Eye Abnormality Automatic Detection using Deep Learning based

Model

Audyati Gany

1,* a

, Meilan Jimmy Hasugian

1,2 b

, Erwani Merry Sartika

Novie Theresia Br. Pasaribu

and Hannah Georgina

Department of Electrical Engineering, Maranatha Christian University, Surya Sumantri 65, Bandung, Indonesia

Department of Electrical Engineering, Chung Yuan Christian University, Zhongli, Taiwan

novie.theresia@eng.maranatha.edu, 1822007@eng.maranatha.edu

Corresponding author

Keywords: Ocular Disease, Fundus Retina, Deep Learning, Convolutional Neural Network.

Abstract: Early detection and diagnosis of ocular pathologies would enable to forestall of visual impairment. One

challenge that limits the adoption of a computer-aided diagnosis tool by the ophthalmologist is, the sight-

threatening rare pathologies such as central retinal artery occlusion or anterior ischemic optic neuropathy and

others are usually ignored. The aim of this research is to develop methods for automatic detection of eye

abnormality caused by the most common ocular disease along with the rare pathologies. For this purpose, we

developed the deep learning-based model trained with Retinal Fundus Multi-disease Image Dataset (RFMiD).

This dataset consists of a 1920 fundus retina images captured using three different fundus cameras with 46

conditions annotated through adjudicated consensus of two senior retinal experts. The model is built on the

top of some prominent pretrained convolutional neural network (CNN) models. From the experiment, the

model could achieve the accuracy level and recall 0.87, whereas precision and F1 score are 0.86, and area

under receiver operating characteristic (AUROC) is 0.90. The proposed model built in deep learning structure

could be a promising model in automatic classification of ocular disease based on fundus retina images.

1 INTRODUCTION

In the World Report on Vision 2019, (WHO, 2019)

stated that approximately 2.2 billion people

worldwide have visually impaired, of whom at least 1

billion have a vision impairment that could have been

prevented. The world faces considerable challenges

in terms of eye care, including inequalities in the

coverage and quality of prevention, treatment, and

rehabilitation services. Early detection and diagnosis

of ocular pathologies would enable to forestall of

visual impairment. One challenge that limits the

adoption of a computer-aided diagnosis tool by the

ophthalmologist is, the sight-threatening rare

pathologies such as central retinal artery occlusion or

https://orcid.org/0000-0002-7389-6667

https://orcid.org/0000-0003-0759-6663

https://orcid.org/0000-0003-3720-3584

https://orcid.org/0000-0001-7774-9675

https://orcid.org/0000-0002-5885-3503

anterior ischemic optic neuropathy and others are

usually ignored.

In the past two decades, many publicly available

datasets of color fundus images have been collected

with a primary focus on diabetic retinopathy,

glaucoma, and age-related macular degeneration, and

few other frequent pathologies. There are several

researches that use the fundus retina images as the

intake in their system (Qummar, et al., 2019)

(Soomro, et al., 2019) (Sarki, Ahmed, Wang, &

Zhang, 2020). However, most of them applied for

diabetic retinopathy disease.

This study is a preliminary research that aims to

develop methods in automatic detection of eyes

abnormality caused by not only a frequent ocular

disease but also along with the rare pathologies. For

370

Gany, A., Hasugian, M., Sartika, E., Pasaribu, N. and Georgina, H.

Eye Abnormality Automatic Detection using Deep Learning based Model.

DOI: 10.5220/0010754400003113

In Proceedings of the 1st International Conference on Emerging Issues in Technology, Engineering and Science (ICE-TES 2021), pages 370-375

ISBN: 978-989-758-601-9

this purpose, we developed the deep learning-based

model trained with Retinal Fundus Multi-disease

Image Dataset (RFMiD).

Therefore, the contribution of this research is its

capability in auto detection of eye abnormality for not

only most common ocular disease such as diabetic

retinopathy, but also for other rare ocular diseases.

2 METHODS AND MATERIALS

2.1 Fundus Retina Datasets

In this research we deployed RFMiD dataset as

provided in Retina Image Analysis Challenge

(Pachade, et al., 2021). These datasets just published

in April 2021. Total images provided during our

experiment are 1920 fundus images. These images

were captured using three different fundus cameras

with 46 conditions annotated through adjudicated

consensus of two senior retinal experts. The

advantage of this dataset is its wide variety of diseases

that appear in clinical settings.

Several diseases/abnormalities that are included

in the dataset: diabetic retinopathy (DR), age-related

macular degeneration (ARMD), media haze (MH),

drusen (DN), myopia (MYA), branch retinal vein

occlusion (BRVO), tessellation (TSLN), epiretinal

membrane (ERM), laser scar (LS), macular scar

(MS), central serous retinopathy (CSR), optic disc

cupping (ODC), and many more total 42 types of eye

abnormalities. Visualization of these disease is shown

Figure 1.

Figure 1: Fundus retina images for several eye diseases

The data distribution between the normal eyes and

eyes with abnormalities is presented in the diagram

below.

Figure 2: Distribution between normal and abnormal eyes.

Around 20.9% (401 images) are normal eyes and

the rest of it are from eyes with several diseases.

2.2 Method

The usage of deep learning in computer vision

problem especially in biomedical imaging analysis

become more compelling (Liu, et al., 2018),

(Tajbakhsh, Shin, Gurudu, & Hurst, 2016) (Roy, et

al., 2020). In this paper, we report a deep-network

based model to develop a system for detecting eye

abnormality automatically. The model was built on

the top of some prominent pretrained convolutional

neural network (CNN) e.g., VGG-16 and Inception-

v3. These networks were trained by using ImageNet

dataset of over 14 million images belonging to 1000

classes. Therefore, it’s very suitable for common

computer vision problem. However, for fundus retina

images, we deployed the network in a transfer

learning fashion and without or with image-

augmentation to improve the generalization of the

model in detecting eyes abnormality.

The basic structure of VGG16 model (Simonyan

& Zisserman, 2015) is depicted in Figure 3. This

model consists of 16 layers, with 13 of convolutional

layers and 3 fully connected (FC) layers.

Eye Abnormality Automatic Detection using Deep Learning based Model

371

Figure 3: The basic structure of VGG16 model.

This fully connected layers parts of the model was

adjusted for the purpose of this research. The

complete configuration of the model, without and

with image-augmentation process is reported in Table

1 and Table 2 respectively.

Table 1: The configuration of the fully-connected layers in

VGG16, without image-augmentation.

Layers Model A1 Model A2

FC 1 1024, ReLU 1024, ReLU

Dro

out n.a. 0.5

FC 2 512, ReLU 512, ReLU

Drop out 0.5 0.5

Output 1, sigmoi

1, sigmoi

We deployed Rectifier Linear Unit (ReLU) as the

activation function in FC1 and FC2, and sigmoid

activation function at the output, for binary

classification system.

Table 2: The configuration of the fully-connected layers in

VGG16, with image-augmentation.

Layers Model

Model

FC 1 512,

ReLU

512,

ReLU

512,

ReLU

512,

ReLU

Drop

out

0.3 n.a. 0.3 n.a.

FC 2 512,

ReLU

512,

ReLU

512,

ReLU

512,

ReLU

Drop

out

n.a. 0.3 0.3 n.a.

Output 1,

sigmoi

To avoid overfitting during the training, drop-out

technique was used as the regularization for the deep

learning model.

The basic structure of Inception-v3 network

(Szegedy, Vanhoucke, Ioffe, Shlens, & Wojna, 2016)

is illustrated in Figure 4. This model contains 42

layers. The detail structure in Stem, Inception A,

Reduction A, Inception B, Reduction B, and

Inception C blocks are depicted in Figure 5-10. This

model employs the Inception network (Szegedy, et

al., 2015).

Figure 4: The basic structure of Inception-v3 model.

Figure 5: The Stem block in Inception-v3 model.

Figure 6: The Inception A block in Inception-v3 model.

ICE-TES 2021 - International Conference on Emerging Issues in Technology, Engineering, and Science

372

Figure 7: The Inception B block in Inception-v3 model.

Figure 8: The Inception C block in Inception-v3 model.

Figure 9: The Reduction A block in Inception-v3 model.

Figure 10: The Reduction B block in Inception-v3 model.

The fully connected layers parts of this model was

modified in order to be applicable in our purpose. The

complete configuration of the model, without and

with image-augmentation process is described in

Table 3 and Table 4, respectively.

Table 3: The configuration of the fully-connected layers in

Inception-v3, without image-augmentation.

ers Model C1

FC 1 4096, ReLU

Drop out 0.5

FC 2 512, ReLU

Drop out 0.3

Out

ut 1, si

moi

Table 4: The configuration of the fully-connected layers in

Inception-v3, with image-augmentation.

Layers Model

Model

FC 1 512,

ReLU

512,

ReLU

512,

ReLU

512,

ReLU

1024,

ReLU

Drop

out

n.a. n.a. 0.3 0.3. n.a.

FC 2 512,

ReLU

512,

ReLU

512,

ReLU

512,

ReLU

512,

ReLU

Drop

out

n.a. 0.3 n.a. 0.3 0.3

Output 1,

sigmoi

From the total 1920 fundus images, is split into

1305 for training images, 231 images for validation

data, and 384 for testing data. Stratified sampling

method is used for splitting the images, on order to

maintain the similar distribution for each class

(normal and abnormal eyes).

In order to improve the ability of the model to

preserve the generalization principles, we applied

data augmentation process. The parameter for this

augmentation is described in Table 5.

Table 5: Parameters for image-augmentation process.

Paramete

Value

rotation angle 30

width shift 0.2

height shift

range

0.2

shear range 0.1

horizontal fli

True

2.3 Evaluation Metrics

The performance of the model is measured in terms

of the ability of the model to correctly detect the eyes

with abnormality and the normal eyes as true positive

(TP) and true negative (TN) respectively. And

misjudge the normal eye to be abnormal as well as

abnormal eye classified as normal as false positive

(FP) and false negative (FN) respectively.

The metrics for evaluation are precision (Pre),

recall (Re), accuracy (Acc), and F1 score.

Pre = TP / (TP + FP) (1

)

Re = TP / (TP + FN) (2

)

Acc = (TP + TN) / (TP + TN + FP + FN) (3

)

F1-score = 2 * Pre * Re / (Pre + Re) (4

)

And another metrics that are prevalent in binary

classification is area under receiver operating curve

(AUROC) or sometimes called as AUC for

Eye Abnormality Automatic Detection using Deep Learning based Model

373

simplicity. That is the area under the curve between

false positive rate and true positive rate.

3 RESULTS AND DISCUSSION

The result for each model based on VGG16 structure

is reported in Table 6. It can be seen in the table that

model A1 outperformed all other models based on

VGG16 network in every evaluation metric.

Table 6: Model performance based on VGG16 network.

Model Acc Pre Re F1 AUC

A1 0.87 0.86 0.87 0.86 0.87

A2 0.85 0.84 0.85 0.84 0.84

B1 0.86 0.85 0.86 0.85 0.84

B2 0.84 0.83 0.84 0.82 0.83

B3 0.83 0.81 0.83 0.81 0.83

B4 0.84 0.83 0.84 0.83 0.84

The comparation of ROC curve for each model is

presented in Figure 11. Figure 11 shows that the AUC

of model A1 is more superior than the other models.

Figure 11: The comparation of ROC curve result for each

model based on VGG16 structure.

Table 7 presents all metrics for each model based

on Inception-v3 structure.

Table 7: Model performance based on Inception-v3

network.

Model Acc Pre Re F1 AUC

C1 0.85 0.84 0.85 0.84 0.89

D1 0.86 0.85 0.86 0.85 0.90

D2 0.85 0.85 0.85 0.85 0.90

D3 0.86 0.85 0.86 0.85 0.90

D4 0.86 0.85 0.86 0.85 0.89

D5 0.83 0.84 0.83 0.84 0.89

In Table 7, model D1 and D3 shows a same result.

Both surpass the other models in every metric.

The ROC curve comparation between all models

based on Inception-v3 is shown in Figure 12.

Figure 12: The Reduction B block in Inception-v3 model.

Figure 12 shows that the AUC of each model is

different not too significant. Therefore, it can be said,

the variation of fully connected layers do not affect

the results essentially.

VGG16 structure is very simple compared to

Inception-v3. The number of neurons in FC layers

layer should be able to accommodate the features as

the output of the convolutional layer. Therefore, the

sudden change in numbers will affect the

performance. This causes the model A1 outperforms

more than the other model. The image augmentation

process seems not to help very much to improve the

performance.

On the other hand, Inception-v3 structure is

already complex. The variation of neuron numbers in

FC layers would not affect significantly. As it is seen

in Figure 12, all models have quite similar ROC.

However, the image augmentation gives more impact

to the performance.

To the best of our knowledge, since these datasets

were launched in April 2021, we haven’t found yet

any reported publication that used the same dataset in

their model. Therefore, at this current moment we

couldn’t provide any comparison from others’ model

that are trained with the same dataset.

4 CONCLUSIONS

From the experiment it can be concluded that VGG16

based model is not sensitive in image augmentation

but to the fully connected layers. On the other hand,

Inception-v3 based model is more impacted by image

ICE-TES 2021 - International Conference on Emerging Issues in Technology, Engineering, and Science

374

augmentation. However, by analysing the ROC

curve, both structures are still promising in

development of eye abnormality detection in further

research.

REFERENCES

Liu, J., Pan, Y., Li, M., Chen, Z., Tang, L., Lu, C., & Wang,

J. (2018, Mar). Applications of deep learning to MRI

images: a survey. Big Data Mining and Analytics, 1(1).

Pachade, S., Porwal, P., Thulkar, D., Kokare, M.,

Deshmukh, G., Sahasrabuddhe, V., . . . Mériaudeau, F.

(2021). Retinal fundus multi-disease image dataset

(RFMiD): a dataset for multi-disease detection

research. Data, 6(2).

doi:https://doi.org/10.3390/data6020014

Qummar, S., Khan, F. G., Shah, S., Khan, A.,

Shamshirband, S., Rehman, Z. U., . . . Jadoon, W.

(2019). A deep learning ensemble approach for diabetic

retinopathy detection. IEEE Access, 7, 150530-150539.

Roy, S., Menapace, W., Oei, S., Luijten, B., Fini, E., Saltori,

C., . . . Demi, L. (2020, Agt). Deep learning for

classification and localization of COVID-19 markers in

point-of-care lung ultrasound. IEEE Trans. on Medical

Imaging, 39(8).

Sarki, R., Ahmed, K., Wang, H., & Zhang, Y. (2020).

Automatic detection of diabetic eye disease through

deep learning using fundus images: a survey. IEEE

Access, 151133 - 151149.

Simonyan, K., & Zisserman, A. (2015). Very deep

convolutional networks for large-scale image

recognition. International Conference on Learning

Representations (ICLR). San Diego.

Soomro, T. A., Afifi, A. J., Zheng, L., Soomro, S., Gao, J.,

Hellwich, O., & Paul, M. (2019). Deep learning models

for retinal blood vessels segmentation: a review. IEEE

Access, 71696 - 71717.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., . . . Rabinovich, A. (2015). Going deeper

with convolutions. IEEE Conference on Computer

Vision and Pattern Recognition (CVPR).

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna,

Z. (2016). Rethinking the inception architecture for

computer vision. IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), (pp. 2818-2826).

Tajbakhsh, N., Shin, J. Y., Gurudu, S. R., & Hurst, R. T.

(2016, May). Convolutional neural networks for

medical image analysis: full training or fine tuning.

IEEE Trans. on Medical Imaging, 35(5).

WHO. (2019). World report on vision. World Health

Organization.

Eye Abnormality Automatic Detection using Deep Learning based Model

375