DEEP LEARNING BASED MULTI RETINAL

DISEASES IDENTIFICATION USING CNN

*Note: Sub-titles are not captured in Xplore and should not be used

line 1: 1st Given Name Surname line 1: 2nd Given Name Surname line 1: 3rd Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

line 1: 4th Given Name Surname line 1: 5th Given Name Surname line 1: 6th Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

Abstract—Eye diseases play a significant role in causing few common eye conditions. Blurred or distorted vision,
visual impairments and even blindness worldwide. To light sensitivity, discomfort, redness, or discharge from the
prevent eye illness, and such vision losses that are caused eye are a few symptoms these disorder might produce. If a
due to eye diseases, also to improve the overall quality of person notices any of these symptoms or changes inherit
life of patient, a method for early detection is needed so that vision, its crusial to consult a doctor right away so that they
timely treatment of these diseases can be done. Currently, get a necessary diagnosis and therapy. In order to correctly
diagnosis of eye diseases requires manual examination by a diagnose an eye disease, it is important to analyse all the
trained opthalmologist. However, this whole method can different range of symptoms.
take up a lot of time as well as can cost a lot of money, also Deep learning (DL), machine learning (ML), and artificial
may not readily available in certain regions.in recent years, intelligence (AI) have the potential to play significant roles
deep learning based technique have proven tremendous in the prevention, diagnosis, and treatment of eye diseases,
promise in the field of medicine and medical images according to studies and research done in recent years with
analysis. This paper proposes a novel approach for the growth in technology. These technologies have aided in
identifying various eye illeness using deep learning based the early diagnosis of diseases; The incorporation of
technique such as CNN. This study uses convolutional artificial intelligence (AI) and deep learning
neural networks to categorize eye disease using a publicly technologies,particularly Convolutional Neural Networks
available dataset. Five different pre-trained models based on (CNNs), has significantly advanced medical image
convolutional neural networks (CNNs), including VGG-16, processing in recent years. The detection of eye diseases has
VGG-19, ResNet-50, ResNet-152, and DenseNet-121, were gained significantly from these technologies. Medical
used in this study. We were able to detect eye diseases at the experts, especially ophthalmologists, can use these
cutting edge using the refined VGG-19. With testing algorithms to diagnose and classify eye diseases at an early
accuracy of 95% on the dataset, this model accurately stage, resulting in more effective and specific treatment
predicted eye diseases due to the effective and same strategies. There are several crucial phases involved in using
weighted precision, recall, and F1 score of 95%. The model retinal images to diagnose eye diseases, including feature
also significantly reduces training loss while improving extraction, categorization, and image pre-processing.
accuracy Accurate diagnoses can be made with the use of deep
learning algorithms used in conjunction with machine
Keywords—Eye Disease, Pre-Trained, Fine-Tuning, learning and image processing techniques.
Convolutional Neural Networks Globally, there are regional variations in the occurrence of
eye issues, which are determined by factors like age, gender,
occupation, lifestyle, economic status, habits, and cultural
I. INTRODUCTION
standards. For instance, tropical populations may experience
There are various conditions that affect the eyes and may higher rates of eye infections due to environmental factors
result in vision issues or even blindness and such conditions like dust, humidity, and sunshine. Both industrialized and
are casued due to eye diseases. From infants to the elderly, developing countries experience a high prevalence of eye
they can effect anyone, and they can be brought on by a disorders, with large optical morbidity rates found in several
number of things, such as genetics, envirornment influences,
Asian countries. However, these diseases are still
and way of life decisions refractive errors, cataracts,
underdiagnosed in some areas and are not adequately
glucoma, dry eye, conjunctivitis, agre-related macular
degeneration (AMD), and agre related macular edema are a treated. According to the World Health Organization

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

(WHO), approximately 285 million people worldwide approaches for analysis and feature extraction, often
experience visual problems, with 246 million having poor utilizing limited and small-scale datasets [13]. A network
eyesight and 39 million being blind[1-2]. It is crucial to named ReLayNet [14], structured as an encoder-decoder,
offer affordable or free comprehensive eye care services to has been developed to segment various layers of the retina,
residents of underprivileged neighbourhoods and slums. including the identification of accumulated fluid in
Moderate-to-severe distance nearsightedness or blindness images. The efficient utilization of deep neural networks
can result from a variety of eye conditions, including for screening age-related macular degeneration (ARMD)
untreated presbyopia, unresolved refractive errors, cataracts, has been achieved through the analysis of color fundus
glaucoma, corneal opacities, diabetic retinopathy, and images. Experiments were conducted in the study [15]
trachoma. The facts indicate that these diseases impact using a customized VGG-16 architecture that incorporates
roughly 1 billion people worldwide. It is believed that batch normalization. A proficient convolutional neural
around half of these cases could have been avoided or network (CNN) was crafted for analyzing the optic disc
properly treated if people had better access to eye care and through digital fundus images. This involved a swift and
treatment. automated deep learning approach for detecting glaucoma
[16]. Various traits of abnormal eyes were identified by
2. LITERATURE REVIEW utilizing pre-trained convolutional neural networks (CNN)
Deep learning has found application in numerous aspects of and other deep learning-based neural networks. These
cataract management, encompassing both clinical and models were employed for the classification of diseases
surgical domains. Its applications range from the diagnosis and the early detection of abnormalities in their initial
of cataracts to enhancing biometry for precise calculation stages [17-18]. The major contributions are enlisted below.
of intraocular lens (IOL) power. In order to identify and 
categorize eye diseases, Xu [3] used the AlexNet and
Visual-DN CNN-based algorithms. They achieved an
accuracy of 86.2% using 8,030 fundus images for eye 3. METHODOLOGY
cataracts identification. Another study conducted by Zhang A. Dataset In this study, we utilized the publicly available
[4] also focused on cataract detection and grading. They Ocular Disease Intelligent Recognition (ODIR) dataset from
reported a significant improvement in accuracy, achieving Kaggle (see Figure 1). This dataset consists of colored
an impressive 93%. Their analysis was based on 1,352 fundus images from a diverse group of
fundus images. In the study proposed by Gosh [5], they individuals,encompassing 8 distinct ocular disease diagnosis
utilized a CNN model and a diverse dataset that included categories. The ODIR dataset comprises comprehensive
glaucoma, retinal disease, and normal eye cataracts. ophthalmic data from 5,000 patients, including age, color
Unfortunately, further details about their findings were not fundus images of both left and right eyes, and diagnostic
provided in the current text. Their CNN achieved an keywords provided by medical professionals. The dataset
accuracy of 82%, which is considered acceptable based on aims to present a comprehensive and authentic
CNN standards. In a recent study conducted by Ahmed representation of patient data gathered from multiple
they focused on cataract analysis using a CNN model hospitals and medical centers across China by Shang Gong
with VGG-19 architecture. Impressively, their = research Medical Technology Co., Ltd[19].
achieved an outstanding overall accuracy of 97.47%, with
a precision rate of 97.47% and a relatively low loss of
5.27% [6]. The performance of the pretrained models
cannot be neglected due to the effective performance in
vision problems[7-9], and [10]. This paper's primary focus
lies in detecting cataract disease through transfer learning.
It also presents a comparative analysis of five different
deep learning models, namely DenseNet121, ResNet50,
ResNet152, VGG-16 and VGG-19 InceptionV3, and
InceptionResNetV2 for cataract disease detection. The
research explores various approaches and hyperparameters
used to implement these models to achieve accurate
identification of eight different types of eye diseases An
advanced and automated image processing algorithm was
suggested for diagnosing glaucoma from fundus images in
the study [11]. This method employs bend point detection
and tracks blood vessels to enhance accuracy and
reliability in the diagnosis process. In the study [12], a
tailored threshold-based algorithm was crafted to identify
red lesions associated with diabetic retinopathy (DR) in
fundus images. Each image was processed independently
in this approach. Diabetic retinopathy (DR) leads to
diverse retinal abnormalities, such as hard exudates,
hemorrhages, microaneurysms, and various symptoms.
Various machine-learning methods have been created to
detect DR and other diseases. These methods employ B. Transfer Learning In this study, we have considered the
different image processing and computer vision transfer learningtechnique[20] where a pre-existing model,
trained on one task, is utilized to tackle a different but  DenseNet-201: With 201 layers and approximately
related task. Instead of training a model from scratch, 20.6 million trainable parameters, DenseNet-201's
transfer learning takes advantage of the knowledge and dense connection pattern makes it an excellent choice
learned representations already present in the pre-trained for various computer vision tasks. It allows for optimal
model to accelerate and improve learning on the new target feature reuse and smooth gradient flow across the
task. Deep neural networks trained on extensive datasets, network [23].
like ImageNet for image recognition or BERT for natural
language processing, capture generic features and patterns
that can be transferred to various tasks. These models learn
hierarchical representations that can be repurposed for D. Hyper Parameters
different problems. By taking a pre-trained model, initially To obtain optimal performance, we carefully set the
designed for a complex task, and applying it to a related yet hyperparameters for training the deep CNN models
distinct task or a smaller dataset, significant time and employed in this work. Iterative experimentation was used
computational resources can be saved, all while enhancing to fine-tune the hyperparameters in order to strike the
the model's performance on the new task. In this study, we correct balance between model complexity and training
leverage pre-trained models that were trained on large efficiency.
datasets and then fine-tune them for specific tasks defined We fixed the number of epochs for all models to 200 to
within our target dataset. By doing so, we aim to leverage ensure that the models get enough iterations over the
the insights and learned representations captured by these training data without ovlearning rate of 0.001 was used to
pre-trained models to enhance performance on our specific achieve constant convergence during gradient updates,
task. In this study, we have used all the ImageNet weights avoiding abrupt oscillations that could impede training.To
for the eye disease classification. further regularize the models and prevent overfitting, we
employed a batch size of 32, balancing computational
efficiency with the ability to capture meaningful gradients
from each minibatch. Additionally, we utilized the Adam
C. Pretrained Model Architectures optimizer due to its adaptive learning rate and momentum
features, which help in efficient weight updates during
Pre-trained models are models that have been trained on
training, see Table 1 for hyper parameters values.
large datasets, such as ImageNet, which contains millions of
records and images. These models are already created and
trained by others, that is why referred to as pre-trained
models. For instance, in the context of eye disease
classification in this study, we utilize pre-trained models like
VGG16, RESNET50, and others, which have been trained
on extensive datasets. Among the deep CNN architectures
used for eye disease classification in this research are:

 ResNet50: It is a well-known convolution neural

network with 50 layers and approximately 23.5 million
trainable parameters. Its design, incorporating residual
blocks, has led to its widespread adoption as a
preferred model for various computer vision tasks [21].

 ResNet152: This model consists of 152 layers and

around 60.2 million trainable parameters. Its use of
residual blocks enhances its capacity to learn and
represent complex visual patterns, making it a powerful
model for diverse computer vision applications [21].

 VGG16: It is a 16-layer variation of VGG models,

comprising 13 convolutional layers and three fully
connected layers. VGG16 has been widely used in
various computer vision tasks for its effectiveness and
simplicity [22].

 VGG-19: It is a variation of the VGG model with 19

layers, comprising 16 convolution layers, 3 fully
connected layers, 5 MaxPooling layers, and 1 SoftMax
layer. VGG-16, on the
other hand, is a 16-layer version of VGG models, containing
13 convolutional layers and three fully connected layers
[22].
 4. RESULTS
The evaluation of our eye disease detection model on the
test set has provided us with valuable insights into its
performance, offeriring implications for our project's
objectives. Remarkably, the model achieved an accuracy of
95%, showcasing its competence in precisely classifying
various e ye diseases and bolstering its relevance in real
clinical settings. A deeper analysis of precision, rscores for
each eye disease category sheds light on the model's
effectiveness in identifying specific conditions. In the
training process, accuracy is increasing while loss is 5. CONCLUSION
decreasing, see Figure 3.The confusion matrix further
enhances ouunderstanding of the model's predictive In this study, we explore the performance of various pre-
behavior, revealing potential misclassifications or biases trained classification models in the context of classifying
towards certain classes (see Figure proposed model tested eye diseases from a multiclass perspective. Detecting and
for approximately 973 images. By scrutinizing the classifying eye problems automatically remains a
distribution of predicted labels in comparison to the ground challenging task, particularly for early diagnosis. However,
truth labels, we can pinpoint areas where the model by leveraging transfer learning of existing models, we were
mightbenefit from further refinement or targeted able to achieve high accuracy while reducing the workload
improvement. These findings empower us to refine and on the new model. We evaluated the performance of five
optimize the model, ensuring it delivers accurate and pre-trained models: VGG-16, VGG-19, ResNet-50, ResNet-
reliable results in the challenging crealm of eye disease 152, and DenseNet-121, on the task of classifying eye
detection. diseases. To assess the classification process, we used
several evaluation metrics, including recall, precision,
accuracy, and F1-score, across all five models. Among the
models tested, VGG-19 emerged as the top-performing one,
achieving an impressive 95% accuracy, recall, precision, and
F1-score. The results shown that transfer learning is
effective to gain effective results with limited data and
limited training process. The computerized classification of
eye diseases has advanced significantly as a result.
However, in order to improve the methods and carry out a
more thorough examination.The ultimate goal is to improve
diagnostic precision and help identify and treat eye diseases
more effectively. The future direction could involve
implementing advanced classification methods with
extensive datasets. This approach can be applied to address
other challenges in medical imaging, aiming for prompt and
dependable results.

REFERENCES
[1] S. Resnikoff “Global data on visual impairment in the
year 2002,” SciELO Public HealthS Resnikoff, D Pascolini,
D Etya’Ale, I Kocur, R Pararajasegaram, GP Pokharel, SP
MariottiBulletin of the world health organization,
2004•SciELO Public Health, vol. 82, no. 11, 2004,
Accessed: Dec. 26, 2023. [Online]. Available:
https://www.scielosp.org/pdf/bwho/v82n11/v82n11a09.pdf
[2] D. Pascolini, S. Mariotti, … G. P.-O., and undefined
2004, “2002 global update of available data on visual
impairment: a compilation of population-based prevalence
studies,” Taylor & FrancisD Pascolini, SP Mariotti, GP
Pokharel, R Pararajasegaram, D Etya’Ale, AD
NégrelOphthalmic epidemiology, 2004•Taylor & Francis,
vol. 11, no. 2, pp. 67–115, Apr. 2004, doi:
10.1076/opep.11.2.67.28158. [3] X. Xu, L. Zhang, J. Li, …
Y. G.-I. journal of biomedical, and undefined 2019, “A
hybrid global-local representation CNN model for automatic [6] M. Khan, M. Ahmed, … R. R.-2021 I. W. A. I., and
cataract grading,” ieeexplore.ieee.orgX Xu, L Zhang, J Li, Y undefined 2021, “Cataract detection using convolutional
Guan, L ZhangIEEE journal of biomedical and health neural network with VGG-19 model,”
informatics, 2019•ieeexplore.ieee.org, Accessed: December. ieeexplore.ieee.orgMSM Khan, M Ahmed, RZ Rasel, MM
26, 2023. [Online]. Available: Khan2021 IEEE World AI IoT Congress (AIIoT),
https://ieeexplore.ieee.org/abstract/document/8705272/ 2021•ieeexplore.ieee.org, Accessed: Dec. 26, 2023.
[4] H. Zhang, K. Niu, Y. Xiong, W. Yang, … Z. H.-C. [Online]. Available:
methods and, and undefined 2019, “Automatic cataract https://ieeexplore.ieee.org/abstract/document/9454244/
grading methods based on deep learning,” Elsevier, [7] M. Mubeen, M. A. Arshed, and H. A. Rehman,
Accessed: December. 26, 2023. [Online]. Available: “DeepFireNet - A Light-Weight Neural Network for Fire-
https://www.sciencedirect.com/science/article/pii/S0169260 Smoke Detection,” pp. 171–181, 2022, doi: 10.1007/978-3-
719307163 031-10525-8_14.
[5] C. L. Dondapati*, A. Ghosh, and Dr. TYJ. N. [8] “Multiclass Brain Tumor Classification from MRI
Malleswari, “Classification of Eye Disorders based on Deep Images using Pre-Trained CNN Model,” VFAST
Convolutional Neural Network,” International Journal of Transactions on Software Engineering, doi:
Innovative Technology and Exploring Engineering, vol. 9, 10.21015/VTSE.V10I4.1182.
no. 6, pp. 1388–1393, Apr. 2020, doi:
10.35940/IJITEE.F4209.049620.