2023 2nd International Conference on Futuristic Technologies (INCOFT)

Karnataka, India. Nov 24-26, 2023

Eye Disease Classification Using ResNet-18 Deep

Learning Architecture
1 2 3
Gurjot Kaur Neha Sharma Rahul Chauhan
Chitkara University Institute of Engineering Chitkara University Institute of Engineering Computer Science and Engineering
and Technology, Chitkara University, and Technology, Chitkara University, Graphic Era Hill University
Punjab, India Punjab, India Dehradun, Uttarakhand, India
[email protected] [email protected] [email protected]
4 5
Sanjeev Kukreti Rupesh Gupta
Computer Science & Engineering Chitkara University Institute of Engineering
Graphic Era Deemed to be University, and Technology, Chitkara University,
Dehradun, Uttarakhand, India, 248002 Punjab, India
[email protected] [email protected]
2023 2nd International Conference on Futuristic Technologies (INCOFT) | 979-8-3503-0884-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/INCOFT60753.2023.10425690

Abstract— The present study aims to investigate the crucial may threaten both the ability to see and ocular well-being.
topic of automated categorization of eye diseases using medical This introduction text provides a complete overview of
photographs by utilizing the capabilities of the ResNet-18 various eye disorders, encompassing a discussion on
model. The purpose of doing this research derives from the prevalent ocular conditions, their associated symptoms,
positive outcomes documented in recent studies that have etiological factors, and methods for detection and
employed ResNet-18 for comparable objectives. This study management. Several frequent eye conditions include
introduces a customized ResNet-18 model architecture Refractive Errors, Cataracts, Glaucoma, Age-Related
explicitly designed for classifying eye illness images into four Macular Degeneration (AMD), Diabetic Retinopathy,
categories with significant medicinal properties. The
Retinal Detachment, Conjunctivitis, Dry Eye Syndrome,
methodology employed in this study utilizes a dataset
consisting of 4,217 photos of eye diseases. The model was
Keratoconus, and Color Blindness. This research paper
trained over 30 iterations, using a batch size 128 and default focuses on three ocular illnesses alongside a healthy eye. The
learning rates. The results demonstrate a noteworthy main goal of eye illness classification is to utilize machine
accomplishment, as the ResNet-18 model suggested in this learning and computer vision methodologies to examine
study achieved a commendable accuracy rate of 94%. This medical photographs and identify the following four
highlights the model's efficacy in distinguishing between diseases: cataract, diabetic retinopathy, glaucoma, and
various eye illnesses, offering substantial enhancements in the normal. Cataract [3] is a prevalent ocular disorder associated
precision and effectiveness of diagnostic protocols. The with aging, wherein the lens becomes clouded, resulting in
research findings are significant, as they lay the foundation for diminished visual acuity and impairment. Surgical
advancing automated systems that can efficiently and precisely intervention is viable for addressing this condition, wherein
identify eye disorders. This has the potential to bring about a the opacified lens is substituted with a synthetic counterpart.
transformative impact on the field of ophthalmic diagnostics This procedure effectively reinstates visual clarity and
and enhance the quality of patient treatment. enhances overall quality of life. Cataracts can cause light
sensitivity, fuzzy or fuzzy vision, and difficulty seeing at
Keywords—Eye Disease, ResNet-18, Deep Learning, night. It is caused by aging, genetics, UV exposure, smoking,
Transfer Learning, Image Classification. and certain medications. One can detect cataracts by doing
I. INTRODUCTION an eye checkup. Surgery to remove the clouded lens and
replace it with a synthetic lens is the best course of treatment.
The field of eye disease classification focuses on the Glaucoma [4] is a collection of ocular disorders that impair
study and development of algorithms and models to the optic nerve, frequently attributed to high intraocular
effectively categorize various forms of ocular illnesses using pressure. The condition progressively results in visual acuity
medical imaging data. The technology plays a crucial role in impairment, initially affecting the peripheral field of vision
aiding ophthalmologists and healthcare professionals in and potentially advancing to a state of total visual loss. The
accurately diagnosing and successfully treating various timely identification, intervention, and continuous
ocular conditions [1]. Eye diseases, called ocular problems or surveillance are crucial in safeguarding visual function and
ophthalmic disorders, involve various medical conditions mitigating permanent harm. Eye conditions include
that impact the eyes and their adjacent structures. The ocular glaucomas. At first, there are no symptoms (often called "the
organs play a crucial role in facilitating the faculty of vision, silent thief of sight"), peripheral vision loss, and tunnel
enabling individuals to see and comprehend the surrounding vision. That damages the visual nerve by raising intraocular
visual environment. Ensuring the optimal condition of our pressure. Regular eye exams that monitor intraocular
ocular organs is paramount for the holistic welfare and pressure (tonometry) can identify glaucoma. Medication,
standard of living. The eyes are commonly regarded as the laser therapy, or surgery may be recommended to lower
channels through which the essence of an individual's inner ocular pressure. Injections of anti-VEGF medication can help
being is revealed. However, they also serve as complex slow the condition's progression, but there is no cure.
physiological structures that play a vital role in our mental Diabetic retinopathy [5] is a pathological condition
interpretation of the surrounding environment [2]. associated with a form of diabetes that impacts the
Unfortunately, similar to other anatomical components, the vasculature of the retina. The condition has the potential to
eyes are susceptible to various illnesses and conditions that

979-8-3503-0884-6/23/$31.00 ©2023 IEEE 1

Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 08:02:14 UTC from IEEE Xplore. Restrictions apply.
 induce visual impairment, characterized by symptoms such framework for Dry Eye Disease (DED), focusing specifically
as blurred or distorted vision, and in more extreme instances, on multi-class classification.
may result in complete loss of sight. The importance of early
detection, routine eye examinations, and effective diabetes The identification of numerous diabetic eye diseases
management cannot be overstated in preventing and (DEDs) using retinal fundus pictures is a significant area of
controlling this condition. Dark spots, fuzzy or erratic vision, research that carries practical implications. The suggested
and floating are signs of diabetic retinopathy. It causes model underwent testing using a diverse set of retinal fundus
uncontrolled, long-term hyperglycemia that impairs the photos obtained from a publically accessible dataset, which
circulation of the retina. Eye conditions can greatly influence an ophthalmologist had annotated. The present study utilized
our independence and quality of life. Nonetheless, eyesight a novel CNN architecture to experiment. The proposed
can frequently be preserved, and the effects of these model for multi-class categorization attained a peak accuracy
disorders are lessened with early detection and proper of 81.33%. Nazir et al. (2020) [11] present a strategy for
therapy [6]. To maintain maximum eye health, following a automated disease localization and segmentation using the
healthy lifestyle, getting medical guidance when needed, and Fast Region-based Convolutional Neural Network (FRCNN)
having regular eye exams are important. A vital first step in algorithm combined with fuzzy k-means (FKM) clustering.
identifying and treating these conditions is speaking with an The FRCNN model is trained using annotated pictures to
eye care specialist, regardless of whether you have perform localization, further segmenting the localized
symptoms or are worried about the health of your eyes. regions using FKM clustering. The methodology's
Remember that maintaining your general well-being is as effectiveness in disease identification and segmentation has
important as preserving your vision. Eye problems can been confirmed through a thorough comparison with the
significantly impact an individual's quality of life and general current methods. Sarki et al. (2021) [12] provide a systematic
well-being. Prompt and precise diagnosis is crucial for examination of the importance of image processing in the
prompt treatment and preventing blindness [7]. Deep classification of dry eye disease (DED). The most favorable
learning methods have demonstrated significant potential in outcomes were achieved by employing conventional image
medical image analysis in recent years. This study processing techniques in conjunction with a novel
investigates the automated identification of several eye convolutional neural network (CNN) structure. Integrating
illnesses using the deep convolutional neural network (CNN) the newly developed Convolutional Neural Network (CNN)
[8] ResNet-18 architecture. A large dataset of ocular pictures with the traditional processing of image methodology
will be gathered for the study, the ResNet-18 model will be showed superior performance in accurately classifying Dry
trained and adjusted, and its ability to classify eye illnesses Eye Disease (DED) cases. The experimental findings
will be assessed. The findings show how deep learning can demonstrated satisfactory levels of accuracy, specificity, and
transform ophthalmic diagnostics by offering quick and sensitivity. Smaida et al. (2021) [13] this study introduce the
accurate evaluations. DCGAN technique for the generation of synthetic medical
images.
The main contributions of this article are as follows:
Furthermore, the GMD (Glaucoma, Myopia, and
x The main goal of classifying eye illnesses is to Diabetic Retinopathy) model is employed to enhance the
identify the four conditions: diabetic retinopathy, classification of eye illnesses, both with and without
glaucoma, cataract, and expected by analyzing including synthetic medical images. Our dataset comprises
medical photographs using machine learning and four distinct categories of ocular disorders. There was a
computer vision techniques. notable improvement in the accuracy of the model, with an
increase from 76.58% in the training set and 76.42% in the
x This work offers empirical support for the ResNet- validation set to 80.45% in the training set and 83.74% in the
18 model's ability to reliably identify and categorize validation set. It is recommended that this study be extended
eye illnesses using photographic data, enabling the to other image classification models, such as Vgg16,
division of images into four classes. Inception v3, and ResNet, to improve the overall accuracy.
x 30 iterations of the model were run across 128 Shamsan et al. (2023) [14] aim to categorize a dataset on eye
batches. To assess the ResNet-18 model, several diseases using hybrid methods that combine fusion and
Confusion Matrix performance indicators are feature extraction techniques. To order CFP photos to
employed. diagnose eye diseases, three methods were developed. Using
features from the MobileNet and DenseNet121 models
II. LITERATURE REVIEW independently, an ANN is used as the first way to categorize
Sharma et al. (2023) [9] present a novel approach a dataset of eye diseases. The second approach uses fused
utilizing a deep Convolutional Neural Network (CNN) features from the MobileNet and DenseNet121 models
model built explicitly upon the ResNet50 architecture. The before and after feature reduction to classify the eye illness
objective is to accurately differentiate between regular and dataset using an artificial neural network. The third approach
cataract-infected classes within the provided photographs. uses artificial neural networks (ANN) to categorize the eye
The dataset chosen for this task is the Ocular Disease illness dataset using hand-crafted features and fused features
Intelligent Recognition dataset. This collection encompasses from the MobileNet and DenseNet121 models separately.
real-time patient information from both eyes. The model has The ANN achieved good accuracy based on the combination
a high level of precision. This model can potentially serve as of handmade features and MobileNet.
a valuable tool within the biomedical or healthcare domain III. METHODOLOGY
due to its innovative approach to medicine. Sarki et al.
(2022) [10] aim to create an automated classification The dataset depicted in Figure 1 [15] is referred to as the
Eye disease dataset, encompassing a total of 4,217
photographs that are categorized into four major eye

2
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 08:02:14 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. Fine Tuned ResNet 18 Architecture

diseases. The photos have been classified into four distinct B. Data Augmentation
categories, with three showing various eye disorders and the Data augmentation is a crucial technique in machine
final category indicating a normal eye. The data utilized in learning and artificial intelligence, significantly improving
this study was obtained through implementing the ResNet-18 models' performance and resilience. At its core, data
model, as documented in Kaggle. The following analysis augmentation encompasses enlarging a dataset by
comprehensively examines sections A, B, and C.A. Dataset implementing diverse modifications on already data points.
A. Description Employing this approach increases the volume of training
data and introduces variability in the dataset, enhancing the
An example of an input dataset may be seen in Figure 2, ability of machine learning models to capture complex
a collection of photographs illustrating different eye patterns and generalize more efficiently. In computer vision,
disorders in their various forms. The dataset can be broken using data augmentation methods, including picture rotation,
down into a total of 4 categories. Cataracts, diabetic flipping, scaling, and color modifications, enhances the
retinopathy, glaucoma, and standard samples were obtained quality and diversity of datasets. This augmentation process
for analysis. The dataset must be used for training, testing, enhances the capability of models to identify objects and
and validation to fulfill its obligations. 85% of the dataset patterns in real-world situations accurately. Likewise, in
was used for training, while the remaining 15% was utilized natural language processing, techniques such as synonym
for validation and testing. substitution and random insertion are employed to enhance
Name of Classes Left Eye Right Eye the language comprehension capabilities of language models.

Original Rotation Zooming Horizontal

Flip
Cataract

Diabetic
Retinopathy

Glaucoma

Normal

Fig. 3.

Fig. 2. Input Dataset

3
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 08:02:14 UTC from IEEE Xplore. Restrictions apply.
 C. ResNet-18 Model
Residual Networks, or ResNets [16] for short, are a kind
of deep neural network architectures that have revolutionized
computer vision and the image classification job. A CNN
architecture known as ResNet-18 has shown to be helpful in
various computer vision applications, such as object
recognition, image segmentation, and picture classification.
In deep learning, the ResNet-18 architecture is well-known
for its remarkable effectiveness and simplicity. The
application of advanced DL models [17] has become a
significant role in medical diagnostics; a prominent instance
is the ResNet-18 model.
Using its architectural strength, ResNet-18 takes on the
difficult task of recognizing and categorizing a range of eye (a) Model Loss
illnesses from medical photographs. Because of its innate
depth and residual connections, the model can distinguish
between various diseases with remarkable ease because of its
ability to collect subtle and nuanced aspects in ocular
images. This model is an excellent option for limited-
resource medical settings because it balances expressive
strength and computational economy with its 18
convolutional layers. This study investigates the application
of the ResNet-18 model in the vital area of classifying eye
illnesses. This research's significance lies in applying
cutting-edge deep learning technology to enhance the speed
and precision of identifying eye disorders. Better patient
outcomes and more effective therapies may result from the
ResNet-18 model's fast, trustworthy, and accurate
assessments. We'll discover more about the methodology, (b) Model Accuracy
results, and consequences of this cutting-edge application as
Fig. 4. Training and Validation Curve (a)Loss and (b) Accuracy
we progress through the study, giving us an understanding of
how ResNet-18 is expanding the body of knowledge in the B. Confusion Matrix
field of ophthalmology.
The representation of the confusion matrix parameter can
IV. RESULTS be shown in Figure 5. The matrix that has been constructed
possesses dimensions of 4x4, which can be attributed to the
The ResNet-18 model employed in this study was
diverse qualities demonstrated by four distinct eye disorders.
evaluated on the Google Colab platform. The ResNet-18
Using a confusion matrix enabled the visual depiction of the
deep learning model underwent training for a duration of 30
classification accuracy. Precision, recall, F1-score, and
epochs. The AdamW optimizer was utilized with the default
accuracy are performance measures that illustrate the
learning rate and a batch size 128 for each epoch. The
evaluation of a system's performance. Table 1 displays the
evaluation of the system's performance is conducted by
performance outcomes.
employing metrics derived from the Confusion Matrix. The
assessment of the ResNet-18 model involves examining its
effectiveness inappropriately categorizing a dataset of photos
into four separate classifications of ocular disorders. The
dataset consists of 4,217 photographs depicting various eye
illnesses. These photographs have been classified into
several groups to facilitate training, validation, and testing.
A. Accuracy and Loss analysis
This section explains the graphical representations of the
Accuracy and Loss plots. The study is conducted with the
ResNet-18 model as the chosen methodology. Figure 4
provides a visually comprehensive illustration of this topic.
Figure 4(a) illustrates the loss of the model. Figure 4(b)
visually represents the model's accuracy, allowing for a
comparative analysis. The graph depicts the representation of
the training accuracy and loss, with the blue line indicating
these metrics. Similarly, the green line shows the validation
loss and accuracy. A practical method for visualizing the
correlation between the number of epochs and the quality of Fig. 5. Confusion Matrix
the model is through a graphical representation that
demonstrates the trade-off between reliability and accuracy
with the progressive increment of epochs.

4
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 08:02:14 UTC from IEEE Xplore. Restrictions apply.
 C. Performance Parameters REFERENCES
Photographs are categorized based on the confusion [1] He, J., Li, C., Ye, J., Qiao, Y., & Gu, L. (2021). Multi-label ocular
matrix in Table 1, which considers the four distinct types of disease classification with a dense correlation deep neural network.
Biomedical Signal Processing and Control, 63, 102167.
eye illnesses. The dataset on eye diseases is organized into
four unique categories, each assigned to a particular group [2] Ramanathan, G., Chakrabarti, D., Patil, A., Rishipathak, S., &
Kharche, S. (2021, October). Eye disease detection using Machine
based on the classification performance of a ResNet-18 Learning. In 2021 2nd Global Conference for Advancement in
model. This performance is evaluated using the Confusion Technology (GCAT) (pp. 1-5). IEEE.
Matrix. The assessment of the ResNet-18 model [3] Khan, M. S. M., Ahmed, M., Rasel, R. Z., & Khan, M. M. (2021,
encompasses the examination of many attributes, such as May). Cataract detection using a convolutional neural network with
accuracy, precision, F1 score, and recall. The analysis of VGG-19 model. In 2021 IEEE World AI IoT Congress (AIIoT) (pp.
these data points yielded an accuracy level of 94%. 0209-0212). IEEE.
[4] Serte, S., & Serener, A. (2019, October). A generalized deep learning
model for glaucoma detection. In 2019 3rd International symposium
TABLE I. PERFORMANCE PARAMETER
on multidisciplinary studies and innovative technologies (ISMSIT)
(pp. 1-5). IEEE.
[5] Bhatia, K., Arora, S., & Tomar, R. (2016, October). Diagnosis of
Name of the Precision Recall F1- Accuracy diabetic retinopathy using machine learning classification algorithm.
Classes Score In 2016 2nd international conference on next generation computing
technologies (NGCT) (pp. 347-351). IEEE.
[6] Kumar, S., Pathak, S., & Kumar, B. (2019). Automated detection of
Cataract 0.96 0.97 0.96 eye related diseases using digital image processing. Handbook of
multimedia information security: techniques and applications, 513-
544.
Diabetic 0.99 0.99 0.99
Retinopathy [7] West, S., & Sommer, A. (2001). Prevention of blindness and
0.94 priorities for the future. Bulletin of the world Health Organization, 79,
244-248.
Glaucoma 0.94 0.89 0.91
[8] Singh, D., Rana, A., Gupta, A., Sharma, R., & Kukreja, V. (2023,
April). An Enhanced CNN-LSTM Based Hybrid Deep Learning
Normal 0.88 0.92 0.90 Model for Corn Leaf Eye Spot Disease Classification. In 2023 IEEE
12th International Conference on Communication Systems and
Network Technologies (CSNT) (pp. 147-151). IEEE.
[9] Sharma, G., Anand, V., & Gupta, S. (2023, July). Harnessing the
V. CONCLUSION Strength of ResNet50 to Improve the Ocular Disease Recognition. In
2023 World Conference on Communication & Computing (WCONF)
This study introduces meaningful progress in medical (pp. 1-7). IEEE.
image processing, explicitly focusing on categorizing ocular [10] Sarki, R., Ahmed, K., Wang, H., Zhang, Y., & Wang, K. (2022).
illnesses utilizing the ResNet-18 neural network model. The Convolutional neural network for multi-class classification of diabetic
efficacy of this architecture has been substantiated by eye disease. EAI Endorsed Transactions on Scalable Information
Systems, 9(4), e5-e5.
organized experimentation, revealing its exceptional ability
[11] Nazir, T., Irtaza, A., Javed, A., Malik, H., Hussain, D., & Naqvi, R.
to attain a praiseworthy accuracy rate of 94% on the dataset A. (2020). Retinal image analysis for diabetes-based eye disease
utilized in our study. The potential of ResNet-18 as a robust detection using deep learning. Applied Sciences, 10(18), 6185.
tool for automated disease diagnosis is shown by its effective [12] Sarki, R., Ahmed, K., Wang, H., Zhang, Y., Ma, J., & Wang, K.
implementation, which involved training over 30 epochs (2021). Image preprocessing in classification and identification of
with a batch size of 128 and leveraging the AdamW diabetic eye diseases. Data Science and Engineering, 6(4), 455-471.
optimizer. This accomplishment has significant promise, as [13] Smaida, M., Yaroshchak, S., & El Barg, Y. (2021). DCGAN for
precise and effective disease categorization is crucial in Enhancing Eye Diseases Classification. In CMIS (pp. 22-33).
clinical practice, where prompt interventions can [14] Shamsan, A., Senan, E. M., & Shatnawi, H. S. A. (2023). Automatic
significantly impact patient outcomes. The findings of this Classification of Colour Fundus Images for Prediction Eye Disease
Types Based on Hybrid Features. Diagnostics, 13(10), 1706.
study have broader implications that go beyond the
[15] Ramzan, F., Khan, M. U. G., Rehmat, A., Iqbal, S., Saba, T.,
numerical data. The ResNet-18 model's remarkable precision Rehman, A., & Mehmood, Z. (2020). A deep learning approach for
provides a look into the potential of medical diagnostics in automated diagnosis and multi-class classification of Alzheimer's
the future. Advanced deep learning models can enhance disease stages using resting-state fMRI and residual neural networks.
healthcare workers' skills, minimize diagnostic errors, and Journal of medical systems, 44, 1-16.
accelerate patient care. As the progression continues, more [16] Sharma, N., Sharma, A., & Gupta, S. (2022, December). A
enhancements and implementations of these models possess Comprehensive Review for Classification and Segmentation of
Gastro Intestine Tract. In 2022 6th International Conference on
the capacity to fundamentally transform the medical imaging Electronics, Communication and Aerospace Technology (pp. 1493-
domain and make substantial contributions to enhancing 1499). IEEE.
healthcare provision. [17] Sharma, G., Anand, V., & Kumar, V. (2023, August). Classification
of Osteo-Arthritis with the Help of Deep Learning and Transfer
Learning. In 2023 5th International Conference on Inventive Research
in Computing Applications (ICIRCA) (pp. 446-452). IEEE.

5
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 08:02:14 UTC from IEEE Xplore. Restrictions apply.