2022 International Conference on Signal and Information Processing (IConSIP)

College of Engineering Pune, India. Aug 26-27, 2022

Ocular Disease Recognition using Deep Learning

Kuldeep Vayadande Varad Ingale Vivek Verma
Department of Artificial Intelligence and Department of Artificial Intelligence and Department of Artificial Intelligence and
Data Science Data Science Data Science
Vishwakarma Institute of Technology Vishwakarma Institute of Technology Vishwakarma Institute of Technology
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected]
2022 International Conference on Signal and Information Processing (IConSIP) | 978-1-7281-6885-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICONSIP49665.2022.10007470

Abhishek Yeole Sahil Zawar Zoya Jamadar

Department of Artificial Intelligence and Department of Artificial Intelligence and Department of Artificial Intelligence and
Data Science Data Science Data Science
Vishwakarma Institute of Technology Vishwakarma Institute of Technology Vishwakarma Institute of Technology
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected]

Abstract— Artificial intelligence holds a significant impact in a Cataracts can be aggravated by inherited genetic defects that
variety of drug-related medical studies, including ophthalmology. create health concerns. The term ocular is used to describe a
Deep literacy styles, in particular, have been successful in detect- tumor that is accompanied by an eye. It can be intraocular (af-
ing clinical signs and bracketing optical conditions. Studies reveal fecting the inside of the eye) or extraocular (affecting the out-
Ocular diseases to be the major contributing reason of childhood side of the eye). Cataracts, diabetic retinopathy, and redness
blindness all over the world. Rapid and automatic illness identifi- level are the most common diseases detected. [5]. Through
cation is vital and urgent in lowering the strain of ophthalmolo- studies including automatic recognition of diseases from iris
gists. Ophthalmologists use pattern recognition to identify disor- and fundus or retinal images, computer vision and deep learning
ders by looking at the eye and its surrounding tissues directly or
have assisted ocular pathology. [6]
indirectly. As a result, can benefit the area of medical greatly.
Each disease has several severity levels that can be identified by Elloumi et al. divide such research into two categories based
confirming the presence of different lesions. Morphological char- on the goal: pathology or severity identification and ocular le-
acteristics identify each lesion, and numerous lesions from differ- sion segmentation. Fundus based images including ocular dis-
ent diseases have similar characteristics. In ophthalmology, deep eases have been included in their analysis of DL-based tech-
literacy techniques have mostly been employed on eye fundus niques, they mention works including Fundus image databases
pictures and optic consonance tomography. In this paper , we that target specific disorders.
have used three models namely CNN, Inception V3 , VGG-19 for
cataract prediction. These properties have allowed deep learning to outperform
standard methods in a variety of computer vision and image
We have got accuracy of 0.9587 for VGG-19 which is perform- analysis tasks. Because of its success, it is now being used to
ing best as compared to other models. analyze medical images, including, of course, ophthalmology
images [8].
Keywords- Clinical Diagnosis, Deep Learning, Image Classifica-
tion, Neural Network, Ocular disease. The primary cause of blindness in the globe is fundus image
disorders. Eye problems, age-related eye problems (AMD),
I. INTRODUCTION glaucoma, and diabetes mellitus are among the most common
According to the World Health Organization (WHO), 2.2 eye illnesses (DR). According to projections, there would be
billion people worldwide are blind or have vision loss, with at 95.4 million people with cataract and 3.36 million of people
least 1 billion having impaired eyesight that might be correct- with myopia by 2030, up from 1.95 billion in 2010. [9] Fur-
ed.[1] Reports reveal that DED, a chronic eye illness can cause thermore, according to various studies, those who develop my-
irreversible vision impairments if left untreated and if ignored opia before the age of 20 are more prone to get cataracts later in
for longer time, and this is estimated to be detected in around life. Despite the fact that the specific cause is uncertain, some
one-third of diabetics.[2] studies suggest that the increased axial eyeball length may hin-
der nutrient transfer to the lenses' rear site. [10]
Fundus pictures can be used to acquire retinal images,
which aid doctors in evaluating lesions or changes in architec- We have focused on one of the ocular diseases that is Cataract.
ture. In such scenario, manual identification has been attempted Firstly, we extracted images from dataset and resized images. Then we
extracted information of whether eye is cataract affected or not. Then
but in longer way. Automation of these operations can save we trained Deep Learning models and compared accuracies of model
time and money.[3] The most serious tumors in the eye are iris to determine best performing model. We have discussed about this in
cataract tumors, commonly known as 'eye tumors.' Cataracts are Section IV Methodology.
caused by a change in the tissue that makes up the lens of the
eye as a result of age or damage. The lens' proteins and fibers
continue to break down, resulting in hazy or poor vision. [4].

978-1-7281-6885-2/22/$31.00 ©2022 IEEE 1

horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 II. LITERATURE REVIEW In [20], they employed a kernel-based VGG network model
In [14], mentioned various methods about how to classify in this paper to achieve better outcomes in deep learning net-
diabetic retinopathy. A recommended profound learning works, and the training model in this proposed VGG network
framework called Deep Convolutional Neural Network would not be merged but would be connected before being
(DCNN) involves spatial examination to give high exactness in transferred to the layers. The fundamental concept was to con-
infection distinguishing proof. A DCNN is a more convoluted nect all levels directly to ensure maximum information transfer
plan in view of human visual discernment. Our proposed engi- across the network.
neering, when joined with dropout layer draws near, accom- In [23], it proposes a profound brain network model to help
plishes a precision of 94-96 percent. location of beginning phases of diabetic retinopathy and glau-
In [25], Sarki looks at a range of automated methods for coma. It can notify people that they should see an ophthalmolo-
identifying diabetic eye condition. They have provided a com- gist for a screening. This system primarily serves as a referral
plete eye detection overview of methods along with cutting trigger, informing the patient that a retinal expert should be
edge field methodology, including cutting-edge field methodol- consulted if a positive result is detected. This newly designed
ogy, with the purpose of delivering vital information to research model is less difficult and has an accuracy of 80%.
communities, healthcare professionals, and diabetes patients. It In [21], the proposed method in this paper uses three sepa-
is classified as a black box since it is unknown. Several re- rate convolution neural network (CNN) models to analyze OCT
searchers have fine-tuned the restrictions of current deep learn- images of the retina to identify the various retinal layers, recov-
ing algorithms like CNN to improve classification efficiency. er critical information, detect any new abnormalities, and pre-
In [16], the ocular image is prepossessed using the HE dict many eye diseases. This research demonstrates how deep
method and the changed image is segmented using k-means learning techniques (CNN) may be utilized to correctly catego-
clustering. SVM and RF are useful for categorizing the typical rize and identify ocular illness traits such CNV, drusen, and
and unusual parts of the ocular picture, as well as reducing hu- DME in comparison to normal.
man error, which reduces false recognition and improves preci- III. METHODOLOGY
sion. When compared to the SVM, the identification rate of RF
is 96.62 percent, indicating a great and steady result. A. Data Collection:
The neural network-based ocular pathology signs and dis- The dataset contains images of left and right eyes of almost
ease identification approach is an innovative, practical, and 5000 patients. Information such as age, gender of each patient is
intelligent alternative to traditional methods for early detection also mentioned along with color fundus images.
of optical diseases. It is a symbol of using image processing for Quality control management was used to mark annotations
the greater good of humanity. It has a high degree of accuracy by trained human readers. Patients are categorized into eight
in detecting all sorts of ocular disorders. It is been implemented categories, including:
practically and user friendly to use. This illness detection meth-
od has been put to the test in real time on a variety of photo- 1. Normal (N),
graphs with various forms of ocular disorders, and thus bears 2. Diabetes (D),
the stamp of practical capability. Medical research and oph-
thalmology will benefit greatly from this study. [17] 3. Glaucoma (G),
Glaucoma is considered as one of the most dangerous eye 4. Cataract (C),
illnesses. Glaucoma causes loss of eye site and is evident in the 5. Age-related Macular Degeneration (A),
human eye. Glaucoma is a very serious illness that can cause
blindness. If it is not curved in its early stages, it may cause 6. Hypertension (H),
blindness. As a result, detect this condition, a mechanism that
relies on the deep neural network learning for the analysis, 7. Pathological Myopia (M),
CNN is suggested. The proposed mechanism is based on a six- 8. Other diseases/abnormalities (O)
layer structure. CNN will work as cauterizing pattern according
to the architecture. The patient's eyes were examined for signs B. Data Preprocessing –
of glaucoma. images of the eyes.[18] 1) We are using information related to cataracts and
In [19], they discorded a method to detect Diabetic Reti- normal in our research. Then we are creating dataset from
nopathy and Glaucoma at an early stage. This system primarily images along with that also resizing images.
serves as a referral trigger, informing the patient that a retinal
IV. OVERVIEW OF PROPOSED MODEL
expert should be consulted if a positive result is detected. A test
set and real-time photos are used to test the less complicated We have implemented 3 models namely CNN, VGG16, and
pre-trained model. The accuracy rate was set at 80%. Using Inception V3, and interpreted results from them.
parameter tweaking and cross-validation techniques, this accu-
racy may be improved even further. The system also offers a
graphical user interface that is simple to use and understand
(GUI).

2
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 A. Convolutional Neural Network: - A. CNN -
A convolutional neural network has an architecture similar
to the human brain. It contains layers where each layer is con-
nected to the next layer through neurons. It uses the convolution
method which produces a function that helps in understanding
how a change in one is affecting change in another.
Finally, ConvNet's task is to compress the images into a
more manageable format while maintaining key components for
a good prediction. Convolutional neural networks are made up
of many layers of artificial neurons. Neurons are mathematical
functions which calculate the weight of each input and give an
activation value which helps in deciding whether a neuron will
fire or not. When you input an image into a ConvNet, each lay-
er develops several activation functions that are passed on to the
next layer.
B. VGG 19: -
A 3x3 receptive field and 19 convolution layers make up the
VGG-19 network. There are five of these levels, each with a
2x2 Max pooling layer. Three fully linked layers follow the
final Max pooling layer,aafter that, three fully connected layers
are added. It uses the softmax classifier as a final layer. All hid-
den layers have their Relu activated. The VGG19 model has the
disadvantage of being expensive to evaluate and requirees a lot
of memory and parameters.
VGG19 contains almost 138 million parameters. The major-
ity of these parameters (about 123 million) are in fully-
connected layers, which are substituted in our model by an
SVM classifier, greatly lowering the number of required pa- Fig. 1. CNN Architecture
rameters.
B. VGG 19 -
C. Inception V3: -
Google Net published Inception V3 in 2014, a CNN-based
Deep Learning model. This model contains 42 layers and has a
lower error rate than its predecessors. It uses an auxiliary classi-
fier to spread label information throughout the network and has
7x7 convolutions. It employs RMPS optimizers and includes
label smoothing, a king of regularizing component that is intro-
duced to the loss formula to keep the network from getting out-
liers in a class and this prevents over-fitting in the model that
gives best results.
V. PROPOSED MODEL ARCHITECTURE
We have proposed 3 deep learning models namely CNN,
INCEPTION V-3, VGG 19 for cataract prediction and follow-
ing is the model architecture summary for each model.
Proposed model architecture: -

Fig. 2. VGG 19 Architecture

3
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 C. INCEPTION V3 - The confusion matrix for VGG-19 which was best perform-
ing model is as follows –

Fig. 3. INCEPTION V3 Architecture

VI. RESULTS AND DISCUSSIONS

The results which we have obtained are as follows –

Fig. 6. Confusion Matrix

VGG-19 was the best performing model followed by Incep-

tion V3 and then CNN. There is a big gap between accuracies
of CNN,VGG-19 and Inception V3. The reason behind it is the
number of layers this model provides are far higher than custom
CNN model.
This fig 7 represents the CNN Model’s loss curve which is
the best performing model.
Fig. 4. Confusion Matrix

The confusion matrix for Inception V3 which was best per-

forming model is as follows –

Fig. 7. Loss Curve

Fig. 5. Confusion Matrix

4
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 This fig 8 represents the CNN Model’s accuracy curve This fig 10 represents the Inception V3 Accuracy curve
which is the best performing model. which is the best performing model.

Fig. 10. Accuracy Curve

Fig. 8. Accuracy Curve This fig 11 represents the VGG 19 Model’s loss curve
which is the best performing model.
This fig 9 represents the Inception V3 loss curve which is
the best performing model.

Fig. 11. Loss Curve

This fig 12 represents the VGG 19 Model’s accuracy curve

which is the best performing model.

Fig. 9. Loss Curve

5
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 VIII. LIMITATIONS AND FUTURE SCOPE
Deep Learning models require a large dataset to determine
model accuracy. Dataset collectionon requires equipment that
may not be available in local hospitals, and this would prevent
data collection on a broader scale. Another restriction of this
study is that any model must be thoroughly tested before being
deployed for medical use, as the subject is extremely sensitive
when dealing with people's lives.
As a future scope to this research, we aim to increase the
model accuracy by collecting more data and improve the model
architecture by ensembling techniques aiming to reduce the
percentage of false negative errors in the model.
IX. CONCLUSIONS
Deep Learning approaches are unique ways for detecting
and classifying distinct anomalies in eye pictures, and they have
a lot of promise for diagnosing ocular diseases efficiently. Us-
ing these deep learning algorithms, decision making becomes
powerful in field of ocular disease.
VGG-19 model stood out with the highest accuracy among
Fig. 12. Accuracy Curve different models. It is giving better accuracy as compared to
CNN and Inception -V3. The creation of new models to show
VGG16 neural network design is good. It is a straightfor- and extract characteristics that aid in the prognosis, diagnosis,
ward stack of convolutional and max-pooling layers, each one and follow-up of ocular disorders is required for novel medical
coming before a final layer that is fully connected. In other equipment utilization for detecting eye diseases. As a result,
words, it cannot extract features that are extremely complex. creating deep learning algorithms that incorporate multi-modal
Inception nets, on the other hand, have inception modules that input is still difficult.
are made up of 1X1 filters, also referred to as point wise convo-
lutions, followed by convolutional layers with various filter Despite the excellent findings, these methodologies face
sizes applied concurrently. Inception nets can now learn more significant open difficulties in terms of interpretability and
intricate characteristics medical professional feedback to the models. Furthermore, the
use of deep learning techniques in medical institutions has the
We can use VGG-19 and Inception V3 to solve complicated potential to enhance the number of individuals diagnosed, re-
issues. In our instance, VGG-19 performs somewhat better than sulting in an improvement in population quality of life. To fully
Inception V3. realize the promise of these tools, engineers and ophthalmolo-
We tested our model on testing dataset and we got the fol- gists must collaborate in a coordinated, multidisciplinary effort
lowing results. centered on the patient in order to reduce medical diagnosis
time and costs.
REFERENCES
[1] Liu, Chenxi, Dian Jiao, and Zhe Liu. "Artificial intelligence (AI)-aided
disease prediction." BIO Integration 1.3 (2020): 130-136.
[2] Li, N., Li, T., Hu, C., Wang, K., Kang, H. (2021). A Benchmark of
Ocular Disease Intelligent Recognition: One Shot for Multi-disease
Detection. In: Wolf, F., Gao, W. (eds) Benchmarking, Measuring, and
Optimizing. Bench 2020. Lecture Notes in Computer Science(), vol
12614. Springer, Cham. https://doi.org/10.1007/978-3-030-71058-3_11
[3] Karthiyayini, R., Shenbagavadivu, N. Retinal Image Analysis for Ocular
Disease Prediction Using Rule Mining Algorithms. Interdiscip Sci
Comput Life Sci 13, 451–462 (2021).
[4] Parampal S. Grewal, Faraz Oloumi, Uriel Rubin, Matthew T.S.
Fig. 13. Prediction Results Tennant,Deep learning in ophthalmology: a review,Canadian Journal of
Ophthalmology, Volume 53, Issue 4,2018,Pages 309-313, ISSN 0008-
4182,https://doi.org/10.1016/j.jcjo.2018.04.019.
VII. SCOPE OF RESEARCH
[5] M. T. Islam, S. A. Imran, A. Arefeen, M. Hasan and C. Shahnaz, "Source
Pattern recognition is used by ophthalmologists to diagnose and Camera Independent Ophthalmic Disease Recognition from Fundus
problems by looking at the eye and its surrounding tissues di- Image Using Neural Network," 2019 IEEE International Conference on
rectly or indirectly. This study focuses on the prediction of cata- Signal Processing, Information, Communication & Systems
(SPICSCON), 2019, pp. 59-63, doi:
ract illness, which is one of the ocular diseases. 10.1109/SPICSCON48833.2019.9065162.
[6] Salem, H., Negm, K.R., Shams, M.Y., Elzeki, O.M. (2022). Recognition
of Ocular Disease Based Optimized VGG-Net Models. In: Hassanien,

6
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap
 A.E., Bhatnagar, R., Snášel, V., Yasin Shams, M. (eds) Medical Signal Processing, Information, Communication & Systems
Informatics and Bioimaging Using Artificial Intelligence . Studies in (SPICSCON), 2019, pp. 59-63, doi:
Computational Intelligence, vol 1005. Springer, Cham. 10.1109/SPICSCON48833.2019.9065162.
https://doi.org/10.1007/978-3-030-91103-4_6 [14] A. Saxena, A. Vyas, L. Parashar and U. Singh, "A Glaucoma Detection
[7] D. Wang and L. Wang, "On OCT Image Classification via Deep using Convolutional Neural Network," 2020 International Conference on
Learning," in IEEE Photonics Journal, vol. 11, no. 5, pp. 1-14, Electronics and Sustainable Communication Systems (ICESC), 2020, pp.
Oct. 2019, Art no. 3900714, doi: 10.1109/JPHOT.2019.2934484. 815-820, doi: 10.1109/ICESC48915.2020.9155930.
[8] PERDOMO CHARRY, Oscar Julián and GONZALEZ, Fabio Augusto. [15] K. Prasad, P. S. Sajith, M. Neema, L. Madhu and P. N. Priya, "Multiple
A Systematic Review of Deep Learning Methods Applied to Ocular eye disease detection using Deep Neural Network," TENCON 2019 -
Images. Cienc. Ing. Neogranad 2020. 2019 IEEE Region 10 Conference (TENCON), 2019, pp. 2148-2153,
[9] Ma, Xiaojun, et al. "Study on a prediction of P2P network loan default doi: 10.1109/TENCON.2019.8929666.
based on the machine learning LightGBM and XGboost algorithms [16] Yaroub Elloumi, Mohamed Akil, Henda Boudegga, "Ocular diseases
according to different high dimensional data cleaning." Electronic diagnosis in fundus images using a deep learning: approaches, tools and
Commerce Research and Applications 31 (2018): 24-39. performance evaluation," Proc. SPIE 10996, Real-Time Image
[10] Singh, Vishal, et al. "Prediction of Modernized Loan Approval System Processing and Deep Learning 2019, 109960T (14 May
Based on Machine Learning Approach." 2021 International Conference 2019); https://doi.org/10.1117/12.2519098.
on Intelligent Technologies (CONIT). IEEE, 2021. [17] Simi Sanya , M Seetha, 2021, Detection of Ocular Diseases using
[11] A. Soni and A. Rai, "A Novel Approach for the Early Recognition of Ensemble of Deep Learning Models, INTERNATIONAL JOURNAL OF
Diabetic Retinopathy using Machine Learning,"2021 International ENGINEERING RESEARCH & TECHNOLOGY (IJERT) Volume 10,
Conference on Computer Communication and Informatics (ICCCI), Issue 09 (September 2021)
2021, pp. 1-5, doi: 10.1109/ICCCI50826.2021.9402566. [18] Tayal, A., Gupta, J., Solanki, A. et al. DL-CNN-based approach with
[12] A. Pandey, P. Patre and J. Minj, "Detection of Glaucoma Disease using image processing techniques for diagnosis of retinal diseases.
Image Processing, Soft Computing and Deep Learning Approaches," Multimedia Systems (2021). https://doi.org/10.1007/s00530-021-00769-7
2020 Fourth International Conference on I-SMAC (IoT in Social, [19] Sarki, R., Ahmed, K., Wang, H., Michalska, S., & Zhang, Y. (2020).
Mobile, Analytics and Cloud) (I-SMAC), 2020, pp. 1-7, doi: 10.1109/I- Early detection of diabetic eye disease from fundus images with deep
SMAC49090.2020.9243596. learning. Lecture Notes in Computer Science, 234–241.
[13] M. T. Islam, S. A. Imran, A. Arefeen, M. Hasan and C. Shahnaz, "Source https://doi.org/10.1007/978-3-030-39469-1_20
and Camera Independent Ophthalmic Disease Recognition from Fundus
Image Using Neural Network," 2019 IEEE International Conference on

7
horized licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:14:28 UTC from IEEE Xplore. Restrictions ap