Received 3 April 2023, accepted 25 May 2023, date of publication 1 June 2023, date of current version 9 June 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3281859

Multi-Class Retinal Diseases Detection Using

Deep CNN With Minimal Memory Consumption
ASIF NAWAZ 1 , TARIQ ALI 1 , GHULAM MUSTAFA 1,

MUHAMMAD BABAR 2 , (Member, IEEE),

AND BASIT QURESHI 3 , (Senior Member, IEEE)
1 University Institute of Information Technology (UIIT), PMAS-Arid Agriculture University, Rawalpindi 46000, Pakistan
2 Robotics and Internet of Things Laboratory, Prince Sultan University (PSU), Riyadh 11586, Saudi Arabia
3 College of Computer and Information Sciences, Prince Sultan University (PSU), Riyadh 11586, Saudi Arabia
Corresponding author: Asif Nawaz ([email protected])
This work was supported by Prince Sultan University, Riyadh, Saudi Arabia, for the payment of Article Processing Charges for this
research work under Grant SEED-CCIS-2023-121.

ABSTRACT Machine Learning (ML) such as Artificial Neural Network (ANN), Deep learning, Recurrent
Neural Networks (RNN), Alex Net, and ResNet can be considered as a broad research direction in the
identification and classification of critical diseases. CNN and its particular variant, usually named U-Net
Segmentation, has made a revolutionary advancement in the classification of medical diseases, specifically
retinal diseases. However, because of the feature extraction complexity, U-Net has a significant flaw in
high memory and CPU consumption while moving the whole feature map to the corresponding decoder.
Furthermore, it can be concatenated to the unsampled decoder feature map avoids reusing pooling indices.
In this research work, a convolutional neural network (CNN) model is proposed for multi-class classification
problems with the efficient use of memory consumption. The proposed model has been evaluated on a
standard benchmark dataset of Eye Net, having 32 classes of retinal diseases. From experimental evaluation,
it has been concluded that the proposed model performs better regarding memory management and accuracy.
The overall comparison has been performed based on precision, recall, and accuracy with different numbers
of epochs and time consumption by each step. The proposed technique achieved an accuracy of 95% on the
Eye-net dataset.

INDEX TERMS Classification, CNN, deep learning, EyeNet, retina, U-Net.

I. INTRODUCTION due to AMD. According to recent research, in the United

Retinal diseases are spreading widely among humans of all States (US), this abnormality is found in about 35% of adults
ages. The retina contains a layer of optic nerve tissue called in the age group of 80 [3]. Identifying retinal diseases is the
photosensitive in the human eye. This layer transforms the most challenging task, as accurate diagnosis needs a highly
light focused by the lens into brain impulses. Macula, posi- experienced ophthalmologist due to the diversity of retinal
tioned in the retina’s middle, performs the sensing process. diseases. Similarly, with computer-aided diagnostic systems
Information acquired by the macula is processed by the retina (CAD), retinal diseases can easily be identified and treated at
and transferred to the brain for visual recognition through the early stages [4].
optic nerve [1]. Different types of diseases can cause abnor- Technology advancements have immense benefits in
mality in perception, such as age-related macular degener- almost every field of life, especially in the medical domain.
ation (AMD), optic disc drusen, Rothspot diabetic macular Several approaches and models have been presented to
edema (DME) [2], etc. In most of the developed countries, improve the efficacy and quality of medical solutions. A sig-
people belonging to the age group of 50 to 60 are losing vision nificant improvement has been observed in the social health
system with the advancement in Automatic Disease Detection
The associate editor coordinating the review of this manuscript and (ADD) [5]. Furthermore, an ADD application, namely retinal
approving it for publication was Senthil Kumar . symptom analysis, provides a unique opportunity to improve
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
56170 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

eye care globally [6]. Recently, many state-of-the-art ML tomography (OCT) for the classification of retinal diseases.
and Deep Learning (DL) models have been proposed for Authors [14] proposed the OCT-NET model, which uses
the classification, segmentation, and identification of reti- a Deep learning approach to classify diabetic-based retinal
nal diseases. We observe that data collection and labeling diseases. The model was evaluated on two public datasets
are significant challenges in the implementation of ADDs, incorporating classes, A2A SD-OCT and SERI-CUHK. Their
as presented by authors in [7] and [8], due to the development proposed model can classify these mentioned diseases with
of several machine learning (ML) and deep learning (DL) significant accuracy. In [15], Mahendran et al. used machine
models, including Recurrent Neural Network (RNN), Convo- learning algorithms to analyze retinal diseases. The proposed
lution Neural Network (CNN), Alex Net ResNet and VGN. methodology uses a decision tree classifier to classify images
These have enabled researchers and physicians to detect and as normal or disease-affected images. Bilateral filtering meth-
categorize such vital disorders [9] readily. An ML-based ods are used to reduce the noise in the dataset images. Next,
Hybrid technique is presented for the classification of retinal Otsu’s segmentation was used for the segmentation of the
diseases automatically. Researchers in [10] have proposed to macula region, and then regional structures were forwarded
use U-Net segmentation for image pre-processing; they have to the classifier that, results in 92% accuracy.
also used a Support Vector Machine (SVM) classifier for the Authors in [16] compare four algorithms for machine
classification. The proposed technique achieved a diagnostic learning (SVM, KNN, C5.0 and random forest) for glau-
accuracy of 89.3%. Yang et al. also provided the first labeled coma prediction [16]. Maximum entropy transformation on
EyeNet dataset containing 32 retinal diseases. It was noted by retinal fundus pictures from an online dataset was used
authors in [10] that the U-Net has a significant flaw of high to identify age-related eye disorders early [17]. A flower
memory consumption in moving the whole feature map to the pollination optimization-configured convolution neural net-
corresponding decoder. Deep learning plays a vitol role in the work (CNN) extracted information from processed photos
classification of images [11], [12], [13]. (FPOA). FPOA adjusted CNN training hyperparameters. The
This research proposes a CNN model based on deep learn- network’s efficiency and accuracy improved. Four pre-trained
ing for classifying multi-class eye disease detection. The CNN algorithms diagnose seven retinal defects in pictures
proposed model has been evaluated on EyeNet Dataset. The with and without illnesses [18]. Bayesian optimization selects
EyeNet dataset includes 32 folders, each containing related suitable hyperparameter values, and image augmentation
images for specific. 70% has been used for training and the improves model applicability. This study evaluates the sug-
rest for validation. From experimental evaluation, it has been gested models. DenseNet201’s classification accuracy on the
observed that the proposed model achieved 95% of accuracy. Retinal OCT Imaging dataset is 99.9%, compared to prior
The deep learning-based CNN model has been applied methods that can only identify a few retinal diseases. The
for retinal-based crucial diseases to boost the conventional authors build a multiclass model-based DED classification
diagnostic method. This is the primary contribution of this system [19]. Multiple Diabetic Eye Disease (DED) diagnoses
study. The key contributions of the paper are as follows. using retinal fundus pictures are essential clinical research.
• A deep learning-based CNN model has been uti- Ophthalmologist-annotated retinal fundus photographs were
lized to strengthen the traditional diagnosis process for assessed.
retinal-based crucial disease. A segmentation approach employing the ensemble classi-
• The proposed CNN model produces better outcomes fier for subsequent classification is presented for identify-
while consuming low memory than standard state-of-art ing and categorizing Oppositional defiant disorder (ODD)
techniques. illnesses. The classification between normal and abnormal
• Experimental evaluation reveals that the performance of blood vessels is observed with an accuracy of 81% [24]. The
the proposed model on the multi-class EyeNet dataset transfer learning technique is used to reduce training time
produces higher accuracy. and resource consumption. A fine-tuned version of VGG-
19 is proposed for feature extraction and classification of a
The remainder of the paper is structured as follows: Section II retinal database in [25]. A CNN based fine-tuned Inception-
presents the related works. In Section III, we offer the pro- Resnet-v2 model is proposed that is implemented on publicly
posed architecture along with a detailed description of the available data called SERI [26] with 100% classification
dataset used. Section IV details the result of the experimental accuracy. A CNN-based ResNet model for the classification
evaluation, including the performance of the given CNN of diabetic retinopathy (DR) severity level has been presented
model. Section V contains the analysis and the discussion. by Zhang et al. Image cropping techniques such as oversam-
Finally, section VI concludes the research work with future pling and cost-sensitive learning with the aid of the Kaggle
directions. dataset are used to handle the issue of imbalance classes.
By introducing extra layers, the increased regularization in
II. LITERATURE REVIEW training and accuracy achieves datside-outputis81% [27].
In the current era, DL and ML models are widely used According to the published literature, it is found
in image processing and classification. Similarly, many that the maximum state-of-the-art models are imple-
models have been introduced to include optical coherence mented and evaluated for the datasets having 2 or
VOLUME 11, 2023 56171
 A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

TABLE 1. Dataset description.

FIGURE 1. Workflow diagram of the proposed model.

4 classes [20], [21], [22], [23]. Moreover, only two of

these are implemented for a dataset having 15 classes and
32 classes [19], providing very low accuracy of 80.93% and
89.3%, respectively. Most of these focused on preprocessing
techniques such as segmentation and transfer learning. Many
of them these work with pre-trained CNN weights and pre-
processed images instead of taking raw images without any
preprocessing. However, all the models presented achieved
high recognition rates greater than 85%.
CNN and its derivative, sometimes called U-Net Segmen- used. The whole model implementation is performed on this
tation, have achieved a breakthrough leap in categorizing dataset. The photos inside the following dataset are catego-
medical disorders, particularly retinal diseases. Due to the rized with the appropriate labels. The dataset was extracted
intricacy of feature extraction, U-Net has a main memory and from the source on GitHub [30].
CPU usage fault when transporting the whole feature map to
the associated decoder. In addition, concatenating it with the B. PROPOSED MODEL
unsampled decoder feature map prevents the reuse of pooling Deep learning is the most widespread technology in the
indices. In this study, a convolutional neural network (CNN) present day. Many processing layers inside the framework
model with optimal memory consumption is developed for of DL enable the computational models to learn data pat-
multi-class classification problems. terns with multiple levels of abstraction. These models
primarily identify voice, objects, visual objects, and sev-
III. METHODOLGY eral other discovery domains. DL technology is inspired
This section also discusses the dataset used and the proposed by the deep structure of the human brain [31]. CNN is
CNN architecture. In Figure 1, the proposed methodology has the most potent and effective DL model [32]. Although
been demonstrated stepwise. researchers have expanded its applicability to other dis-
ciplines, CNN is primarily used in the medical indus-
A. DATASET try [33]. A method for developing a DL-based model
The EyeNet dataset provided by Yang et al. [10] has been for retinal illness categorization in which the DL model
used. They offered a labeled collection of 32 sorts of clinical sequentially processes retinal pictures. At first, the low-level
data. Historically, most model implementations were per- features are extracted then the middle and high-level
formed using STARE or Drive datasets, as these datasets refined features are extracted for classification. These final
are identified with four or two classes at most. On the other extracted features are passed to the trainable classifier for
hand, Table 1 Eye-Net dataset with 32 retinal disease types is classification.

56172 VOLUME 11, 2023

A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

FIGURE 2. The proposed model architecture.

1) CONVOLUTIONAL NEURAL NETWORK MODEL The graphical illustration of the model is given in Figure 2.
CNN contains hidden layers; these layers perform convo- In the proposed model, feature extraction is done in three
lution, a sub-sampling technique to extract features of data steps. The first level includes low-level features of images,
from a low level to a high level. In the proposed model, ten and then these extracted features are passed to the mid-level
convolution layers are used. In Figure 2, the arrangement of for further refinement. The high level consists of the detailed
layers is shown. On the abstract, retina images are input to features which basically involved in the training process and
the CNN model, which gives label prediction for the normal then used for classification.
or affected eyes. The presented model minimized the number
of layers compared to the traditional models. Models such
as AlexNet are implemented with 25 layers, Densnet201 C. DATA AUGMENTATION
with 201 layers, Inception3 with 48 layers, and ResNet- Data augmentation technique is used for enough data to be
10 with 101 layers. In addition, these pre-trained networks available for the training. Data augmentation is also used
are usually implemented with transfer learning techniques in to avoid overfitting. The parameters we have used in aug-
the medical field for classification. A network with fewer mentation are rescaling, zooming, and flipping [28], [29].
layers is presented so the training time can be reduced. Batch We augmented data using various random transformations
normalization layers are used so that higher learning rates so the model could not face the same image again. This
can be achieved and used, which improves the training speed. technique helps the model from overfitting as well as better
Detailed information about the proposed CNN model is given generalization. In Keras, augmentation is done by Image-
in Table 2. Data Generator. Data augmentation in our proposed model

VOLUME 11, 2023 56173

 A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

FIGURE 3. CNN model.

image. This also enhances the visibility of the essential fea-

tures to diagnose the disease better. Figure 4 depicts another
augmentation in which the original image is converted into a
grayscale image. The grayscale technique changes the colors
of the original image into grayed contrast. In Figure 5, the
original image is flipped. In this technique, the direction of
the original image is changed. Figure 6 depicts the brighten-
ing step of the augmentation technique. Figure 7 displays the
FIGURE 4. Saturated. zoom step. The selection of augmentation steps or processes
may vary in every research.

includes six basic steps. Each step transforms the image to D. ADAPTIVE MOMENT ESTIMATION
the new level. The optimization algorithms, based on stochastic gradi-
In Figure 3, data augmentation steps saturation has been ent, have a significant impact and usage in science and
shown. Data augmentation is beneficial for enhancing the per- engineering-related fields [35]. Adaptive moment estimation,
formance and results of machine learning models by adding also called Adam, is one of the optimization algorithms [35].
additional and distinct training samples. If the dataset used The Adam optimization technique is a stochastic gradient
in a machine learning model is extensive and comprehensive, descent predicated on the adaptive estimate of first- and
the model will perform better and more precisely. The orig- second-order moments. The approach is easy to develop, has
inal image is saturated, which increases the opacity of the low memory needs, is computationally efficient, is invariant

56174 VOLUME 11, 2023

A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

FIGURE 8. Zoom.

FIGURE 5. Gray scaled.

non-stationary goals and issues that include highly noisy
TABLE 2. CNN layers distributions. or sparse gradients. The hyper-parameters may be inter-
preted straightforwardly and usually require very little fine-
tuning [35]. It can be used to update the network weights iter-
atively based on training data instead of the classical stochas-
tic gradient descent procedure. It is specifically designed to
work with complex optimization problems. In this research,
Adam is used because of its low memory consumption and
efficiency in computation. The implementation of Adam is
done by using built-in Kara’s library.

IV. PERFORMACE EVALUATION

To determine how well the suggested model performs, the
output of the categorization model has been contrasted with
the labels that correspond to it. The labeling of images is done
by the ophthalmologist [11]. We evaluate the model on the
bases of these labels. In Figure 9, the confusion matrix is plot-
ted to analyze the performance of CNN. Figure 10 depicts the
results of the classification report generated by implementing
the proposed model. The precision is also known as a positive
predictive value, and recall is a true positive rate or actual
values (presented in equation 1 and equation 2)
TP
precision = (1)
TP + FP
TP
Recall = (2)
TP + FN
The dataset contains 32 retinal diseases; therefore, the per-
formance measures are computed against each class. All the
FIGURE 6. Flipped. classes are classified correctly with 100% precision and 99%
recall. The class AMN has 91% precision and 100% recall.
On the other hand, the minimum precision in the classifica-
tion report is 77% of the class Adult Fovemacular Dystrophy
Appearance. Similarly, the class RH has a minimum recall
value of 70% of results analyzed at 10 and 15 epochs. In both
cases model performed outclass.

V. RESULTS AND DISCUSSIONS

The dataset is divided into two subsets such as training and
FIGURE 7. Brighten. validation. The implementation is done in python by using
Keras. The implementation is executed on Intel(R) Core
(TM) i5-7200CPU at 2.70GHz. The Google GPU is used for
to diagonal re-scaling of the gradients, and is a good fit the training. We tested the model with different optimizers;
for applications that involve a significant amount of data among them, Adam gave the best results. The learning rate
or parameters. The approach is also suitable for use with used for the optimizer is 0.001. All the other hyperparameters

VOLUME 11, 2023 56175

 A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

FIGURE 9. Confusion matrix.

FIGURE 10. Classification repor.

are selected after complete attestation. The size of the batch 10 epochs, and epochs per step are 500 with the validation
is 10. We trained data on different epochs and achieved steps two learning rate of the optimizer is 0.001. The valida-
95% validation accuracy. At first, we trained the dataset with tion accuracy is 0.95, and the validation loss is 0.0279. Then

56176 VOLUME 11, 2023

A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

TABLE 3. Comparison of the proposed model with existing models.

FIGURE 13. Validation and training loss comparison at 15 epochs.

FIGURE 11. Model accuracy by implementing 10 Epochs.

the passage of time, validation accuracy fluctuated in steps 7,

8, and 9. The training accuracy at 7, 8, 9, and 10 is 0.9041,
0.9184, 0.9184, and 0.9450, respectively. Validation accu-
racy increases with the model’s training and reaches 0.9500.
In Figure 12, model accuracy has been shown at 15 epochs
with 400 validation steps per epoch. The different epochs
are executed to analyze the performance and validity of the
model entirely. At first, test accuracy is 0.4068, and valida-
tion is 0.0000e+0. Then with each step, validation accuracy
increased. The fluctuation between the different steps has
also been observed from the plotted results. However, the
FIGURE 12. Model accuracy at 15 Epochs. model’s overall performance is considerable compared to the
traditional models with greater layers. In both cases, at 10 and
15 iterations, 0.9500 accuracy has been archived.
we trained the model with 15 epochs with 400 epochs per step In Figure 13, a Comparison between validation and train-
with the same validation steps and optimizer, and again the ing losses has been shown. The training loss must be greater
model validation accuracy is 95%. We also trained the model than the validation loss. If validation loss is greater, the model
with different optimizers, but Adam optimizer gave the best is overfilled and must be fixed. To avoid model overfitting,
results compared to other optimizers. a dropout layer also has been used. If the validation loss and
Figure 11 depicts the model accuracy with 10 epochs. Both training loss are equal, then it will result in model under-
training and testing accuracies are compared by executing the fitting. The number of epochs is 15, and the validation per
model for 10 epochs with 500 per epoch’s steps. The start step is 400. Training loss at the first epoch is 1.9829 at 42s
loss is 1.8199, the training accuracy is 0.4673, the validation 105ms/step time. The model starts learning best weights at
loss is 2.0881, and the validation accuracy is 0.2500. With each epoch, and at last, at 15 epochs, the Training loss is

VOLUME 11, 2023 56177

 A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

FIGURE 14. Time consumption on each step.

posed model requires less time and RAM to complete the

training with dynamic learning weights.
Table 3 lists the summarized comparison of the proposed
CNN model with the existing models in the literature. Most
of the existing models work with a smaller number of
classes. However, two existing models also work with 15 and
32 classes of datasets. The accuracy of the models with a
smaller number of classes is high, whereas the accuracy of
existing models with a high number of classes is very low.
The most relevant comparison of the proposed model is with
U_Net segmentation using an SVM classifier with the same
dataset. U_Net has 89.3% accuracy, and the proposed model
achieved 95% accuracy on the same dataset. The proposed
FIGURE 15. Comparisons with state-of-the-art approaches. model for the multiclass problem gained 95% of validation
accuracy. Besides accuracy, the proposed model works with
low memory consumption as well as work even with limited
resources such as hardware limitations.
0.1670, validation loss is 0.0279 at 40s 99ms/step with the This experiment compares the results of different models
validation accuracy of 0.9500. with the proposed work on the common dataset. Such com-
Figure 14 illustrates the time consumption in training at parisons aim to determine which model or algorithm per-
10 epochs with 500 validation per step. At first, the time forms better on a uniform dataset. Since each model uses its
required to complete the 500 steps is 132s 263ms/step. Then own dataset, getting the same dataset for the proposed work
in steps 2, 3, and 4, time decreases, which is 131s 261ms/step, is complicated. However, on public request, two different
129s 258ms/step, and 130s 261ms/step, respectively. From datasets were obtained from the work of Sinanet al. [37]
the results, time fluctuation has been observed in steps 5 and and Neha et al. [38]. Figure 15 shows the comparisons of
6. However, the overall results show that model required the proposed model with the stated approaches in terms of
less training time than the other state-of-the-art models. The Accuracy and F1-score.
training by implementing U-Net segmentation takes almost
20 hours on a high-standard GPU. On the other hand, local VI. CONCLUSION
machines cannot complete the training because of hardware The classification of the different retinal disorders is
limitations (RAM, CPU). Results demonstrate that, the pro- addressed by presenting a CNN model based on deep

56178 VOLUME 11, 2023

A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

learning. EyeNet, a dataset containing 32 various retinal dis- [14] O. Perdomo, H. Rios, F. J. Rodríguez, S. Otálora, F. Meriaudeau, H. Müller,
eases, is the basis for the model’s implementation. The pro- and F. A. González, ‘‘Classification of diabetes-related retinal diseases
using a deep learning approach in optical coherence tomography,’’ Comput.
posed model is trained on different epochs to test the model’s Methods Programs Biomed., vol. 178, pp. 181–189, Sep. 2019.
accuracy. Initially, the model was trained at 10 epochs and [15] G. Mahendran, M. Periyasamy, S. Murugeswari, and N. K. Devi, ‘‘Analysis
achieved 95% validation accuracy; then, at 15 epochs model on retinal diseases using machine learning algorithms,’’ Mater. Today,
Proc., vol. 33, pp. 3102–3107, Jan. 2020.
again achieved 95% validation accuracy with 0.0279 valida- [16] S. J. Kim, K. J. Cho, and S. Oh, ‘‘Development of machine learning
tion loss which varies in both cases. The model’s total perfor- models for diagnosis of glaucoma,’’ PLoS ONE, vol. 12, no. 5, May 2017,
mance is much superior to that of other models considered to Art. no. e0177726.
[17] P. G. Subin and P. Muthukannan, ‘‘Optimized convolution neural network
be state of the art. There is a possibility that the categorization based multiple eye disease detection,’’ Comput. Biol. Med., vol. 146,
of retinal diseases might benefit from the model that has been Jul. 2022, Art. no. 105648.
provided. Regular model updates and retraining using new [18] M. Subramanian, M. S. Kumar, V. E. Sathishkumar, J. Prabhu, A. Karthick,
S. S. Ganesh, and M. A. Meem, ‘‘Diagnosis of retinal diseases based
data will continue to enhance its performance in the future. on Bayesian optimization deep learning network using optical coher-
This will be achieved by leveraging the advancements in deep ence tomography images,’’ Comput. Intell. Neurosci., vol. 2022, pp. 1–15,
learning techniques and the increasing availability of diverse Apr. 2022.
retinal disease datasets. [19] R. Sarki, K. Ahmed, H. Wang, Y. Zhang, and K. Wang, ‘‘Convolutional
neural network for multi-class classification of diabetic eye disease,’’ EAI
Endorsed Trans. Scalable Inf. Syst., vol. 9, no. 4, p. e5, 2022.
CONFLICT OF INTERESTS [20] D. Marín, A. Aquino, M. E. Gegundez-Arias, and J. M. Bravo, ‘‘A new
supervised method for blood vessel segmentation in retinal images by
Authors have no conflict of interests. using gray-level and moment invariants-based features,’’ IEEE Trans. Med.
Imag., vol. 30, no. 1, pp. 146–158, Jan. 2011.
REFERENCES [21] X. You, Q. Peng, Y. Yuan, Y.-M. Cheung, and J. Lei, ‘‘Segmentation
of retinal blood vessels using the radial projection and semi-supervised
[1] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, approach,’’ Pattern Recognit., vol. 44, nos. 10–11, pp. 2314–2324,
and S. Thrun, ‘‘Dermatologist-level classification of skin cancer with deep Oct. 2011.
neural networks,’’ Nature, vol. 542, no. 7639, pp. 115–118, Feb. 2017. [22] G. B. Kande, P. V. Subbaiah, and T. S. Savithri, ‘‘Unsupervised fuzzy based
[2] K. Shankar, A. R. W. Sait, D. Gupta, S. K. Lakshmanaprabu, A. Khanna, vessel segmentation in pathological digital fundus images,’’ J. Med. Syst.,
and H. M. Pandey, ‘‘Automated detection and classification of fundus vol. 34, no. 5, pp. 849–858, Oct. 2010.
diabetic retinopathy images using synergic deep learning model,’’ Pattern [23] M. A. Palomera-Perez, M. E. Martinez-Perez, H. Benitez-Perez, and
Recognit. Lett., vol. 133, pp. 210–216, May 2020. J. L. Ortega-Arjona, ‘‘Parallel multiscale feature extraction and region
[3] R. Arunkumar and P. Karthigaikumar, ‘‘Multi-retinal disease classification growing: Application in retinal blood vessel detection,’’ IEEE Trans. Inf.
by reduced deep learning features,’’ Neural Comput. Appl., vol. 28, no. 2, Technol. Biomed., vol. 14, no. 2, pp. 500–506, Mar. 2010.
pp. 329–334, Feb. 2017. [24] P. Gowsalya and S. Vasanthi, ‘‘Segmentation and classification of features
[4] T. Shanthi and R. S. Sabeenian, ‘‘Modified Alexnet architecture for clas- in retinal images,’’ in Proc. Int. Conf. Commun. Signal Process., Apr. 2014,
sification of diabetic retinopathy images,’’ Comput. Electr. Eng., vol. 76, pp. 1869–1873.
pp. 56–64, Jun. 2019. [25] A. Das, R. Giri, G. Chourasia, and A. A. Bala, ‘‘Classification of retinal
[5] S. Farsiu, S. J. Chiu, R. V. O’Connell, F. A. Folgar, E. Yuan, diseases using transfer learning approach,’’ in Proc. Int. Conf. Commun.
J. A. Izatt, and C. A. Toth, ‘‘Quantitative classification of eyes with Electron. Syst. (ICCES), Jul. 2019, pp. 2080–2084.
and without intermediate age-related macular degeneration using optical [26] R. M. Kamble, G. C. Y. Chan, O. Perdomo, M. Kokare, F. A. González,
coherence tomography,’’ Ophthalmology, vol. 121, no. 1, pp. 162–172, H. Müller, and F. Mériaudeau, ‘‘Automated diabetic macular edema
Jan. 2014. (DME) analysis using fine tuning with Inception-Resnet-v2 on OCT
[6] R. F. Mullins, S. R. Russell, D. H. Anderson, and G. S. Hageman, ‘‘Drusen images,’’ in Proc. IEEE-EMBS Conf. Biomed. Eng. Sci. (IECBES),
associated with aging and age-related macular degeneration contain pro- Dec. 2018, pp. 442–446.
teins common to extracellular deposits associated with atherosclerosis, [27] D. Zhang, W. Bu, and X. Wu, ‘‘Diabetic retinopathy classification
elastosis, amyloidosis, and dense deposit disease,’’ FASEB J., vol. 14, no. 7, using deeply supervised ResNet,’’ in Proc. IEEE SmartWorld, Ubiqui-
pp. 835–846, May 2000. tous Intell. Comput., Adv. Trusted Comput., Scalable Comput. Commun.,
[7] Y. Kanagasingam, A. Bhuiyan, M. D. Abràmoff, R. T. Smith, L. Gold- Cloud Big Data Comput., Internet People Smart City Innov. (Smart-
schmidt, and T. Y. Wong, ‘‘Progress on retinal image analysis for age World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), Aug. 2017, pp. 1–6.
related macular degeneration,’’ Prog. Retinal Eye Res., vol. 38, pp. 20–42, [28] L. Perez and J. Wang, ‘‘The effectiveness of data augmentation in image
Jan. 2014. classification using deep learning,’’ 2017, arXiv:1712.04621.
[8] D. S. Kermany, ‘‘Identifying medical diagnoses and treatable diseases [29] C. Shorten and T. M. Khoshgoftaar, ‘‘A survey on image data augmentation
by image-based deep learning,’’ Cell, vol. 172, no. 5, pp. 1122–1131, for deep learning,’’ J. Big Data, vol. 6, no. 1, p. 60, Dec. 2019.
Feb. 2018. [30] Accessed: Mar. 2, 2022. [Online]. Available: https://github.com/
[9] M. M. M. S. Fathy and M. T. Mahmoudi, ‘‘A classified and comparative huckiyang/EyeNet
study of edge detection algorithms,’’ in Proc. Int. Conf. Inf. Technol., [31] G. E. Hinton, S. Osindero, and Y.-W. Teh, ‘‘A fast learning algorithm for
Coding Comput., Apr. 2002, pp. 117–120. deep belief nets,’’ Neural Comput., vol. 18, no. 7, pp. 1527–1554, Jul. 2006.
[10] C.-H. H. Yang, J.-H. Huang, F. Liu, F.-Y. Chiu, M. Gao, W. Lyu, [32] J. Gu, ‘‘Recent advances in convolutional neural networks,’’ Pattern
M. D. I.-H. Lin, and J. Tegner, ‘‘A novel hybrid machine learning model Recognit., vol. 77, pp. 354–377, May 2018.
for auto-classification of retinal diseases,’’ 2018, arXiv:1806.06423. [33] M. Bakator and D. Radosav, ‘‘Deep learning and medical diagnosis: A
[11] M. B. Jabra, A. Koubaa, B. Benjdira, A. Ammar, and H. Hamam, ‘‘COVID- review of literature,’’ Multimodal Technol. Interact., vol. 2, no. 3, p. 47,
19 diagnosis in chest X-rays using deep learning and majority voting,’’ Aug. 2018.
Appl. Sci., vol. 11, no. 6, p. 2884, Mar. 2021. [34] F. Gao, Z. Yue, J. Wang, J. Sun, E. Yang, and H. Zhou, ‘‘A novel active
[12] S. Guefrechi, M. B. Jabra, A. Ammar, A. Koubaa, and H. Hamam, ‘‘‘Deep semisupervised convolutional neural network algorithm for SAR image
learning based detection of COVID-19 from chest X-ray images,’’ Multi- recognition,’’ Comput. Intell. Neurosci., vol. 2017, pp. 1–8, Oct. 2017.
media Tools Appl., vol. 80, no. 2021, pp. 31803–31820. [35] D. P. Kingma and J. Ba, ‘‘Adam: A method for stochastic optimization,’’
[13] W. Boulila, A. Ammar, B. Benjdira, and A. Koubaa, ‘‘Securing the clas- 2014, arXiv:1412.6980.
sification of COVID-19 in chest X-ray images: A privacy-preserving deep [36] L. Terry, ‘‘An in vivo investigation of choroidal vasculature in age-related
learning approach,’’ in Proc. 2nd Int. Conf. Smart Syst. Emerg. Technol. macular degeneration,’’ Ph.D. dissertation, School Optometry Vis. Sci.,
(SMARTTECH), May 2022, pp. 220–225. Cardiff Univ., Cardiff, Wales, 2017.

VOLUME 11, 2023 56179

 A. Nawaz et al.: Multi-Class Retinal Diseases Detection Using Deep CNN With Minimal Memory Consumption

[37] S. S. M. Sheet, T.-S. Tan, M. A. As’ari, W. H. W. Hitam, and J. S. Y. Sia, GHULAM MUSTAFA received the Ph.D. degree
‘‘Retinal disease identification using upgraded CLAHE filter and trans- in computer science from PMAS Arid Agriculture
fer convolution neural network,’’ ICT Exp., vol. 8, no. 1, pp. 142–150, University, Rawalpindi. He is currently an Assis-
Mar. 2022. tant Professor with UIIT, PMAS Arid Agricul-
[38] N. Sengar, R. C. Joshi, M. K. Dutta, and R. Burget, ‘‘EyeDeep-Net: ture University, Rawalpindi. His research interests
A multi-class diagnosis of retinal diseases using deep neural network,’’ include data mining and knowledge discovery.
Neural Comput. Appl., vol. 35, pp. 10551–10571, Jan. 2023.

MUHAMMAD BABAR (Member, IEEE) received

the Master of Science and Ph.D. degrees in com-
ASIF NAWAZ received the Ph.D. degree puter software engineering from the National
from International Islamic University Islamabad, University of Sciences and Technology (NUST),
in 2019. He is currently an Assistant Professor Islamabad, Pakistan, in 2018. He is currently with
with UIIT, PMAS Arid Agriculture University, Prince Sultan University (PSU), Saudi Arabia. His
Rawalpindi. He is also a potential researcher in research interests include but is not limited to
the field of data mining, machine learning, social big data analytics, machine learning, the Internet
media analysis, and text mining. He has published of Things (IoT), smart city design and planning,
more than 30 research articles in well reputed security, and the Social Web of Things (SWOT).
international journals. He has published his research work in various IEEE, Elsevier, and Springer
journals.

BASIT QURESHI (Senior Member, IEEE)

received the bachelor’s degree in computer science
from Ohio University, Athens, OH, USA, in 2000,
the master’s degree in computer science from
TARIQ ALI received the master’s degree in com- Florida Atlantic University, in 2002, and the Ph.D.
puter science from Muhammad Ali Jinnah Univer- degree in computer science from the University of
sity, Islamabad, Pakistan, in 2009, and the Ph.D. Bradford, U.K., in 2011. He is currently an Asso-
degree in computer science from Abasyn Uni- ciate Professor and the Chair of the Department
versity, Islamabad, in 2019. He is currently an of Computer Science, Prince Sultan University,
Assistant Professor with UIIT, PMAS Arid Agri- Saudi Arabia. He has published over 60 research
culture University. His research interests include works in the area of robotics, the Internet of Things, machine learning, and
document classification, multi label classification, computational intelligence. He is a member of the IEEE Computer Society,
machine learning, and deep learning. the IEEE Communications Society, and the ACM.

56180 VOLUME 11, 2023