Asian Journal of Convergence in Technology Volume VII and Issue II

ISSN NO: 2350-1146 I.F-5.11

Ocular Disease Detection Using Advanced Neural

Network Based Classification Algorithms
Nadim Mahmud Dipu Sifatul Alam Shohan K.M.A Salam
Department of Electrical and Computer Department of Electrical and Computer Department of Electrical and Computer
Engineering Engineering Engineering
North South University North South University North South University
Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh
nadim.dipu@northsouth.edu sifatul.shohan@northsouth.edu kazi.salam@northsouth.edu

Abstract—One of the most challenging tasks for Meng et al. [13] proposed a two-stage process of
ophthalmologists is early screening and diagnosis of ocular utilizing convolutional neural networks (CNN) on fundus
diseases from fundus images. However, manual diagnosis of images in order to perform Optic Disc (OD) localization.
ocular diseases is difficult, time-consuming and it can be prone Automatic ocular disease classification models have been
to errors. That is why a computer-aided automated ocular proposed by He et al. [14] that are based on knowledge
disease detection system is required for the early detection of distillation. This system is built by training and optimizing
various ocular diseases using fundus images. Due to the two deep networks sequentially.
enhanced image classification capabilities of deep learning
algorithms, such a system can finally be realized. In this study, Roy et al. [15] suggested a fully convolutional deep
we present four deep learning-based models for targeted architecture called ReLayNet for segmenting retinal layers
ocular tumor detection. For this study, we trained the cutting- and fluids from Optical Coherence Tomography (OCT)
edge image classification algorithms such as Resnet-34, scans. This technique utilizes an encoder-decoder network
EfficientNet, MobileNetV2, and VGG-16 on the ODIR dataset for semantic segmentation on OCT scans.
consisting of 5000 fundus images that belong to 8 different
classes. Each of these classes represents a different ocular Liefers et al. [16] used a fully convolutional neural
disease. The VGG-16 model achieved an accuracy of 97.23%; network that had dilated convolution filters in order to
the Resnet-34 model reached an accuracy of 90.85%; the implement a pixel-wise classification on Optical Coherence
MobileNetV2 model provided an accuracy of 94.32%, and the Tomography (OCT) scans. The performance of this model
EfficientNet classification model achieved an accuracy of was evaluated on a dataset consisting of 400 OCT scans of
93.82%. All of these models will be instrumental in building a patients who were affected by varying stages of age-related
real-time ocular disease diagnosis system. macular degeneration.
Keywords—Ocular Disease Classification, Color Fundus Lee et al. [17] proposed a CNN-based model that can
Photography, Ocular Disease Detection, Convolutional Neural detect intra-retinal fluid on OCT images. This model was
Networks, EfficientNet, VGG-16, Resnet-34, MobileNetV2, trained on 1,289 OCT scans, and the images segmented by
Transfer Learning the CNN model received a cross-validated Dice score of
0.911.
I. INTRODUCTION
A novel convolutional multi-task architecture was
Various ocular diseases are capable of causing permanent proposed by Playout et al. [18] that takes a supervised
and irreversible damage to the patient’s vision, and in learning approach. This model is trained to perform three
extreme cases, it can even lead to blindness [1-3]. Although tasks simultaneously and those tasks involve segmentation of
effective treatments are available for these ocular diseases, bright lesions, segmentation red lesions, and lesion detection.
these treatment options can only be implemented if the The area under ROC curve of this model was 0.839.
disease is diagnosed as early as possible. Ocular diseases are
primarily diagnosed using color fundus photography or CFP Hu et al. [19] proposed a retinal vessel segmentation
[4]. This technique is utilized in order to record the interior technique that’s implemented using a convolutional neural
surface of the human eye so that various types of possible network and fully connected conditional random fields
ocular diseases can be detected [5]. (CRFs). The accuracy and effectiveness of this model was
evaluated on the color fundus images taken from STARE
Although this method of diagnosis is effective, it’s still [20] and DRIVE [21] datasets.
quite difficult to detect certain ocular diseases using CFP.
Some of the most prevalent ocular diseases, such as Gulshan et al. [22] proposed a deep learning-based
cataracts, myopia, and diabetic retinopathy are difficult to algorithm for automating the process of diabetic macular
diagnose as they show very few initial symptoms. [6] edema and diabetic retinopathy detection. This task was done
Moreover, the process of manually inspecting and detecting using an optimized neural network-based image
ocular diseases is a laborious task, and this process is not that classification model.
accurate [7].
Li et al. [23] proposed a deep learning-based system to
In recent times deep learning-based neural network detect Glaucomatous Optic Neuropathy (GON). This study
models have shown promising results in medical image was also done on color fundus photographs. The researchers
classification and object detection. [8-10] Moreover, that is had trained a classification model that was trained on 8000
why convolutional neural network-based models have been color fundus images. The model achieved a sensitivity score
extensively studied for ocular disease detection [9] [11-12]. of 95.6%, specificity of 92.00%, and an AUC score of 0.986.

This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 International License
91
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

Karri et al. [24] presented an algorithm that can identify The Resnet-34 model that we used in our study was
different retinal pathologies from optical coherence pretrained on a large image classification dataset known as
tomography images (OCT) images. This algorithm was Imagenet [27]. The Imagenet dataset contains 14 million
developed by fine-tuning a pre-trained convolutional neural images that are categorized into 1000 classes.
network called GoogleNet [25]. The dataset used in this
study had four distinct classes that included dry age-related This model is then further trained on the ODIR dataset so
macular degeneration, diabetic macular edema, and no that it can be used for classifying ocular diseases. This
pathology. process of training a pretrained image classifying model on
custom images is called transfer learning.
Although almost all of these studies have shown
promising results, only a few of the existing studies have
addressed the task of classifying multiple ocular diseases
from fundus images. Furthermore, an automated ocular
disease diagnostic tool will require a robust model that has
been thoroughly trained on multiple ocular diseases so that it
can detect diseases from color fundus images.
The models that we have discussed so far are highly
effective at performing specific classification or
segmentation tasks such as segmenting retinal vessels and
classifying a specific ocular disease. However, they cannot
be used as a generalized ocular disease detection system.
Our task was to classify ocular diseases from color
fundus photographs as effectively as possible. Although
various CNN-based classification models have been used for
ocular disease classification before, the latest, state-of-the-art
classification models such as EfficientNet [28] and VGG-16
[29] have not been extensively studied in this regard. These
models have been highly effective at classification tasks
performed on various other medical imaging datasets. That is
why we chose to use these models in order to determine their
performance of on ODIR dataset. This way we can figure out
which model would be ideal for building an autonomous Fig. 1. A look at the fundus images of the ODIR dataset.
ocular disease detection system.
TABLE I. DISTRIBUTION OF THE IMAGES IN THE DATASET
II. DATASET
No. Labels Training Off-Sie On-Sie All
For this study, we have used the Ocular Disease Cases Training Training Cases
Intelligent Recognition (ODIR) dataset. [26] It is one of the Cases Cases
largest publicly available multiclass ocular disease detection 1 N 1,135 161 324 1,620
datasets in the world. This dataset was compiled by 2 D 1,131 162 323 1,616
Shanggong Medical Technology Co, limited by taking 3 G 207 30 58 307
collecting fundus images from different hospitals in China. 4 C 211 32 64 243
The fundus images of this dataset are split into eight different 5 A 171 25 47 295
ocular disease classification categories. These categories 6 H 94 14 30 138
include seven disease classes that are diabetes (D), cataract 7 M 177 23 49 249
(C), glaucoma (G), age-related muscular degeneration (A), 8 O 944 134 268 1,346
myopia (M), hypertension (H), and other
abnormalities/diseases (O). In total, this dataset contains
5000 cases of color fundus photographs (CFPs), and it is split
into training and testing subsets. Roughly 3500 cases are
used for training, and the rest are used for testing. Some
sample images of the ODIR dataset can be observed in figure
1.
The class distribution of the images is illustrated in the
bar chart shown in figure 2. And we can see the details
regarding the image distribution of the dataset in table I.
III. METHODOLOGY

A. Classification Using Resnet-34

Resnet refers to a convolutional neural network
architecture that’s extensively used as a classification model.
Fig. 2. Bar chart representing the distribution of the dataset.
[31]

92
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

This way, the model was able to learn about all of the
necessary pretrained features. After that the model was run
for 50 epochs. We used a training callback function called
“early stopping” that stops the training process if the
validation loss of the model does not decrease for more than
20 epochs.
After that, we unfroze the model’s parameters and then
proceeded to calculate the ideal learning rate. The model was
further training for 50 epochs to ensure our classification
model provided the maximum performance.
B. Classification Using EfficientNet
EfficientNet is one of the most sophisticated models out
there when it comes to custom image classification. It is an
open-source, state-of-the-art CNN-based model that was
developed by Google Brain. In order to create this model, we
used the Keras deep learning framework, and we
implemented it in Google Colab. We used a supervised
learning approach to training the EfficientNet model on the
ODIR dataset.
This model was trained by passing the features of the
training images into the deep neural network, and its task is
to provide the probabilities of the test images belonging to a
particular class. In this case, the class that has the highest
probability according to the model is considered to be the
model’s prediction. The architecture of the EfficientNet
model is illustrated in figure 5.

Fig. 5. EfficientNet Architecture

EfficientNet was developed in order to test how to

effectively scale the overall size of the convolutional neural
networks (CNNs). The comparison of the various scaling
methods used in EfficientNet is shown in figure 6. Just like
Resnet-34, the EfficientNet model is also pretrained and
benchmarked on the ImageNet dataset. That’s why it has a
Fig. 3. Resnet-34 Architecture.
strong understanding of the general features that are required
The overall architecture of the Resnet-34 model is shown to classify the images.
in figure 3. The architecture of the Resnet-34 model might
seem a bit complicated, which is why a more simplified
illustration is shown in figure 4. At first, we loaded the
dataset into a colab notebook for training the Resnet-34
model. For creating this model, we utilized the FastAi
library. After that, we downloaded a custom pretrained
Resnet image classification model.

Fig. 6. Scaling of EfficientNet architecture.

While creating and running this model, we enabled the

GPU environment on our Google Colab notebook. Next, we
had to ensure our model was running on the TensorFlow 1.x
Fig. 4. Simplified representation of Resnet-34 model.
environment and Keras 2.3.1 was installed. After that, we
imported the EfficinetNetB0 model from the Keras library.
Next, we initialized our classification model by fine- Then we set the input resolution of the images to be 150 x
tuning its final layer while the rest of the model was frozen. 150.

93
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

The next task was to import our ocular disease dataset D. Classification Using VGG16 Model
into the notebook and then utilize transfer learning in order The VGG-16 model was developed by researchers of
to classify ocular diseases from color fundus images. After University of Oxford and the VGG-16 paper was published
importing the fundus images, we passed the data through a in 2015. It’s widely regarded as one of the best image
training generator function to prepare it for training. classification models out there, and it achieved 92.7%
Next, we set the number of epochs for which our model accuracy on the ImageNet dataset. It is built using a large
will train on the dataset to be 150, and the final layer of the number of tiny convolutional filters that allow the model to
EfficientNet model was removed so that it can be replaced learn about complicated pixel relational data.
by eight layers that correspond to the eight classes of our At first, we download the required libraries and
dataset. Finally, we trained the model and then evaluated its dependencies to make sure that the environment is
performance using the images inside the test directory. compatible with the VGG-16 classification model. The
C. Classification Using MobileNetV2 VGG-16 model expects the data to have an input size of 224
x 224. Moreover, that is why we had to resize our training
MobileNetV2 is an image classification model that was images accordingly.
developed by Google, and its task is to provide efficient real-
time classification even in constrained computing After that, we performed some preprocessing on the
environments such as smartphones. [30] This model is quite images in order to make them suitable for the VGG-16
similar to the previous two models in the sense that it also model. This was done using the ImageDataGenerator module
utilizes transfer learning, and it’s pretrained on the ImageNet of the Keras library. Those preprocessing steps involved
dataset as well. The architecture of the MobileNetV2 model setting the re-scale value to 1/255, shear range to 0.2, zoom
is illustrated in figure 7. range to 0.2, and the value of horizontal flip to true. The task
of the ImageDataGenerator function is to generate the
This image classification framework uses an inverted
preprocessed images based on the parameters that we have
residual structure in which the input and output layers of the set so that those images can be fed to the VGG-16 model.
residual blocks comprise thin bottleneck layers. Moreover, Some samples of the preprocessed images can be seen in
the convolutions used in this model are quite lightweight and figure 8.
it does not have non-linearities in its narrow layers.
In order to implement the MobileNetV2 model on the
ODIR dataset, we had to import the dataset into our Colab
notebook and then convert the images into a TensorFlow
dataset. We built the TensorFlow dataset by using the
ImageForlder API provided by the TensorFlow framework.
After that, we instantiated the MobileNetV2 classification
model in which the classification layers were dependent on
the last layer before the flatten operation was performed on
it. Then we set the compiled model using categorical cross
entropy as the loss function and accuracy as the evaluation
metric. Finally, we generated the Accuracy vs. Epoch and
Cross Entropy vs. Epoch graphs in order to evaluate the
efficiency of our model.

Fig. 8. A look at the preprocessed ODIR datasets.

In order to make our VGG-16 model train and predict on

eight different ocular disease classes, we had to append two
Dense layers to the existing VGG-16 architecture. The
overall architecture of the VGG-16 model after modification
is shown in table II.
This model is set to use the Adam optimizer and
categorical cross entropy loss function. It also used the
softmax activation function. After those parameters were set,
we ran the model on the training set for 150 epochs. When
the training was finished, we evaluated its performance on
the validation set.
Fig. 7. MobileNetV2 architecture.

94
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

E. Evaluation of The Models 25088)

In order to analyze the results provided by the fc1 (Dense) (None, 4096) 102764544
classification models in a comprehensible way, we have fc2 (Dense) (None, 4096) 16781312
used evaluation metrics such as accuracy, precision, recall,
F1 score, etc. The formulas of these evaluation metrics are Total params: 134,260,544
Trainable params: 134,260,544
shown below: Non-trainable params: 0
(1)

IV. EXPERIMENTAL RESULTS

(2)
In total we had to perform four different experiments in
order to determine the performance of the four classification
(3) models (Resnet-34, EfficientNet, MobileNetV2, VGG-16).
A. Resnet-34 Classification Model
(4) We trained the Resnet-34 model for up to 200 epochs for
the ocular disease classification task. Its performance was
impressive, and it achieved an accuracy of 93.47% on the
Here, training set. Moreover, it achieved an accuracy of 90.85% on
TP = True Positive (The total number of images that are the test set, which comprised of previously unseen images.
correctly detected to be positive) We had split the dataset consisting of 5000 color fundus
photographs into a training set comprising 3500 images
FP = False Positive (The total number of images that are (70% of the total images) and a test set of 1500 images (30%
predicted to be positive but actually are negative) of the total images) of the total MRI scans). We primarily
TN = True Negative (The number of images that are evaluated our four models on the test set in order to make it
accurately predicted to be negative) easy for us to compare and contrast the models with one
another. This was done to simulate how the Resnet-34 model
FN = False Negative (The number of images that are would perform in a real-life scenario with previously unseen
incorrectly predicted to be negative) color fundus images. The output generated by the Resnet-34
model on the test set is shown in figure 10.
TABLE II. VGG-16 ARCHITECTURE
The confusion matrix for test set is shown in Figure 9.
Output Out of the eight classes the most successful prediction made
Layer (type) Param #
Shape
(None, 224, on the other diseases class. Overall, the performance of this
block1 conv1 (Conv2D) 1792 model was quite satisfactory on the test set.
224, 64)
(None, 224,
block1 conv2 (Conv2D) 36928
224, 64)
(None, 112,
block1_pool (MaxPooling2D) 0
112, 64)
(None, 112,
block2 conv1 (Conv2D) 73856
112, 128)
(None, 112,
block2_conv2 (Conv2D) 147584
112, 128)
(None, 56,
block2_poo] (MaxPooling2D) 0
56, 128)
(None, 56,
block3 conv1 (Conv2D) 295168
56, 256)
(None, 56,
block3_conv2 (Conv2D) 590080
56, 256)
(None, 56,
block3 conv3 (Conv2D) 590080
56, 256)
(None, 28,
block3_pool (MaxPooling2D) 0
28, 256)
(None, 28,
block4 conv1 (Conv2D) 1180160
28, 512)
(None, 28,
block4_conv2 (Conv2D) 2359808 Fig. 9. Confusion matrix produced by the Resnet-34 classification model
28, 512)
(None, 28,
block4_conv3 (Conv2D) 2359808 A confusion matrix is a type of layout that provides
28, 512)
block4_pool (MaxPooling2D)
(None, 14,
0
visualization of the performance of an algorithm. Each row
14, 512) of the confusion matrix represents the instances in a
(None, 14, true/actual class. And each column of the matrix represents
block5_conv! (Conv2D) 2359808
14, 512)
the instances in a predicted class. The values located at the
(None, 14,
block5_conv2 (Conv2D)
14, 512)
2359808 main diagonal of the matrix represents the instances at which
(None, 14, the model was able to accurately predict the class to which
block5_conv3 (Conv2D) 2359808
14, 512) an image belongs to. On the other hand, all of the other
block5_pool (MaxPooling2D)
(None, 7, 7,
0 nonzero values in the confusion matrix represents the
512) instances at which the model had incorrectly classified an
flatten (Flatten) (None, 0 image.

95
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

The precision of the Resnet-34 model is 93.70%, and its

recall was 92.65%. Furthermore, the model achieved an F1
score of 93.17%. The class-wise accuracy, precision, recall
and F1 score of this model is given in table number III.

TABLE III. PERFORMANCE OF THE RESNET-34 MODEL

Class Accuracy Precision Recall F1 Score
AMD 99.64% 0.99 0.98 0.98
Cataract 99.64% 0.98 0.99 0.98
Diabetes` 92% 0.57 0.97 0.72
Glaucoma 92.07% 0.98 0.61 0.75
Hypertension 99.34% 0.98 0.99 0.98
Normal 99.39% 0.98 0.98 0.98
Myopia 99.52% 0.99 0.97 0.98
Other 99.82% 0.98 1.0 0.99

B. EfficientNet Classification Model

Just like the Resnet-34 model we also trained the
Efficient-NetB0 model for 500 epochs. It performed
exceptionally well on the training set as well as the test set. It
achieved an accuracy of 94.97% on the training set. And it
achieved an accuracy of 93.82% on the test set.
Fig. 10. Output generated by Restnet-34.
The confusion matrix produced by the EfficientNetB0
model on the test set is shown in Figure 11. This model
provided the most successful prediction on the — class.
The EfficientNet model had an overall precision of
92.73%, and its recall was 96.25%. Furthermore, this model
achieved an F1 score of 93.74%. The class-wise analysis of
the accuracy, precision, recall, and F1 score of the
EfficientNet model is given in table number IV. As
mentioned before, we used the softmax function as a loss
function for the EfficientNet model. The figure 12 shows the
training and validation accuracy of the model. Here the dots
represent the training accuracy and the curve below shows
the validation accuracy. We can observe from this figure that
the training accuracy drastically increases as the number of
epochs increases. The validation accuracy increases as well
but it sometimes declines as well during the training process.
And the graph shown in figure 13 illustrates the training
and validation loss of the EfficientNet model. Here the x-axis
Fig. 11. Confusion Matrix generated by EfficientNet
represents the number of epochs and the y-axis represents the
training and validation loss of the model. The MobileNetV2 model had a precision score of
We can clearly see that both the training and validation 93.33%, and its recall was 89.67%. Furthermore, this model
losses of the EfficientNet model drastically decreases as the achieved an F1 score of 91.46%. The details of the class-
number of epochs increase. wise accuracy, precision, recall and F1 score of the model is
shown in table number V.
C. MobileNetV2 Classification Model
The performance of the MobileNetV2 model was fairly TABLE IV. PERFORMANCE OF THE EFFICIENTNET MODEL
close to the previous two classification models. It achieved Class Accuracy Precision Recall F1 Score
an accuracy of 95.56% on the training set. And it achieved an AMD 99.27% 0.95 0.98 0.96
accuracy of 94.32% on the test set. Cataract 98.93% 0.98 0.92 0.95
The training and validation accuracy graph as well as the Diabetes 99.27% 0.95 0.99 0.97
training and validation loss graph is illustrated in figure 14. Glaucoma 98.27% 0.95 0.99 0.97
Hypertension 98.73% 1.0 0.96 0.98
Normal 99.53% 0.94 0.97 0.98
Myopia 98.67% 0.93 0.92 0.92
Other 99% 0.99 0.92 0.96

96
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

Fig. 12. Training and Validation Accuracy Graph of the EfficientNet

model.

Fig. 14. Training and Validation Accuracy and Loss Graph of the VGG-16
Model.

TABLE V. PERFORMANCE OF THE MOBILENETV2 MODEL

Class Accuracy Precision Recall F1 Score
AMD 98.73% 0.94 0.94 0.94
Cataract 97.73% 0.97 0.91 0.94
Diabetes 98.47% 0.91 0.96 0.93
Glaucoma 98.67% 0.94 0.99 0.93
Hypertension 92.47% 0.92 0.93 0.91
Normal 98.13% 0.93 0.95 0.94
Myopia 98.20% 0.92 0.88 0.90
Fig. 13. Training and Validation Loss of EfficientNet Model. Other 98.87% 0.99 0.91 0.95

D. VGG-16 Classification Model TABLE VI. PERFORMANCE OF THE VGG-16 MODEL

Out of the four classification models the VGG-16 model Class Accuracy Precision Recall F1 Score
had the best performance in terms of accuracy. It achieved an AMD 98.27% 0.94 0.90 0.92
accuracy of 98.65% on the training set. The accuracy Cataract 97.6% 0.91 0.88 0.89
achieved it achieved on the test set was 97.23%. The Diabetes 97.87% 0.88 0.93 0.91
confusion matrix of the VGG-16 model is shown in figure Glaucoma 97.67% 0.92 0.95 0.93
15. Hypertension 96.89% 0.87 0.92 0.90
Normal 97.93% 0.93 0.94 0.93
We can observe from figure 14 that the training as well
Myopia 97.00% 0.85 0.84 0.94
as the validation accuracy rises exponentially as the number
Other 98.33% 0.98 0.89 0.93
of epochs increase. Furthermore, both training and validation
loss (cross entropy) decreases as the number of epochs rises.
This happens because as the model gets trained for more and
more epochs it learns more about the features of the images
and it gets better at differentiating between the images
belonging to different classes, thus increasing its accuracy.
The VGG-16 model had a precision score of 96.73%, and
its recall was 93.76%. Furthermore, this model achieved an
F1 score of 95.22%. The class-wise details of this model are
included in table number VI.
Our models have outperformed some of the existing
solutions to ocular disease detection and classification. For
instance, He et al. [9] had achieved an F1 score of 90.4% of
the ODIR dataset using their ResNet-34 model.
However, our Resnet-34 model achieved an F1 score of
93.17%.
Fig. 15. Confusion matrix produced by VGG-16.

97
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

This was because we had trained our model for more [11] Tayal, A., Gupta, J., Solanki, A., Bisht, K., Nayyar, A., & Masud, M.
epochs and we had fine-tuned our model for better gaining (2021). DL-CNN-based approach with image processing techniques
for diagnosis of retinal diseases. Multimedia Systems, 1-22.
accuracy.
[12] Akil, M., Elloumi, Y., & Kachouri, R. (2020). Detection of retinal
abnormalities in fundus image using CNN deep learning networks.
V. CONCLUSION
[13] Meng, X., Xi, X., Yang, L., Zhang, G., Yin, Y., Chen, X. (2018).
In this study, we have developed four neural network- Fast and effective optic disk localization based on convolutional
based ocular disease, classification models. Those models are neural network.Neurocomputing,,312,285–295.
Resnet-34, EfficientNet, MobileNetV2 and VGG-16. Out of https://doi.org/10.1016/j.neucom.2018.05.114
which, the VGG-16 provided the best accuracy of 97.23% [14] He, J., Li, C., Ye, J., Qiao, Y., Gu, L. (2021). Self-speculation of
when it comes to classifying ocular diseases from fundus clinical features based on knowledge distillation for accurate ocular
disease classification. Biomedical Signal Processing and Control, 67,
photographs. The performance of the other models was also 102491. https://doi.org/10.1016/j.bspc.2021.102491
satisfactory. We have performed extensive experiments on [15] Roy, A. G., Conjeti, S., Karri, S. P., Sheet, D., Katouzian, A.,
the publicly available ODIR- 2019 dataset to validate our Wachinger, C., Navab, N. (2017). ReLayNet: retinal layer and fluid
proposed method's effectiveness. Our proposed method can segmentation of macular optical coherence tomography using fully
generate more impressive results than the existing CNN- convolutional networks. Biomedical Optics Express, 8(8), 3627.
https://doi.org/10.1364/boe.8.003627
based ocular disease classification models while at the same
time requiring lower computational power. [16] Lee, C. S., Tyring, A. J., Deruyter, N. P., Wu, Y., Rokem, A., Lee, A.
Y. (2017). Deep-learning based, automated segmentation of macular
The best part about our proposed method is that it can edema in optical coherence tomography. Biomedical Optics Express,
8(7), 3440. https://doi.org/10.1364/boe.8.003440
easily be extended to other types of medical image-based
disease classification. Furthermore, the models described in [17] Playout, C., Duval, R., Cheriet, F. (2019). A Novel Weakly
Supervised Multitask Architecture for Retinal Lesions Segmentation
this study can be used in order to build a user-friendly, real- on Fundus
time ocular diseases classification system. Such a system will [18] Images. IEEE Transactions on Medical Imaging, 38(10), 2434–2444.
be a great help to the medical professionals and it will https://doi.org/10.1109/tmi.2019.2906319
revolutionize the field of ocular disease diagnosis. [19] Hu, K., Zhang, Z., Niu, X., Zhang, Y., Cao, C., Xiao, F., Gao, X.
(2018). Retinal vessel segmentation of color fundus im-ages using
REFERENCES multiscale convolutional neural network with an im-proved cross-
[1] Bourne, R. R., Stevens, G. A., White, R. A., Smith, J. L., Flaxman, S. entropy loss function. Neurocomputing, 309, 179–191.
R., Price, H., Jonas, J. B., Keeffe, J., Leasher, J., Naidoo, K., https://doi.org/10.1016/j.neucom.2018.05.011
Pesudovs, K., Resnikoff, S., & Taylor, H. R. (2013). Causes of vision [20] M. D. M. Goldbaum, STARE Dataset Website, Clemson University,
loss worldwide, 1990–2010: a systematic analysis. The Lancet Global Clemson, SC, USA, 1975
Health, 1(6). https://doi.org/10.1016/s2214-109x(13)70113-x [21] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B.
[2] Sommer, A., Tielsch, J. M., Katz, J., Quigley, H. A., Gottsch, J. D., van Ginneken, Digital Retinal Image for Vessel Extraction (DRIVE)
Javitt, J. C., Martone, J. F., Royall, R. M., Witt, K. A., & Ezrine, S. Database, Image Sciences Institute, University Medical Center
(1991). Racial Differences in the Cause-Specific Prevalence of Utrecht, Utrecht, The Netherlands, 2004
Blindness in East Baltimore. New England Journal of Medicine, [22] Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D.,
325(20), 1412–1417. https://doi.org/10.1056/nejm199111143252004 Narayanaswamy, A., Venugopalan, S., Widner, K., Madams, T.,
[3] Congdon, N., O'Colmain, B., Klaver, C. C., Klein, R., Muñoz, B., Cuadros, J., Kim, R., Raman, R., Nelson, P. C., Mega, J. L., Webster,
Friedman, D. S., Kempen, J., Taylor, H. R., Mitchell, P., & Eye D. R. (2016). Development and Validation of a Deep Learning
Diseases Prevalence Research Group (2004). Causes and prevalence Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus
of visual impairment among adults in the United States. Archives of Photographs. JAMA, 316(22), 2402.
ophthalmology (Chicago, Ill. : 1960), 122(4), 477–485. https://doi.org/10.1001/jama.2016.17216
[4] Application of Ocular Fundus Photography and Angiography. (2014). [23] Li, Z., He, Y., Keel, S., Meng, W., Chang, R. T., He, M. (2018).
Ophthalmological Imaging and Applications, 154–175. Efficacy of a Deep Learning System for Detecting Glaucomatous
https://doi.org/10.1201/b17026-12 Optic Neu-ropathy Based on Color Fundus Photographs.
[5] Rowe, S., MacLean, C. H., & Shekelle, P. G. (2004). Preventing Ophthalmology, 125(8), 1199–1206.
Visual Loss From Chronic Eye Disease in Primary Care. JAMA, https://doi.org/10.1016/j.ophtha.2018.01.023
291(12), 1487. https://doi.org/10.1001/jama.291.12.1487 [24] Karri, S. P., Chakraborty, D., Chatterjee, J. (2017). Transfer learning
[6] Kessel, L., Erngaard, D., Flesner, P., Andresen, J., Tendal, B., & based classification of optical coherence tomography images with dia-
Hjortdal, J. (2015). Cataract surgery and age‐related macular betic macular edema and dry age-related macular degeneration.
degeneration. An evidence‐based update. Acta Ophthalmologica, Biomed-ical Optics Express, 8(2), 579.
https://doi.org/10.1364/boe.8.000579
93(7), 593–600. https://doi.org/10.1111/aos.12665Li, N., Li, T., Hu,
C., Wang, K., & [25] Zhong, Z., Jin, L., & Xie, Z. (2015, August). High performance
[7] Li, N., Li, T., Hu, C., Wang, K., & Kang, H. (2021). A Benchmark of offline handwritten chinese character recognition using googlenet and
Ocular Disease Intelligent Recognition: One Shot for Multi-disease directional feature maps. In 2015 13th International Conference on
Detection. Benchmarking, Measuring, and Optimizing, 177–193. Document Analysis and Recognition (ICDAR) (pp. 846-850). IEEE.
https://doi.org/10.1007/978-3-030-71058-3_11 [26] Li, N., Li, T., Hu, C., Wang, K., & Kang, H. (2021). A Benchmark of
[8] Miranda, E., Aryuni, M., & Irwansyah, E. (2016, November). A Ocular Disease Intelligent Recognition: One Shot for Multi-disease
survey of medical image classification techniques. In 2016 Detection. Benchmarking, Measuring, and Optimizing, 177–193.
International Conference on Information Management and https://doi.org/10.1007/978-3-030-71058-3_11
Technology (ICIMTech) (pp. 56-61). IEEE. [27] J. Deng, W. Dong, R. Socher, L. Li, Kai Li and Li Fei-Fei,
[9] He, J., Li, C., Ye, J., Qiao, Y., & Gu, L. (2021). Multi-label ocular "ImageNet: A large-scale hierarchical image database," 2009 IEEE
disease classification with a dense correlation deep neural network. Conference on Computer Vision and Pattern Recognition, 2009, pp.
Biomedical Signal Processing and Control, 63, 102167 248-255, doi: 10.1109/CVPR.2009.5206848.
[10] Li, C., Ye, J., He, J., Wang, S., Qiao, Y., & Gu, L. (2020, April). [28] Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model
Dense correlation network for automated multi-label ocular disease scaling for convolutional neural networks. In International
detection with paired color fundus photographs. In 2020 IEEE 17th Conference on Machine Learning (pp. 6105-6114). PMLR.
International Symposium on Biomedical Imaging (ISBI) (pp. 1-4). [29] Qassim, H., Verma, A., & Feinzimer, D. (2018, January).
IEEE Compressed residual-VGG16 CNN model for big data places image
recognition. In 2018 IEEE 8th Annual Computing and

98
Asian Journal of Convergence in Technology Volume VII and Issue II
ISSN NO: 2350-1146 I.F-5.11

Communication Workshop and Conference (CCWC) (pp. 169-175). [31] Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model
IEEE. scaling for convolutional neural networks. In International
[30] Targ, S., Almeida, D., & Lyman, K. (2016). Resnet in resnet: Conference on Machine Learning (pp. 6105-6114). PMLR.
Generalizing residual architectures. arXiv preprint arXiv:1603.08029.

99