You are on page 1of 5

Diagnosis of Eye Retinal Diseases Based on

Convolutional Neural Networks Using Optical


Coherence Images
Mehmet Emre Sertkaya1, Burhan Ergen1, Mesut Togacar2
1
Computer Engineering, Faculty of Engineering, Firat University,
Elazig, Turkey
2
Computer Technology Program, Technical Vocational High School, Firat University,
Elazig, Turkey
181129208@firat.edu.tr

Abstract—In this study, the diagnosis of some diseases in the aging process and may develop due to another disease [11].
retina of the eye by using deep learning architectures is The Drusen structure needs to be carefully examined to
intended to be diagnosed. Optical Coherence Tomography determine if it is related to another disease.
device from Choroidal Neovascularization, Diabetic Macular
Choroidal neovascularization (CNV) is one of the leading
Edema, Drusen and healthy eye retinal images were examined.
LeNet, AlexNet and Vgg16 architectures of deep learning were causes of blindness in developed countries, especially age-
used. In each architecture, the hyper parameters were changed related macular degeneration (AMD) [12], [13].
to diagnose these diseases. Results of the implementation The Bruch's membrane (BM), together with the retinal
showed that exhibit successful results in Vgg16 and AlexNet pigment epithelium, sees a functional and physical barrier
architecture. Dropout layer structure in AlexNet has been that separates the vascular choroid from the avascular
shown to reduce the loss by minimizing loss.
external retina [14]. CNV retinal edema caused by
Index Terms—Biomedical optical imaging; Convolutional
disruption of this barrier causes blindness in the future if the
neural networks; Medical diagnosis; Supervised learning. loss of vision due to bleeding, lower retinal fluid and
bleeding is not treated [15].
I. INTRODUCTION In the last five years, the use of deep learning techniques
in the field of computer vision has started to increase. Deep
Optical Coherence Tomography (OCT) has become a key
learning methods, especially convolutional neural networks
diagnostic imaging technique for the diagnosis of retinal
(CNNs), have recently led to breakthroughs in many
diseases [1]. The ability to visualize the internal structure of
computer vision tasks, including object detection and
retina provides a qualitative and quantitative assessment of
recognition and image segmentation and captioning [16].
morphological changes associated with underlying diseases
Choroidal tissue segmentation using CNN demonstrated the
[2]. Measurements derived from OCT are considered to be
importance of analysing OCT images on multiple scales
very important markers in evaluating treatment response and
[17].
disease progression in clinical practice and in clinical trials
In this study, the study was performed to determine the
[3], [4]. In particular, retinal thickness or central macular
retinal diseases of the eye using CNN models. In this
thickness (CMT) measured in OCT has been shown to be
direction, Deep Learning architecture is described in the
associated with pathological changes and treatment
second chapter and convolutional neural network models are
outcomes for various ocular diseases [5]–[7].
used in the third part. In the fourth chapter, the data used
Many eye diseases and brain, cardiovascular system and
and the fifth part of the studies are mentioned experimental
some other systemic diseases manifest itself in the retina.
results.
Examples include diabetes mellitus, and the most feared
complication is diabetic retinopathy, age-related macular
II. DEEP LEARNING
degeneration, glaucoma, and part of all cardiovascular
diseases [8]. Deep learning is a type of machine learning that is
Diabetes Mellitus (DM), a group of metabolic diseases connected to artificial neural networks based on the
characterized by hyperglycemia caused by insulin structure and functions of a human brain [18]. In the 1980s,
dysfunction, causes eye retinal disorder called Diabetic with the emergence of parallel distributed architecture, the
Macular Edema (DME) [9]. The eye is the most frequently foundations of deep learning were established depending on
affected organ in patients with DM. DR is now the leading the artificial neural networks [19]. So deep concept of
cause of blindness in all age groups in countries [10]. learning neural networks that arise from time to time to
Drusen is the formation of abnormal extracellular matter research the “new generation neural networks” is also
accumulation under the retina pigment epithelium (RPE). defined.
The occurrence of Drusen disease may develop due to the The deeper concept of deep learning refers to layers that
are interconnected. These layers aim to find high-level
Manuscript received 7 January, 2019; accepted 30 March, 2019. representations as they move toward each layer. The deep

978-1-7281-2209-0/19/$31.00 ©2019 IEEE


learning structure is intended to make inferences by using neurons in the network and prevents memorization.
numerical data, images, videos, texts or even audio data by
providing a human-like and decision-making structure. a)
b)
Therefore, this diversity of deep learning is used in
particular for classification, recognition, detection [20]. In
general, the studies have been shaped in this direction and X X

now it has started to occupy a big place in our daily life.


Deep learning structure is formed in more than one layer
as a structure, and it only fulfils its task. Through these X X

tasks, the input data that enters the network is extracted, and
the extracted attributes are sent to the next layer to extract
X X

more detailed attributes. Deep learning as the basic structure


Fig. 1. Standard neural net (a); after applying dropout (b).
represents the advanced structure of convolutional artificial
neural network [21].
III. METHOD
A. Convolution Layer LeNet is the first network model in which convective
The convolutional layer, which is the basic layer of deep neural networks exhibit successful results. LeNet was
learning, is the layer where the attributes are extracted. This developed in 1998 by Yann LeCun. Initially designed to
layer makes the images from the previous layer convolution classify handwritten digits 8-layer structure was then being
with the filter matrix and, depending on these results, the used in other areas [23].
output image to be transmitted to the next layer is generated. AlexNet is a convolutional neural network model
The filter matrix maintained by this convolution is 2 × 2, 3 × developed by Alex Krizhevsky and his team in 2012,
3, 5 × 5, 7 × 7, 11 × 11 and so on size. Depending on the Compute Unified Device Architecture (CUDA) supported
size of the image, RGB is subjected to a 3D filter and by Graphics Processing Unit (GPU) in ImageNET Large
grayscale to a one-dimensional filter matrix. ƒ main data Scale Visual Recognition Competition (ILSRVRC)
(image), the convolution operation including ℎ filter shown competition. This network entered the top five in the
in (1) competition with a 15.3 % error rate. AlexNet was designed
f • h  f ( k, l ) h ( i − k, j − l ).
= (1) by the SuperVision group consisting of Alex Krizhevsky,
kl Geoffrey Hinton and Ilya Sutskever [24]. Alexnet optionally
B. Rectified Linear Units(ReLu) Layers includes Dropout layers to prevent possible memorization.
Therefore, the Dropout layer is used after the last
This layer normalizes the values in the incoming image
convolution layer is used after the last convolution layer in
after the convolution layers. That is, if the values in the
deep learning architectures.
incoming image are less than zero, normalizes to zero.
VGG16 was designed by Visual Geometry Group in
C. Pooling Layer 2014, consisting of Karen Simonyan and Andrew Zisserman
The pooling layer, also called subsampling, is usually [25]. AlexNET is similar to the network structure and the
placed after the ReLu layer. This layer reduces the size of weight configuration is public. This network has achieved
the network by reducing the number of parameters and success rate of up to 7.5 %. However, VGGNET consists of
calculations. The pooling layer is usually between two 138 million parameters.
convolution layers or between the ReLu layers. This layer
mostly uses two methods: maximum pooling and average IV. DATASET
pooling. In a specific entry window, the maximum pooling The dataset used is the retinal images composed of OCT
takes maximum values within the window, whereas the images and formed as a result of the work of institutions for
average pooling takes the mean value. many years [26]–[28]. This image dataset consists of four
D. Full Connected Layer classes and this classes CNV, Drusen DME and Normal
images.
Converts the pre-defined multidimensional matrices into a
one-dimensional matrix. Neurons are fully attached to this TABLE I. DATASET TRAIN, VALIDATION AND TEST IMAGE
layer, each neuron binds to the previous neuron, so it is RATIOS.
known as a fully connected layer [22]. Sample
Dataset
Number NORMAL CNV DME DRUSEN
Category
E. Dropout Layer (%)
Train 80 21,581 30,837 8,859 6,537
Sometimes the network can be memorizing the other by Validation 15 4,000 4,000 4,000 4,000
over-memorizing some reasons. Dropout layer is used to Test 5 1,000 1,000 1,000 1,000
eliminate this. The dropout layer randomly removes the Total 100 26,581 35,837 13,859 11,537

Fig. 2. Formations of CNV, DME and DRUSEN Diseases on Retina.


Table IV and Fig. 5 also examined the dropout layer in
V. EXPERIMENTAL RESULTS AlexNet (Dropout) has prevented the loss of memorization,
In this study, it is aimed to diagnose eye retinal diseases but did not show much effect on performance. This shows
using LeNet, AlexNet (Dropout Layer), Alexnet and Vgg16 that the models have not gone by memorization, but it can
using this dataset. In this direction, each model was run in be observed that the Dropout layer exhibits better results in
200 iterations and successful results were obtained. Table II layers that do not use this layer in minimizing loss.
and Fig. 3 show the results of the training success of these TABLE IV. RESULTS OF TRAINING LOSS.
models. In the study, Vgg16 resulted in 94.16 percent and Last Average Lowest
Method Epochs
AlexNet 94.28 percent performance. Loss Loss Loss
LeNet 200 0.2282 0.4384 0.1925
TABLE II. RESULTS OF TRAINING ACCURACY. AlexNet 200 0.2080 0.3554 0.1452
Method Epoch Last Average Best VGG16 200 0.1715 0.4087 0.1423
Accuracy Accuracy Accuracy
LeNet 200 91.05 84.35 93.72
AlexNet 200 92.12 87.23 94.28
VGG16 200 90.01 84.62 94.16 1,5
1,4
1,3 LeNet
1,2 Vgg16
1,1 AlexNet(Dropout)

Loss
1 AlexNet
1 0,9
0,9 0,8
0,8 0,7
Accuracy

0,7 0,6
0,6 0,5
0,4
0,5 0,3
0,4 0,2
0,3 0,1
0,2 LeNet Vgg16 0
0,1 AlexNet(Dropout) AlexNet 1 26 51 76 101 126 151 176 201
0 Fig. 5. LeNet, AlexNet, AlexNet (Dropout Layer) and Vgg16 Train Loss
Graphics. Epochs
1 26 51 76 101 126 151 176 201
Epochs VI. CONCLUSIONS
Fig. 3. LeNet, AlexNet, AlexNet (Dropout Layer) and Vgg16 Train
Accuracy Graphics. In this study, using optical coherence images, eye retinal
diseases were diagnosed using convolutional neural
In parallel with the train accuracies, verification results networks. The proposed method was used to classify the
showed high results in these models as shown in Table III existing dataset using the AlexNet, LeNet, VGG-16
and Fig. 4. AlexNet achieved the highest verification score architectures. In all processing steps, the AlexNet
with percent 94.02 accuracy. architecture yielded a good classification result of 94.28 %.
TABLE III. RESULTS OF VALIDATION ACCURACY. In the future studies, by using deep learning methods, it is
Last Average Best aimed to determine the deformed region by removing the
Method Epoch
Accuracy Accuracy Accuracy heat map.
LeNet 200 82.66 78.72 83.76
AlexNet 200 93.61 86.58 94.02
VGG16 200 92.58 84.31 93.01

1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
LeNet Vgg16
Accuracy

0,2
0,1 AlexNet(Dropout) AlexN
0 et
1 26 51 76 101 126 151 176 201
Epochs

Fig. 4. LeNet, AlexNet, AlexNet (Dropout Layer) and Vgg16 Validation


Accuracy Graphics.
10.1016/j.ajo.2003.09.042.
VII.REFERENCES [14] K. Nguyen, C. Fookes, A. Ross, S. Sridharan, “Iris recognition with
[1] S. M. Waldstein et al., “Correlation of 3-dimensionally quantified off-the-shelf CNN features: a deep learning perspective”, IEEE
intraretinal and subretinal fluid with visual acuity in neovascular age- Access, vol. 6, pp. 18848–18855, 2017. DOI:
related macular degeneration”, JAMA Ophthalmol, vol. 134, no. 2, 10.1109/ACCESS.2017.2784352.
pp. 182–190, 2016. DOI: 10.1001/jamaophthalmol.2015.4948. [15] X. Sui et al., “Choroid segmentation from Optical Coherence
[2] S. Sharma et al., “Macular morphology and visual acuity in the Tomography with graph-edge weights learned from deep
second year of the comparison of age-related macular degeneration convolutional neural networks”, Neurocomputing, vol. 237, pp. 332–
treatments trials”, Ophthalmology, vol. 123, no. 4, pp. 865–875, 2016. 341, 2017. DOI: 10.1016/j.neucom.2017.01.023.
DOI: 10.1016/j.ophtha.2015.12.002. [16] T. S. Lee, D. Mumford, “Hierarchical Bayesian inference in the visual
[3] A. Wood et al., “Retinal and choroidal thickness in early age-related cortex”, J Opt Soc Am A Opt Image Sci Vis, vol. 20, no. 7, pp. 1434–
macular degeneration”, Am. J. Ophthalmol., vol. 152, no. 6, pp. 1030– 1448, 2003.
1038, 2011. DOI: 10.1016/j.ajo.2011.05.021. [17] J. L. McClelland, D. E. Rumelhart, Computational models of
[4] M. Fleckenstein et al., “Tracking progression with spectral-domain cognition and perception. Explorations in parallel distributed
optical coherence tomography in geographic atrophy caused by age- processing: A handbook of models, programs, and exercises.
related macular degeneration”, Invest. Ophthalmol. Vis. Sci., vol. 51, Cambridge, MA, US: The MIT Press, 1988.
no. 8, pp. 3846–3852, 2010. DOI: 10.1167/iovs.09-4533. [18] L. Deng, “Deep Learning: methods and applications”, Foundations
[5] P. A. Keane et al., “Evaluation of optical coherence tomography and Trends® in Signal Processing, vol. 7, no. 3–4, pp. 197–387,
retinal thickness parameters for use in clinical trials for neovascular 2014. DOI: 10.1561/2000000039.
age-related macular degeneration”, Invest. Ophthalmol. Vis. Sci., [19] F. Dogan, I. Turkoglu, “Derin Ogrenme Algoritmalarının Yaprak
vol. 50, no. 7, pp. 3378–3385, 2009. DOI: 10.1167/iovs.08-2728. Sınıflandırma Başarımlarının Karşılaştırılması”, Sakarya University
[6] M. Sonka, M. D. Abramoff, “Quantitative analysis of retinal OCT”, Journal of Computer and Information Sciences, vol. 1, no. 1, pp. 10–
Medical Image Analysis, vol. 33, pp. 165–169, 2016. DOI: 21, 2018. (in Turkish)
10.1016/j.media.2016.06.001. [20] A. Adler, M. Elad, M. Zibulevsky, “Compressed learning: a deep
[7] American Diabetes Association, “Diagnosis and classification of neural network approach”, Computer Vision and Pattern Recognition,
diabetes mellitus”, Diabetes Care, vol. 34, pp. S62–S69, 2011. DOI: 2016.
10.2337/dc10-S062. [21] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, “Gradient-based learning
[8] “Blind and partial sight registration in Avon. - PubMed - NCBI.” applied to document recognition”, Proceedings of the IEEE, vol. 86,
[Online]. Available: https://www.ncbi.nlm.nih.gov/pubmed/2930763 no. 11, pp. 2278–2324, 1998. DOI: 10.1109/5.726791.
[9] D. Ardeljan, C.-C. Chan, “Aging is not a disease: distinguishing age- [22] A. Krizhevsky, I. Sutskever, G. E. Hinton, “ImageNet classification
related macular degeneration from aging”, Prog Retin Eye Res, with deep convolutional neural networks”, in Proc. 25th Int. Conf.
vol. 37, pp. 68–89, 2013. DOI: 10.1016/j.preteyeres.2013.07.003. Neural Information Processing Systems, New York, NY, USA, 2012,
[10] N. Congdon et al., “Causes and prevalence of visual impairment pp. 1097–1105. DOI: 10.1145/3065386.
among adults in the United States”, Arch. Ophthalmol., vol. 122, [23] K. Simonyan, A. Zisserman, “Very deep convolutional networks for
no. 4, pp. 477–485, 2004. DOI: 10.1001/archopht.122.4.477. large-scale image recognition”, Computer Vision and Pattern
[11] K. Michalska-Malecka, A. Kabiesz, M. Nowak, D. Spiewak, “Age Recognition, 2014.
related macular degeneration – challenge for future: Pathogenesis and [24] D. S. Kermany et al., “Identifying medical diagnoses and treatable
new perspectives for the treatment”, European Geriatric Medicine, diseases by image-based deep learning”, Cell, vol. 172, no. 5,
vol. 6, no. 1, pp. 69–75, 2015. DOI: 10.1016/j.eurger.2014.09.007. pp. 1122–1131, 2018. DOI: 10.1016/j.cell.2018.02.010.
[12] P. A. Campochiaro, “Molecular pathogenesis of retinal and choroidal [25] D. Kermany, K. Zhang, M. Goldbaum, “Labeled optical coherence
vascular diseases”, Prog Retin Eye Res, vol. 49, pp. 67–81, 2015. tomography (OCT) and chest x-ray images for classification”,
DOI: 10.1016/j.preteyeres.2015.06.002. Mendeley Data, v2, 2018. DOI: 10.17632/rscbjbr9sj.2.
[13] H. E. Grossniklaus, W. R. Green, “Choroidal neovascularization”, Retinal OCT Images (optical coherence tomography). [Online].
Am. J. Ophthalmol., vol. 137, no. 3, pp. 496–503, 2004. DOI: Available: https://www.kaggle.com/paultimothymooney/kermany2018
[26]

You might also like