2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

Modified EfficientNetB3 Deep Learning Model to

Classify Colour Fundus Images of Eye Diseases
2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA) | 979-8-3503-3828-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCCMLA58983.2023.10346769

Riya Sharma Jayesh Gangrade Shweta Gangrade

Dept of Computer Science & Dept. of Artificial Intelligence & Dept of Information Technology Manipal
Engineering Machine Learning University Jaipur Jaipur, India
Manipal University Jaipur Manipal University Jaipur [email protected]
Jaipur,India Jaipur, India
[email protected] [email protected]

Ashish Mishra *Gautam Kumar Vinit Kumar Gunjan

Dept of Computer Science & Dept. of Artificial Intelligence & Department of Computer Science &
Engineering Machine Learning Engineering
Jaypee Institute of Information Manipal University Jaipur CMR Institute of Technology, Hyderabad,
Technology, Noida, India Jaipur, India India [email protected]
[email protected] [email protected]

Abstract— The only way to prevent blindness from eye This study aims to leverage the power of deep learning to
problems is by early detection and prompt treatment. Although build a reliable and effective system for the automated
colour fundus photography (CFP) is useful for fundus classification of eye diseases [3-4]. We shall pay particular
inspection, there is a need for computer-assisted automated attention to glaucoma, cataract, diabetic retinopathy, and a
diagnosis tools due to the similarities between the early normal category for healthy eyes [5]. We aim to develop a
symptoms of many eye disorders. The suggested approach uses model capable of precisely identifying and classifying these
cutting-edge deep learning model to categorize images into disorders by training a deep convolutional neural network
several disease categories by learning distinguishing features (CNN) using a huge dataset of annotated eye images [6].
from the input images. The high-resolution fundus photos from
individuals with diabetic retinopathy (DR), glaucoma, cataract, Despite substantial advancements in deep learning-based
and healthy eyes make up most of the dataset used for this medical image analysis, there are still several problems with
research. The experimental findings show that the suggested the classification of eye diseases. significant, labelled datasets
system achieve 97% accuracy with modified efficientNetB3 are difficult to come by since doing so requires highly skilled
model and surpasses current approaches for categorizing eye ophthalmologists to annotate a significant number of images
diseases. This approach may help doctors diagnose and treat eye [7]. Furthermore, differentiating between various eye disease
conditions earlier, leading to better patient outcomes. stages and subtypes can be challenging and calls for careful
feature extraction.
Keywords— Eye Diseases; Deep learning model; Colour
Fundus images; EfficientNetB3 model. By applying transfer learning strategies and utilizing
already-built, pre-trained CNN models, this research seeks to
I. INTRODUCTION address these difficulties [8]. With the help of transfer
Deep learning algorithms have emerged as powerful learning, our approach can achieve excellent classification
instruments for precise and effective diagnosis in medical accuracy even with little labelled data by taking advantage of
imaging, building on the tremendous advances made in this the knowledge accumulated by CNN models trained on
sector in recent years. Eye illnesses are one of the many massive image datasets. Additionally, to improve the
medical conditions that could benefit from these performance of the deep learning model, we will research
developments and are a major worldwide health issue. various architectures and optimization techniques [9-10].
Particularly common eye conditions that impact millions of
people globally are cataract, diabetic retinopathy, and The approach used in this work, including dataset
collection and pre-processing, model architecture design,
glaucoma [1-2]. It has the potential to revolutionize patient
treatment, lower healthcare costs, and avert irreparable training, and assessment measures, will be presented in the
parts that follow. In addition, we'll talk about the outcomes
eyesight loss by creating a reliable and automated method for
their early diagnosis and classification. and assess how well our model performs in comparison to
other approaches. Finally, we will discuss the results of our
Prior studies on the diagnosis of eye diseases have mostly research, their possible applications, and future research
used conventional methods of image processing and manual directions.
interpretation by ophthalmologists. However, these methods
frequently have drawbacks including subjectivity, laborious
analysis, and inter-observer variability. On the other hand,
deep learning algorithms have showed enormous potential in
automating the diagnosis procedure, enabling quick and
precise classification of eye illnesses from medical
photographs.

979-8-3503-3828-7/23/$31.00 ©2023 IEEE

632
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

a) Cataract b) diabetic retinopathy c) glaucoma d) normal ophthalmologists to diagnose, especially in their early stages.
Fig. 1. Types of Eye Diseases For this reason, the researchers in this study were able to
isolate diagnostically useful biological markers from the
In summary, this study uses deep learning approaches to retina, blood vessels, optic cup, and optic disc.
meet the urgent demand for precise and effective classification
of eye diseases. Our goal is to create a model that can III. PROPOSED MODEL
automatically diagnose cataract, diabetic retinopathy, A. Data Preparation
glaucoma, and a normal category, assisting medical
professionals in early detection and intervention. The results In this stage, collects and normalises data for training
of this study have the potential to have a major influence on purposes. This model was trained using the Eye disease
patient care, lessen the burden of eye illnesses, and enhance dataset 2022 [15]. The data was obtained via Kaggle. In the
outcomes for visual health. dataset, there was a significant class imbalance. To avoid
model underperformance owing to class imbalance, the
II. RELATED WORK dataset was rebalanced and normalised to improve the
performance of the deep learning model.
Jiang et al. [11] identified the dataset as having referable or
non-referable diabetic retinopathy using three pre-trained B. Model Selection
CNNs (Inception-v3, ResNet152, and Inception-ResNet-v2). It is essential to select the right deep learning model to
Before CNN training, Adaboost was used to resize and classify eye diseases accurately and effectively. We will use
improve pictures, as well as merge models. The Adam the EfficientNetB3 [16] model in this study as our model of
optimizer achieved an AUC of 0.946 and a precision of choice. Modern deep learning architecture EfficientNetB3 is
88.21%. renowned for its effectiveness and superior performance in
Zago et al. [12] integrated the capabilities of two image classification applications.
convolutional neural networks (CNNs) to identify diabetic The foundation of Efficient Net models is a scalable
and non-diabetic retinopathy based on the likelihood of red compound coefficient that grows the model's depth, width,
lesion patches. After being trained on DIARETDB1, the and resolution equally. Due to its scalability, EfficientNetB3
model was tested on IDRiD, Messidor, Messidor-2, DDR, may successfully strike a balance between model size and
DIARETDB0, and Kaggle. On Messidor, the model's performance, making it appropriate for applications with
sensitivity was 0.94 and its AUC was 0.91 2. limited resources. The depth of the model allows it to
Gupta and Chhikara [13] investigated DR detection strategies recognise complicated characteristics, the width gives it the
such as Adaboost, Random Forest, SVM, and others, ability to recognise a variety of patterns, and the resolution
gradually demonstrating the gap that these standard aids in the recognition of smaller details in the images[17,18].
techniques provide in learning new disease-related features. Additionally, EfficientNetB3 is pre-trained on a sizable
The fundus image quality is being compared here. dataset like ImageNet, which offers a useful foundation of
Alyoubi et al. [14] examined 33 papers on deep learning for understanding about different visual ideas. Pre-training on
DR classification and stressed the need for continual model these datasets enables the model to pick up general visual
development due to the global rise in diabetes. Data representations that can then be modified for the particular
augmentation in model training reduces overfitting, the purpose of classifying eye diseases. The pre-trained weights
authors noted. of EfficientNetB3 will be a great place to start because they
To identify multi-class fundus images for an Ocular Disease are already familiar with the low-level edges, textures, and
Intelligent Recognition (ODIR) dataset featuring alterations features of images that are important for our classification
in the optic disc, macula, and blood vessels, Neha et al. [11] challenge[19].
developed four CNN models with different optimisers. When
C. Transfer Learning
compared to other fundus image classification models,
VGG16 with SGD optimisers performed best. We used transfer learning to properly utilise the pre-
To identify pathologies of the eye in fundus images, Xiong et trained EfficientNetB3 model. The process of using a model's
al. [12] employed a deep learning model with a mixture loss knowledge gained from one task and applying it to another is
known as transfer learning. In our example, we used the pre-
function. Classification of a dataset including eye diseases
trained EfficientNetB3 model's expertise to our objective of
was enhanced using a deep learning model with a loss
classifying eye diseases. This model has been trained on a
function and loss robustness training on the focus loss. sizable dataset.
Kai et al. [13] used CNN models to segregate retinal
vasculature from CFP images. CNN models with loss We modified EfficientNetB3 model by adding the batch
functions generate probability maps. Feature maps were used normalization stage, 3 dense layer and dropout layer as shown
to build the CNN model. in Fig. 2.
Clement et al. [14] suggested a convolutional structure and Using dataset to fine-tune the pre-trained
supervised learning system for the Messidor dataset. Red and EfficientNetB3 model, we enable the model to learn
bright lesions are detected and segmented concurrently. The characteristics particular to a certain disease while taking
system generates slices with red and bright lesions and
validates them pixel by pixel. Its AUC was 83.9%.

The progression of eye illness is a common topic of study.

Many different systems for categorizing eye diseases were
developed in this study. Eye illnesses are difficult for

633
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

Fig 2. System Generated Modified EfficientNetB3 IV. DATASET

advantage of the initial knowledge gained during pre-training. The dataset consists of 4214 total colour fundus images of
To make precise predictions for the current classification task, eye. The dataset utilized in this research paper consists of three
the model's parameters are fine-tuned to match the sections which are train, tests and validation which are further
characteristics of the dataset on eye diseases as shown in categorized into four distinct classes: cataract, glaucoma,
figure 3 . Transfer learning allows us to take advantage of diabetic retinopathy, and normal consists of colour fundus
EfficientNetB3's powerful feature extraction skills, greatly images of eye. We divided the dataset and considered 3370
lowering the amount of training time required on small dataset images for training (80%) , 422 images for testing (10%) and
of eye diseases. By using this method, computational time and 422 images for validation phase(10%). Table 1 show the
resources are reduced without sacrificing the model's ability distribution of the Kaggle dataset.
to reliably categorise cataract, diabetic retinopathy, glaucoma,
and healthy eyes[20]. TABLE I. SPLITTING OF CFI OF EYE DISEASES.

Fig 3. System Generated Modified EfficientNetB3 Classes Training Testing Validation

80% 10% 10%
Cataract 830 104 104
Diabetic_retinopathy 878 110 110
Glaucoma 805 101 101
Normal 857 107 107

V. RESULT
We used the Google colab environment and keras library
to perform the experiment.
The model is trained for epochs with EarlyStopping
callback. An early termination technique is used to regulate
the training length and stop pointless computations. Utilising
the EarlyStopping callback, training loss value is monitored.
The training process is terminated if the loss does not decrease
for the specified number of epochs, as determined by the
patience parameter (in this case, 15 epochs). This method

634
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

avoids overtraining and aids in determining the ideal point of

convergence. The training, validation and testing accuracy
obtained in evaluation through model training. Model
accurately classified the 409 images out of 422 images; hence
we achieved the 96.91% accuracy as shown in figure 4 &5 and
confusion matrix as shown in figure 6.

Fig 4. Training and Validation Accurac Fig 6. Confusion Matrix

A. Data Augmentation
Training CNN models requires enormous datasets; the
more data there are, the better the model performs.
Unfortunately, there are few medical images available, and
privacy considerations prevent their collection from hospitals.
Because of this Using a CFP image from a collection of eye
diseases without enough medical images to train CNN models
is a challenge, as this often leads to overfitting. During the
training process, we used data augmentation to solve this
problem. This method includes many methods for artificially
increasing images from the same dataset. Images can be
increased by CLAHE image enhancement, Vertical and
horizontal flipping them, rotating and Grid distortion at
various angles, relocating them, and so forth[. TABLE II.
describes the training dataset before and after the image
increase.

TABLE II. AUGMENTTED IMAGES OF THE EYE

DISEASES DATASET IN TRAINING PHASE

Classes Cataract Diabetic Glaucoma Normal

retinopathy
Before_Aug 830 878 805 857
After_Aug 2830 2878 2805 2857

After the data augmentation, we achieved the 97.5%

accuracy.

Fig 5. Training and Validation Loss

635
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

B. Deployment
The trained model is deployed in an android app. So, a
facility is provided in the app to predict the results based on
the trained model which is embedded in the app. The app
allows users to upload images of eyes with suspected diseases,
and the app will classify the image into different categories of
eye diseases as shown in figure 7.8.9 and 10.

Fig 9. User chooses an image from here album

Fig 7. This is the first screen that user encounters which giving two options:
i) Take Picture ii) Launch Gallery

Fig 10. Model Classified the Eye diseases

CONCLUSION
Fig 8. User clicks on Launch Gallery Option, then all those apps are shown In this study, we successfully explored the use of deep
from where he/she can select photo learning, specifically the Convolutional Neural Network
(CNN) architecture, for the classification of images of eye
diseases. The foundation of our classification model was built

636
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

using the modified EfficientNetB3 model, which is renowned 10. Thanki, R. “A deep neural network and machine learning
for its effectiveness and accuracy in image recognition approach for retinal fundus image classification’. Healthc.
applications. A suitable platform for quick testing was created Anal. Vol. 3,2023.
by integrating the trained model into an Android application 11. Jiang, H., Yang, K., Gao, M., Zhang, D., Ma, H. Qian W
that allowed users to upload fundus image for classification. “An interpretable ensemble deep learning model for diabetic
This research paper concludes by highlighting the effective retinopathy disease classification”, in 41st Annual
application of a deep learning-based strategy for eye diseases International conference of the IEEE engineering in medicine
image classification. The implementation of the Android and biology society (EMBC), pp. 2045–2048,2019.
application, transfer learning, and the use of the
12. Zago, GT., Andreao. RV., Dorizz,i B. and Teatini Salles
EfficientNetB3 model all work together to promote automated
EO‘Diabetic retinopathy detection using red lesion
eye disease diagnosis. The accomplishments and learnings
from this study lay the groundwork for advancements in localization and convolutional neural networks’, Computers
computer-aided diagnosis and provide a foundation for in Biology and Medicine,2020.
additional research into increasing the accuracy, efficiency, 13. Gupta A. and R. Chhikara “Diabetic retinopathy:
and accessibility of eye disease classification systems. Present and past,” Proc. Comput. Sci., vol. 132, pp. 1432–
1440, Jan. 2018.
REFERENCES 14. Alyoubi W. L., W. M. Shalash, and M. F. Abulkhair,
1. Bourne, R.R.; Stevens, G.A.; White, R.A.; Smith, J.L. ‘Diabetic retinopathy detection through deep learning
Flaxman, S.R.; Price, H.; Taylor, H.R.” Causes of vision loss techniques: A review,’ Informat. Med. Unlocked, vol. 20, Jan.
worldwide, 1990–2010: A systematic analysis” Lancet Glob. 2020, Art. no. 100377.
Health Vol.1,2013 e339–e349. 15. Eye disease dataset:
2. Ahmed, M., Ansari, M. D., Singh, N., Gunjan, V. K., BV, Available on :
S. K., & Khan, M. (2022). Rating-based recommender system https://www.kaggle.com/datasets/kondwani/eye-disease-
based on textual reviews using iot smart devices. Mobile dataset, Accessed on : 204/04/2023
Information Systems, 2022. 16. A. Batool and Y. -C. Byun, "Lightweight EfficientNetB3
3. Li, T.; Bo, W.; Hu, C.; Kang, H.; Liu, H.; Wang, K.; Fu, H Model Based on Depthwise Separable Convolutions for
‘Applications of deep learning in fundus images’, A review. Enhancing Classification of Leukemia White Blood Cell
Med. Image Anal. Vol.69,2021. Images," in IEEE Access, vol. 11, pp. 37203-37215, 2023,
4. Gaddam, D. K. R., Ansari, M. D., Vuppala, S., Gunjan, V. doi: 10.1109/ACCESS.2023.3266511.
K., & Sati, M. M. (2022). A performance comparison of 17. Luis Perez and Jason Wang,The Effectiveness of Data
optimization algorithms on a generated dataset. Augmentation in Image Classification using Deep Learning,
In ICDSMLA 2020: Proceedings of the 2nd International arXiv preprint arXiv:1712.0462, 2017. Link:
Conference on Data Science, Machine Learning and https://arxiv.org/abs/1712.04621
Applications (pp. 1407-1415). Springer Singapore. 18. Fernandes, J.B., Narayan, V., Sammilitha, P.K.,
5. Orfao, J.; van der Haar, D.“A Comparison of Computer Koundinya, P.S., Krishna, R.R. (2022). Blockchain-Based
Vision Methods for the Combined Detection of Glaucoma, Privacy Securing G-Cloud Framework for E-Healthcare
Diabetic Retinopathy and Cataracts’, In Lecture Notes in Service. In: Garcia Diaz, V., Rincón Aponte, G.J. (eds)
Computer Science (including Subseries Lecture Notes in Confidential Computing. Advanced Technologies and
Artificial Intelligence and Lecture Notes in Bioinformatics); Societal Change. Springer, Singapore.
Springer: Cham, Switzerland; 2021, Volume 12722, pp. 30– https://doi.org/10.1007/978-981-19-3045-4_8
42. 19. A Suorong Yang, Xiaolin Hu, Zhiqiang Shen, and Jie
6. Siddiquee, K. N. E. A., Islam, M. S., Singh, N., Gunjan, V. Yang,Image “Data Augmentation for Deep Learning:”. arXiv
K., Yong, W. H., Huda, M. N., & Naik, D. B. (2022). preprint arXiv:2204.08610 2022. Link:
Development of algorithms for an iot-based smart agriculture https://arxiv.org/abs/2204.08610
monitoring system. Wireless Communications and Mobile
20. Nirmala, T., Kumar, S.A., Rithvik Rao, P., Raviteja
Computing, 2022, 1-16.
Reddy, P., Poojitha, T. (2022). Disaster Analysis on
7. Prananda, A.R.; Frannita, E.L.; Hutami, A.H.T.; Maarif,
Government Data. In: Garcia Diaz, V., Rincón Aponte, G.J.
M.R.; Fitriyani, N.L.; Syafrudin, M. “Retinal Nerve Fiber
(eds) Confidential Computing. Advanced Technologies and
Layer Analysis Using Deep Learning to Improve Glaucoma
Societal Change. Springer, Singapore.
Detection in Eye Disease Assessment”. Appl. Sci. Vol.13, No.
https://doi.org/10.1007/978-981-19-3045-4_3
37,2023.
8. Lakshmanna, K., Shaik, F., Gunjan, V. K., Singh, N.,
Kumar, G., & Shafi, R. M. (2022). Perimeter degree
technique for the reduction of routing congestion during
placement in physical design of VLSI
circuits. Complexity, 2022, 1-11.
9. Pahuja, R.; Sisodia, U.; Tiwari, A.; Sharma, S.;Nagrath, P.
“A Dynamic Approach of Eye Disease Classification Using
Deep Learning and Machine Learning Model”, Lect. Notes
Data Eng. Commun. Technol. Vol. 90, pp.719–736,2022.

637
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA)

638
Authorized licensed use limited to: National Changhua Univ. of Education. Downloaded on November 26,2024 at 07:47:00 UTC from IEEE Xplore. Restrictions apply.