1

(a) (b) (c) (d)

Figure 1: Glaucoma in retinal fundus images. (a)normal disc, (b)glaucoma onset, (c)critical glaucoma, (d)advance glaucoma.

tecting the optic disc, is helpful for clinical evaluation

by the ophthalmologists. Nevertheless, automated op-

tic disc segmentation methods that used for glaucoma

detection should be sensitive. Cause a small error in

recognize Of optic disc may affect the diagnosis and

treatment seriously (Mookiah et al., 2012). Image

segmentation is a fundamental part of many optical

understanding systems that includes division images

into numerous segments or regions (Szeliski, 2010).

Image segmentation has been used in many applica-

tions, especially in medical image analysis such as

tumor boundary extraction and optic disc segmenta-

tion(Forsyth and Ponce, 2002). The other method for

optic disc segmentation is utilizing a novel vibrational

level set function on the red channel of the retinal fun-

dus images (Wong et al., 2008). In another algorithm,

localized the optic disc was applied by using template

matching; after that, morphological ﬁltering removed

the blood vessels. At last, the boundary information

combines with the local edge vector to operate the de-

formable contour was used to detect the optic disc re-

gions (Yu et al., 2012; Zhang et al., 2008). For op-

tic disc segmentation task, a method based on math-

ematical morphology is proposed to detect and seg-

ment the optic disc in images (Welfer et al., 2010).

This method is expanded by combining a multiscale

morphological approach (Welfer et al., 2013). A

template-based approach for OD segmentation is uti-

lized edge detection and morphological methods con-

formed by circular Hough transformation to estimated

circular objects (Aquino et al., 2010). However, in the

past few years, deep learning networks have used as

a new efﬁciency method in image segmentation tasks

with an extraordinary performance that attaining the

highest accuracy and speed rates. The deep convo-

lutional neural network can extract indicated features

from the input images automatically. There are var-

ious models developed for medical image segmenta-

tion, which based on FCNs (Long et al., 2015) mod-

els. A U-shaped convolutional neural network was

proposed to segment optic disc, and advancement was

achieved by in comparison with the exert of old meth-

ods (Sevastopolsky, 2017). For optic disc segmenta-

tion task, the polar transformation and multi-label loss

function method were applied in a U-shaped (Fu et al.,

2018). The team extended this algorithm one year

later and suggested a Stack-U-Net network architec-

ture (Sevastopolsky et al., 2018), which is based on a

U-Net (Ronneberger et al., 2015) network. The other

optic disc segmentation method has a U-Shape with

Densely connected convolutional blocks (Al-Bander

et al., 2018), based on DenseNet (Huang et al., 2017).

In optic disc segmentation tasks, low time and high

accuracy are essential. In this paper, a network de-

signed based on deep learning and segmented the op-

tic disc with the low time, which can help the oph-

thalmology clinic to evaluate the retina disease like

glaucoma.

2 THE PROPOSED METHOD

This paper develops a deep learning algorithm for op-

tic disc segmentation and designs a new network ar-

chitecture based on a residual model. In this paper, a

new approach called residual-like convolutional neu-

ral network applied for optic disc segmentation in reti-

nal color fundus images. In the structure of this net-

work, there are some layers which based on residual

layers in ResNet (He et al., 2016). For optic disc seg-

mentation, in pixel-level where features can be take

out from various sized windows, but at the identical

time, passing some features from ﬁrst layers to resul-

tant layers, as residual layers do, should be useful.

Residual block structure seems to be suitable for such

roles when needed to construct data that be similar

to the input image. Another beneﬁt of using residual

blocks is the amplify and improved gradient circula-

tion, which has a positive effect on network conver-

gence. Figure 2 is the residual layer that proposed in

this paper. As shown in Figure 1 in this paper, in this

convolutional network, the RELU activation function

is used after the end of each convolution layer, and

after that batch normalization is used. The Relu func-

tion gives an output x if x is positive and 0 otherwise.

Also, the Sigmoid activation function is applied to the

DeLTA 2020 - 1st International Conference on Deep Learning Theory and Applications

Figure 2: Residual layer used in proposed network.

last layer. As shown in Figure 2 the network at ﬁrst,

extracts the features from an input image and then up-

samples the feature map. In this paper, the output of

the proposed network is a binary image shown in Fig-

ure 5. For evaluating the results, loss function deﬁned

as:

l(A, B) = −logd (A, B) (1)

A is a predicted output, comprising probabilities

that each predicted pixel appertains to the foreground,

and also B is a correct binary output. For binary im-

ages d(A; B) is an expansion of Dice score. Dice

score calculates the expanse of overlapping regions

between any two images. Dice (Dice, 1945) to gauge

the similarity of two samples, such as image and

ground truth, is deﬁned as:

Dice(A, B) =

2|A ∩ B|

(|A| + |B|)

(2)

The ranges value of the Dice coefﬁcient is be-

tween 0 and 1. In this study RMSprop (Tieleman

and Hinton, 2012) optimizer is used with a learn-

ing rate of 0.0002. There are different datasets

Figure 3: Intersection-over-Union (IOU).

for optic disc segmentation. In this paper, we have

used two well-known datasets and then compare the

results of the proposed method with the other meth-

ods that have used these datasets. As shown in Fig-

ure 4 DRIONS-DB (Carmona et al., 2008) and RIM-

ONE v.3 (Fumero et al., 2011) (110 and 159 im-

ages,respectively) datasets used to evaluate the re-

sults, which comprise the manual segmentation of

the optic disc. Another parameter for the quality of

the trained algorithm evaluated by Intersection-over-

Original image Ground truth

(a) An example of DRIONS-DB dataset.

Original image Ground truth

(b) An example of RIM-ONE v.3 dataset

Figure 4: Some examples from RIM-ONE v.3 and

DRIONS-DB datasets.

Deep Learning Residual-like Convolutional Neural Networks for Optic Disc Segmentation in Medical Retinal Images

Figure 5: Proposed network structure.

(a)

(b)

Figure 6: (a) Examples from RIM-ONE v.3 for OD segmentation by the proposed method. The green contour refers to the

ground truth, and blue is prediction, (b) Examples from RIM-ONE v.3 for OD segmentation by the proposed method. The

green contour refers to the ground truth, and blue is prediction.

Union, as well as called the Jaccard Index, is one

of the well-known metrics in image segmentation.

Intersection-over-Union(IOU) is the region of overlap

between the ground truth and predicted segmentation,

and deﬁned as:

IOU =

|A ∩ B|

|A ∪ B|

(3)

3 EXPERIMENTS AND RESULTS

The Dice coefﬁcient and IOU score do not depend on

the object scale and image scale.

IOU is a score that used in image segmentation.

As shown in Figure 3 IOU score is necessary to eval-

uate the percentage of overlap between ground-truth

and predicted segmentation. IOU score is similar to

the Dice coefﬁcient that often used for loss function

DeLTA 2020 - 1st International Conference on Deep Learning Theory and Applications

Table 1: Comparison of the proposed method with existing methods(on DRIONS-DB).

DRIONS-DB Dataset

Methods Dice IOU predict time (s)

Proposed method 0.9452 0.8853 0.008

(Walter et al., 2002) 0.6813 - -

(Morales et al., 2013) 0.9084 - -

(Abdullah et al., 2016) 0.9102 0.851 43.2

(Rehman et al., 2019) 0.8990 0.8210 31.10

(Zahoor and Fraz, 2017) - 0.8862 1.60

(Fan et al., 2017) 0.9137 0.8473 -

(Ramani and Shan-

thamalar, 2020)

0.8962 0.8217 1.41

DRIU (Maninis et al.,

2016)

0.94 0.89 0.1

(Sevastopolsky, 2017) 0.97 0.88 0.13

Table 2: Comparison of the proposed method with existing methods(on RIM-ONE v.3).

RIM-ONE v.3 Dataset

Methods Dice IOU predict time (s)

Proposed method 0.9371 0.87 0.008

(Zilly et al., 2017) 0.94 0.98 5.3

(Maninis et al., 2016) 0.96 0.89 0.13

(Joshua et al., 2019) 0.96 0.88 0.03

(Sevastopolsky, 2017) 0.95 0.89 0.1

(Civit Masot et al.,

2019)

0.97 - -

in the training network. IOU ranges from 0-1, which

1 (100) indicate fully overlapping segmentation. For

training this algorithm, we use free GPU service of

the Google Colab framework.

4 CONCLUSIONS

In this paper, we prove that our algorithm based on

Residual-like CNN can detect OD in shorter time,

better than other reported methods on retinal fun-

dus images. The best preponderance of the proposed

method, simple programming, accurate, and the low-

est prediction time, which is 0.008 per second. The

results by IOU and DICE scores were evaluated and,

great performances for optic disc segmentation were

achieved. The lowest prediction time and experiment

results express that optic disc segmentation can be

done automatically in ophthalmology clinics as on-

line predictions medical assistive tool.

ACKNOWLEDGEMENTS

The authors thanks Mr.Hooman Misaghi for his sup-

port and helps.

REFERENCES

Abbas, Q. (2017). Glaucoma-deep: detection of glaucoma

eye disease on retinal fundus images using deep learn-

ing. Int J Adv Comput Sci Appl, 8(6):41–5.

Abdullah, M., Fraz, M. M., and Barman, S. A. (2016). Lo-

calization and segmentation of optic disc in retinal im-

Deep Learning Residual-like Convolutional Neural Networks for Optic Disc Segmentation in Medical Retinal Images

ages using circular hough transform and grow-cut al-

gorithm. PeerJ, 4:e2003.

Al-Bander, B., Williams, B. M., Al-Nuaimy, W., Al-Taee,

M. A., Pratt, H., and Zheng, Y. (2018). Dense fully

convolutional segmentation of the optic disc and cup

in colour fundus for glaucoma diagnosis. Symmetry,

10(4):87.

Aquino, A., Geg

undez-Arias, M. E., and Mar

ın, D. (2010).

Detecting the optic disc boundary in digital fundus im-

ages using morphological, edge detection, and feature

extraction techniques. IEEE transactions on medical

imaging, 29(11):1860–1869.

Bajwa, M. N., Malik, M. I., Siddiqui, S. A., Dengel, A.,

Shafait, F., Neumeier, W., and Ahmed, S. (2019).

Two-stage framework for optic disc localization and

glaucoma classiﬁcation in retinal fundus images using

deep learning. BMC medical informatics and decision

making, 19(1):136.

Bourne, R. R., Flaxman, S. R., Braithwaite, T., Cicinelli,

M. V., Das, A., Jonas, J. B., Keeffe, J., Kempen, J. H.,

Leasher, J., Limburg, H., et al. (2017). Magnitude,

temporal trends, and projections of the global preva-

lence of blindness and distance and near vision im-

pairment: a systematic review and meta-analysis. The

Lancet Global Health, 5(9):e888–e897.

Carmona, E. J., Rinc

on, M., Garc

ıa-Feijo

o, J., and

Mart

ınez-de-la Casa, J. M. (2008). Identiﬁcation of

the optic nerve head with genetic algorithms. Artiﬁ-

cial Intelligence in Medicine, 43(3):243–259.

Chen, X., Xu, Y., Yan, S., Wong, D. W. K., Wong, T. Y.,

and Liu, J. (2015). Automatic feature learning for

glaucoma detection based on deep learning. In In-

ternational Conference on Medical Image Computing

and Computer-Assisted Intervention, pages 669–677.

Springer.

Civit Masot, J., Luna Perej

on, F., Dur

an L

opez, L.,

Dom

ınguez Morales, J. P., Vicente D

ıaz, S.,

Linares Barranco, A., and Civit Balcells, A. (2019).

Multi-dataset training for medical image segmentation

as a service. In UCCI 2019: 11th International Joint

Conference on Computational Intelligence (2019), p

542-547. ScitePress Digital Library.

Dice, L. R. (1945). Measures of the amount of ecologic

association between species. Ecology, 26(3):297–302.

Fan, Z., Rong, Y., Cai, X., Lu, J., Li, W., Lin, H., and Chen,

X. (2017). Optic disk detection in fundus image based

on structured learning. IEEE journal of biomedical

and health informatics, 22(1):224–234.

Federation, I. (2013). International diabetic federation atlas;

chapter 2 the global burden.

Forsyth, D. A. and Ponce, J. (2002). Computer vision: a

modern approach. Prentice Hall Professional Techni-

cal Reference.

Fraz, M. M., Welikala, R., Rudnicka, A. R., Owen, C. G.,

Strachan, D., and Barman, S. A. (2015). Quartz:

Quantitative analysis of retinal vessel topology and

size–an automated system for quantiﬁcation of reti-

nal vessels morphology. Expert Systems with Appli-

cations, 42(20):7221–7234.

Fu, H., Cheng, J., Xu, Y., Wong, D. W. K., Liu, J., and Cao,

X. (2018). Joint optic disc and cup segmentation based

on multi-label deep network and polar transformation.

IEEE transactions on medical imaging, 37(7):1597–

1605.

Fumero, F., Alay

on, S., Sanchez, J. L., Sigut, J., and

Gonzalez-Hernandez, M. (2011). Rim-one: An open

retinal image database for optic nerve evaluation.

In 2011 24th international symposium on computer-

based medical systems (CBMS), pages 1–6. IEEE.

Gao, Y., Yu, X., Wu, C., Zhou, W., Wang, X., and Chu, H.

(2019). Accurate and efﬁcient segmentation of optic

disc and optic cup in retinal images integrating multi-

view information. IEEE Access, 7:148183–148197.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional net-

works. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 4700–

4708.

Jackson, A. and Radhakrishnan, S. (2014). Understanding

and living with glaucoma. Glaucoma Research Foun-

dation.

Joshua, A. O., Nelwamondo, F. V., and Mabuza-Hocquet,

G. (2019). Segmentation of optic cup and disc

for diagnosis of glaucoma on retinal fundus im-

ages. In 2019 Southern African Universities Power

Engineering Conference/Robotics and Mechatron-

ics/Pattern Recognition Association of South Africa

(SAUPEC/RobMech/PRASA), pages 183–187. IEEE.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 3431–3440.

Maninis, K.-K., Pont-Tuset, J., Arbel

aez, P., and Van Gool,

L. (2016). Deep retinal image understanding. In In-

ternational conference on medical image computing

and computer-assisted intervention, pages 140–148.

Springer.

Mookiah, M. R. K., Acharya, U. R., Lim, C. M., Petznick,

A., and Suri, J. S. (2012). Data mining technique

for automated diagnosis of glaucoma using higher or-

der spectra and wavelet energy features. Knowledge-

Based Systems, 33:73–82.

Morales, S., Naranjo, V., Angulo, J., and Alca

niz, M.

(2013). Automatic detection of optic disc based on

pca and mathematical morphology. IEEE transactions

on medical imaging, 32(4):786–796.

Quigley, H. A. and Broman, A. T. (2006). The number of

people with glaucoma worldwide in 2010 and 2020.

British journal of ophthalmology, 90(3):262–267.

Ramani, R. G. and Shanthamalar, J. J. (2020). Improved im-

age processing techniques for optic disc segmentation

in retinal fundus images. Biomedical Signal Process-

ing and Control, 58:101832.

Rehman, Z. U., Naqvi, S. S., Khan, T. M., Arsalan, M.,

Khan, M. A., and Khalil, M. (2019). Multi-parametric

DeLTA 2020 - 1st International Conference on Deep Learning Theory and Applications

optic disc segmentation using superpixel based fea-

ture classiﬁcation. Expert Systems with Applications,

120:461–473.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Sevastopolsky, A. (2017). Optic disc and cup segmentation

methods for glaucoma detection with modiﬁcation of

u-net convolutional neural network. Pattern Recogni-

tion and Image Analysis, 27(3):618–624.

Sevastopolsky, A., Drapak, S., Kiselev, K., Snyder, B. M.,

Keenan, J. D., and Georgievskaya, A. (2018). Stack-

u-net: Reﬁnement network for image segmentation on

the example of optic disc and cup. arXiv preprint

arXiv:1804.11294.

Szeliski, R. (2010). Computer vision: algorithms and ap-

plications. Springer Science & Business Media.

Tieleman, T. and Hinton, G. (2012). Lecture 6.5-rmsprop:

Divide the gradient by a running average of its recent

magnitude. COURSERA: Neural networks for ma-

chine learning, 4(2):26–31.

Walter, T., Klein, J.-C., Massin, P., and Erginay, A. (2002).

A contribution of image processing to the diagnosis

of diabetic retinopathy-detection of exudates in color

fundus images of the human retina. IEEE transactions

on medical imaging, 21(10):1236–1243.

Welfer, D., Scharcanski, J., Kitamura, C. M., Dal Pizzol,

M. M., Ludwig, L. W., and Marinho, D. R. (2010).

Segmentation of the optic disk in color eye fundus im-

ages using an adaptive morphological approach. Com-

puters in Biology and Medicine, 40(2):124–137.

Welfer, D., Scharcanski, J., and Marinho, D. R. (2013). A

morphologic two-stage approach for automated optic

disk detection in color eye fundus images. Pattern

Recognition Letters, 34(5):476–485.

Wong, D., Liu, J., Lim, J., Jia, X., Yin, F., Li, H., and Wong,

T. (2008). Level-set based automatic cup-to-disc ratio

determination using retinal fundus images in argali.

In 2008 30th Annual International Conference of the

IEEE Engineering in Medicine and Biology Society,

pages 2266–2269. IEEE.

Xu, Y., Liu, J., Lin, S., Xu, D., Cheung, C. Y., Aung, T.,

and Wong, T. Y. (2012). Efﬁcient optic cup detec-

tion from intra-image learning with retinal structure

priors. In International Conference on Medical Im-

age Computing and Computer-Assisted Intervention,

pages 58–65. Springer.

Yu, H., Barriga, E. S., Agurto, C., Echegaray, S., Pattichis,

M. S., Bauman, W., and Soliz, P. (2012). Fast lo-

calization and segmentation of optic disk in retinal

images using directional matched ﬁltering and level

sets. IEEE Transactions on information technology in

biomedicine, 16(4):644–657.

Zahoor, M. N. and Fraz, M. M. (2017). Fast optic disc seg-

mentation in retina using polar transform. IEEE Ac-

cess, 5:12293–12300.

Zhang, Y., Matuszewski, B. J., Shark, L.-K., and Moore,

C. J. (2008). Medical image segmentation using new

hybrid level-set method. In 2008 ﬁfth international

conference biomedical visualization: information vi-

sualization in medical and biomedical informatics,

pages 71–76. IEEE.

Zilly, J., Buhmann, J. M., and Mahapatra, D. (2017). Glau-

coma detection using entropy sampling and ensemble

learning for automatic optic cup and disc segmenta-

tion. Computerized Medical Imaging and Graphics,

55:28–41.

Deep Learning Residual-like Convolutional Neural Networks for Optic Disc Segmentation in Medical Retinal Images