4th International Conference on Industrial Artificial

Intelligence (IAI), Aug. 24-27, 2022, Shenyang, China

Retinal Medical Image Classification Based on

Deep Convolutional Neural Network AlexNet
Zehan Tian, Jing Wang*, Meng Zhou Yanzhu Zhang Mingyu Shi
School of Electrical and Control School of Automation Department of Ophthalmology,
Engineering, North China and Electrical Engineering, The Fourth Affiliated Hospital
University of Technology Shenyang Ligong University of China Medical University,
Beijing, China Shenyang , China Eye Hospital of China Medical
[email protected] [email protected] University, The Key Laboratory
[email protected] of Lens in Liaoning Province
[email protected] Shenyang, China
2022 4th International Conference on Industrial Artificial Intelligence (IAI) | 978-1-6654-5120-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/IAI55780.2022.9976745

[email protected]

Abstract—Eye diseases will have a very serious impact on detachment refers to the separation of retinal neuroepithelium
the life, study and work of patients. In order to better assist and pigment epithelium. Patients will have symptoms such as
doctors in their work, it is very meaningful to use deep learning spotted vision, light spots shaking or shadows covering their
neural networks for medical image analysis and auxiliary medical
diagnosis. In this paper, we use deep neural network AlexNet eyes [10]. (6) Asteroid hyalosis is a common degenerative
combined with Adam optimization algorithm to classify images of process in which fat calcium globules accumulate in the
four common eye diseases: vitreous opacity, vitreous opacity with vitreous humor. Although its cause and mechanism are still
retinal detachment, asteroid hyalosis and vitreous hemorrhage. unclear, it is related to aging and some systemic diseases
Use confusion matrix, accuracy, precision, recall, specificity and [11]. (7) Vitreous hemorrhage is a relatively common eye
other evaluation indicators to evaluate its classification effect.
The application results of the above methods on ophthalmic diseasewhich leads to refractive interstitial clouding and cause
ultrasound images from actual hospitals show that AlexNet has vision loss. The most common cause is proliferative diabetic
high classification accuracy for actual ultrasound pattern, and retinopathy, followed by ocular trauma [12]. This article
can be used to assist doctors in ophthalmic disease diagnosis. focuses on the diagnosis of four types of diseases: vitreous
Index Terms—ultrasound pattern; computer aided diagnosis; opacities, vitreous opacities with retinal detachment, asteroid
deep learning; AlexNet;
hyalosis and vitreous hemorrhage.
With the increasing of ophthalmic patients and the mixed
I. I NTRODUCTION
occurrence of various diseases, more and more refined treat-
Eye is one of the most important sensory organs in the ment methods are required. However, the work intensity of
human body. Eye disease can have a serious impact on a doctors is getting higher, so the urgent need is to develope
person’s life, studies and work [1, 2]. It is essential to prevent, new technology to assist doctors in diagnosis [13–15]. AlexNet
diagnosis and treat the eye disease in a timely manner. Many won the 2012 ILSVR (ImageNet Large-Scale Visual Recog-
people around the world are suffering from eye diseases and nition Challenge) competition, which was the first time that
even facing the risk of blindness, so it is very necessary to use a deep learning neural network participated the competition.
new technical means to find and diagnosis the eye disease as ResNet won in the 2015 ILSVRC competition with a Top5
earlier as possible [3, 4]. error rate of 3.57% on the classification track, which has
Common eye diseases include as follows: (1) Glaucoma exceeded the error rate that the human eye can achieve.
is a group of diseases that damage the optic nerve, vision Lots of research has made good progress linking neural
loss and visual field loss. According to recent estimates, the networks to the analysis of medical images and auxiliary
number of glaucoma patients worldwide will increase from medical diagnosis. Ting DSW et al. built a deep learning
76.5 million in 2020 to 111.8 million in 2040 [5]. (2) Cataract system to detect eye diseases such as diabetic retinopathy,
is a visual disorder caused by the opacity of the lens caused sight-threatening diabetic retinopathy and possible glaucoma
by protein degeneration inside the lens, which can lead to [16]. The deep learning system they built has high sensitivity
varying degrees of vision loss or even blindness [6]. (3) and specificity in identifying diabetic retinopathy and related
Macular degeneration is a disease caused by the atrophy of the eye diseases. It is worth noting that the above method divides
macula in the retina or the accumulation of metabolites and the data into multiple layers rather than the entire image, so
the function deterioration. Age-related macular degeneration it is difficult to directly apply in practice. When assessing the
is the most prevalent retinal disease in the Western world severity of age-related macular degeneration (AMD), it can be
[7, 8]. (4) Vitreous opacity is caused by the development of time consuming if all handled manually by an expert. So Yifan
opacities, liquefaction and shrinkage of the vitreous over time, Peng et al. proposed a deep learning model: DeepSeeNet [17].
resulting in an opaque body within the vitreous [9]. (5) Retinal This method can directly target color fundus photographs and

978-1-6654-5120-8/22/$31.00©2022 IEEE
1
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.
 automatically classify patients according to the severity of age- contains 60,000,000 parameters and 65,000 neurons. The
related macular degeneration. The deep learning algorithms for general optimization algorithm for AlexNet was stochastic gra-
classifying glaucoma lesions often required extensive human dient descent, while the adaptive moment estimation (Adam)
labelling of the dataset. Medeiros FA proposed to use spec- optimization algorithm was chosen in this experiment. The
tral domain optical coherence tomography (SDOCT) data to detailed structure can be seen in Fig. 1. The feature extraction
train deep learning algorithms to quantify glaucoma structural is finished at convolutional layer, which contains several
damage on optic disc photographs, so the glaucoma disease convolutional kernels that generate their own feature maps
can be better diagnosed [18]. and its shared weight largely reduces the amount of training
Deep learning not only helps ophthalmologists in the aux- of the model parameters. Pooling is applied to reduce the
iliary diagnosis of eye diseases, but also makes good progress feature matrix extracted from the convolutional neural network
in other disease diagnosis. Lakshmanaprabu S.K. et al. used and the maximum pooling is selected here. As shown in
optimal deep neural network (ODNN) to extract deep features Fig. 2, the 4*4 feature matrix can be obtained as a 2*2 matrix
from CT lung images and used linear discriminant analysis after the maximum pooling layer, which largely reduces the
(LDA) to reduce the dimension of features [19]. ODNN computational effort when training the network.
effectively reduced the manual labeling time and human errors, Before a deep convolutional neural network can correctly
and the specificity reached 94.2%. Hassan Ali Khan et al. classify the target data it needs to be trained extensively with
used the edge detection technology to find and crop the a training dataset, during which the parameters and weights in
region of interest in the brain tumor image, then proposed a the network are continuously iterated. To speed up the training
simple convolutional neural network for efficient brain tumor of the network, AlexNet can be computed on the GPU, and
classification [20]. This proposed network is very small, and a number of optimisations have been made to improve the
the training time for each round is about 200 seconds, which is network’s classification ability. First, deep neural networks all
much lower than the existing network methods: 456 seconds need activation functions to improve their nonlinearization and
for VGG-16 and 606 seconds for ResNet-50. But how well AlexNet chooses to use ReLU as the activation function at the
such a small network will perform when applied to multi- convolutional layer, the formula is
disease classification is unknown. Therefore, we choose a
smaller AlexNet network for multi-disease classification, and ReLU = max(x, 0) (1)
it is also one of the goals of this study to examine whether
a smaller network can complete the work of multi-disease Previous networks often used sigmoid or tanh as the activa-
classification. tion function, but these activation functions are complicated to
In view of the wide variety of eye diseases and the location derive during training calculations and tend to cause gradient
of the lesions are very close, this paper uses the deep learning disappearance, while ReLU avoids these problems by increas-
neural network AlexNet combined with the Adam optimizer ing the nonlinearity of the network. Secondly, Dropout module
to classify four types of eye disease pictures. AlexNet includes is added to AlexNet, where Dropout randomly deactivates neu-
technologies such as convolution layer, pooling layer, rectified rons at a set ratio during each round of training. For example, if
linear unit (ReLU) and Dropout. On this basis, the Adam a ratio of 50% is set, Dropout will randomly deactivate 50% of
optimization algorithm is combined to further enhance the the neurons in that layer during training, which is equivalent
training effect of the model. This paper evaluates the clas- to randomly cutting off 50% of the pathways in that layer.
sification effect of AlexNet network in multiple ophthalmic Each round of training is scaled to randomly deactivate the
diseases through indicators such as confusion matrix, accuracy, network, which greatly increases the generalization ability of
precision, recall and specificity, and verifies the feasibility of deep convolutional networks and better prevents overfitting.
AlexNet for practical auxiliary diagnosis. Fig. 3 shows Dropout randomly deactivating neurons in the
The paper is divided into the following sections: section II fully connected layer.
contains introduce the experimental model and evaluation in- Deep convolutional neural networks require the use of a loss
dex, section III describes the ophthalmic data set, environment function to measure the error between the predicted and actual
and results of the experiment, section IV the experimental labels, and reduce the error by passing the update parameters
results and analyzes the challenges in future work. backwards. This experiment was conducted for a classification
task, and so the CrossEntropy Loss function was chosen as a
II. M ODEL AND E VALUATION M ETRICS
tool to measure the error between the predicted labels and the
A. Deep Convolutional Neural Network Model actual labels. The CrossEntropy Loss is defined as follows,
AlexNet [21] is a deep convolutional neural network jointly
n
developed by Hinton and his student Alex Krizhevsky, which X
won the championship in the ImageNet LSVRC-2010 com- Hi = − Oi∗ log (Oi ) (2)
i=1
petition. AlexNet consists of five convolutional layers, three
pooling layers and three fully connected networks. The final where Hi is the cross entropy between the predicted and true
output layer is a fully connected layer with 1000 channels probabilities of the data in class i, and Oi∗ is the true value of
and an activation function of softmax. The entire network the classification label. Oi is the probability of the ith label

2
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. The structural framework of AlexNet

of the ith neuron of the Softmax pre-stage. Softmax calculates

the probability that each outcome is distributed between (0, 1)
and sums to 1, corresponding to each data label.
The CrossEntropy Loss input is the result of the forward
pass of the convolutional neural network. The predicted label
probabilities of the AlexNet forward pass output undergo a
new backward pass after each round of training. When the
predicted output of the network is closer to the real data
label, the smaller the cross entropy and the smaller the value
updated by the reverse pass, the higher the accuracy of the
Fig. 2. Maximum pooling layer network prediction. Backward transfer is an important process
for updating the internal parameters of deep convolutional
neural network iterations, and choosing the right optimisation
algorithm can save network training time and computational
resources.
Adam algorithm comprehensively considers the first-order
moment and second-order moment of the gradient. It requires
little memory, so is computationally fast and effective. Its
iterative formula is

g = hθ x i − y i x i

wt = β1 wt−1 + (1 − β1 ) ∗ g
vt = β2 vt−1 + (1 − β2 ) ∗ g 2
wt
wt∼ = (4)
1 − β1t
vt
vt∼ =
1 − β2t
Fig. 3. Dropout randomly deactivated neurons
α
θj = θj−1 − wt∼ ∗ √ ∼
vt + ϵ
predicted by the neural network, calculated from the Softmax
function of the fully connected layer, where g is the calculated gradient, wt is the first-order moment
of the gradient, β1 is the decay coefficient of the first-order
exp (yi ) moment, vt is the second-order moment of the gradient g, β2
O i = Pn (3) is the decay coefficient of the second-order moment, θ is the
j=1 exp (yj )
parameter to be updated or solved for, wt∼ and vt∼ are the
where n is the total number of all labels and yi is the output offset corrections for wt and vt .

3
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.
 B. Evaluation indicators III. E XPERIMENT
The evaluation metrics include Confusion Matrix, Accuracy, A. Datasets
Precision, Recall and Specificity. The sample data in the test The experiment data was obtained from the colour ultra-
set are divide into four main categories according to their sound cropping process provided by the hospital. The ocular
true and predicted labels: TP (True Positive), the number of ultrasound examines the vitreous and lens to determine the
samples that are predicted by the model to be positive and anterior and posterior diameters of the lens and length of
whose true labels are also positive. TN (True Negative), the intraocular visual axis, and to look for echogenicity at the edge
number of samples that are predicted by the model to be of lens. The vitreous humor may also be examined to look
negative and whose true labels are also negative. FP (False for conditions such as infection or blood accumulation in the
Positive), the number of positive samples predicted by the vitreous humor. The raw data were first manually cropped to
model but the true labels are negative samples. FN (False remove gaps and interfering information and sorted, and finally
Negative), the number of positive samples predicted by the reviewed by a medical professional before being applied to the
model, but the real label is the number of positive samples. experiment. The data set consisted of 1,966 images, with five
The confusion matrix can then be drawn from the above four main categories: vitreous opacity 600 images, vitreous opacity
values, as shown in Fig. 4. retinal detachment 227 images, asteroid hyalosis 304 images,
vitreous hemorrhage 185 images and normal 600 images. The
data were divided into training set and test set in a ratio of
9:1 for training. Fig. 5 shows a display of the five categories.

Fig. 4. Diagram of the confusion matrix Fig. 5. Presentation of the 5 categories in the dataset

Accuracy, precision, recall and specificity of the model

are calculated based on the values in the confusion matrix. B. Experimental environment and results
Accuracy is the most common index to evaluate the model, This experiment was run on python 3.7 with an AMD
generally the higher the better, Ryzen 7 5800H with Radeon Graphics 3.20 GHz CPU and
an NVIDIA GeForce RTX 3060 Laptop GPU. The specific
TP + TN parameters of the AlexNet interlayer structure for this exper-
Accuracy = (5)
P+N iment can be found in Tab I, where k is kernels, s is stride
Precision is the proportion of samples with a positive and and p is padding. During training, the input images were first
correct prediction to those with a positive prediction, the randomly cropped to a size of 227*227 and randomly inverted
formula is horizontally. This is to enhance the data to prevent overfitting,
TP Fig. 6 shows the images processed as input to the network. The
Precision = (6)
TP + FP loss function set for this experiment is the CrossEntropy Loss,
Recall is the number of all samples with a positive true label and the backward transfer optimization algorithm is chosen as
that are correctly predicted, the formula is the Adam optimization algorithm with a learning rate set at
0.0002. The learning rate should not be too high or the training
TP
Recall = (7) effect will be affected. The average training time for a round
TP + FN of input 32 images for AlexNet was about 4.5 seconds.
Precision and recall are a contradictory combination of indi- AlexNet was initially designed with a stochastic gradient
cators, but they can also evaluate the function of the model descent (SGD) algorithm. If trained with a stochastic gradient
in more detail. Specificity is the proportion of negative cases descent algorithm with a learning rate of 0.002, all other
identified out of all negative cases and the formula is things being equal, the accuracy of AlexNet for this eye
TN disease 5 classification test was 72.8% after 300 rounds. Tab II
Specificity = (8) shows that, all other experimental conditions being equal, the
FP + TN

4
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
A LEX N ET INTERNAL PARAMETERS

1 data Image Input 227x227x3 images

2 conv1 Convolution k=3x96x11x11, s=4, p=0
3 relu1 ReLU ReLU
4 pool1 Max Pooling k=3x3, s=2, p=0
5 conv2 Convolution k=48x256x5x5, s=1, p=2
6 relu2 ReLU ReLU
7 pool2 Max Pooling k=3x3, s=2, p=0
8 conv3 Convolution k=256x384x3x3, s=1 , p=1 Fig. 6. Presentation of images pre-processed before input to the network
9 relu3 ReLU ReLU
10 conv4 Convolution k=384x384x3x3, s=1, p=1
11 relu4 ReLU ReLU
12 conv5 Convolution k=384x256x3x3, s=1 , p=1
13 relu5 ReLU ReLU
14 pool5 Max Pooling k=3x3, s=2, p=0
15 fc6 Fully Connected 4096 fully connected layer
16 relu6 ReLU ReLU
17 drop6 Dropout 50% dropout
18 fc7 Fully Connected 4096 fully connected layer
19 relu7 ReLU ReLU
20 drop7 Dropout 50% dropout
21 fc8 Fully Connected 5 fully connected layer
22 prob Softmax Softmax
23 output Classification Output 5 disease categories

TABLE II
C OMPARISON OF THE TWO OPTIMISATION ALGORITHMS

RUN SGD Adam

10 29.20% 69.70%
20 30.8 71.9%
30 33.30% 75.90%
40 37.40% 76.40%
50 41.50% 78.50%
... ... (Completed)
300 72.80% Fig. 7. Confusion matrix for test data: (a) AlexNet-Adam 10 epochs, (b)
AlexNet-Adam 30 epochs, (c) AlexNet-Adam 50 epochs, (d) AlexNet-SDG
300 epochs

Adam optimisation algorithm requires fewer training rounds

and achieves a higher accuracy rate. effect is the most stable.
After resetting the internal parameters of the AlexNet model
and using the Adam optimizer to train 10, 30, 50 and 70
rounds, the accuracy of this eye disease5 classification test was IV. C ONCLUSION
69.70% for 10 rounds, 75.9% for 30 rounds and 78.5% for 50
rounds. According to the comparison AlexNet is the strongest AlexNet with Adam optimization is used to complete the
classifier after 50 rounds of training, which shows that deep 5 classification of eye disease images. This algorithm can
neural networks can only have better potential after sufficient achieve an accuracy of 78.5% after training, and effectively
training. The accuracy of the test after 70 rounds of training extract the features of eye disease images and classify them.
was 74.3%, which was not even as high as the accuracy after One round of training time is about 4.5 seconds, and it takes
30 rounds of training, showing that overtraining can lead to about 4 minutes in total. The specificity index is high, the
overfitting. So deep neural networks need a sufficient amount results show that it can be used for computer-aided diagnosis
of training to show the advantages of having large internal in practical work. The recall rate of AlexNet in classifying
parameters, but also to avoid the overfitting that can occur vitreous opacity is relatively low compared to other categories,
with too much training. so the vitreous opacity and retinal detachment samples and
Tab III shows the precision, recall and specificity calculated vitreous hemorrhage samples and vitreous opacity samples still
by AlexNet using the Adam optimization algorithm for 50 and need to be refined in extracting features. Also deep learning
70 rounds of training and after 300 rounds of training with the neural networks perform better when trained on large data sets,
SGD algorithm for testing. According to the comparison, it can as the number of samples for individual diseases is limited, so
be seen that using the Adam optimization algorithm to train the accuracy is likely to continue to increase with the amount
50 rounds of training time is the shortest, and the classification of data. The above conclusions suggest that deep learning
ability for 5 types of data is the most comprehensive and the neural networks have very great potential to assist in diagnosis.

5
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.
 TABLE III
T HE PRECISION , RECALL AND SPECIFICITY OF A LEX N ET FOR TESTING AFTER TRAINING WITH DIFFERENT OPTIMIZATION ALGORITHMS FOR DIFFERENT
EPOCHS

Adam SGD
Epochs=50 Epochs=70 Epochs=300
Precision Recall Specificity Precision Recall Specificity Precision Recall Specificity
normal 0.897 0.867 0.982 0.737 0.933 0.939 0.667 0.933 0.915
asteroid hyalosis 0.800 0.800 0.911 0.820 0.683 0.933 0.797 0.783 0.911
vitreous opacities 0.769 0.556 0.983 0.833 0.556 0.989 0.583 0.389 0.972
vitreous opacity,
0.703 0.750 0.859 0.643 0.750 0.815 0.684 0.650 0.867
retinal detachment
vitreous hemorrhage 0.828 0.889 0.970 0.840 0.778 0.976 0.840 0.778 0.976

R EFERENCES [12] R. Rabinowitz, R. Yagev, A. Shoham, and T. Lifshitz,

“Comparison between clinical and ultrasound findings in
[1] E. C. O’Neil, M. Henderson, M. Massaro-Giordano, and patients with vitreous hemorrhage,” Eye, vol. 18, no. 3,
V. Y. Bunya, “Advances in dry eye disease treatment,” pp. 253–256, Mar. 2004.
Current Opinion in Ophthalmology, vol. 30, no. 3, pp. [13] F. Zhang, Z. Li, B. Zhang, H. Du, B. Wang, and X.
166–178, May 2019. Zhang, “Multi-modal deep learning model for auxiliary
[2] C. Matossian, M. McDonald, K. E. Donaldson, K. K. diagnosis of Alzheimer’s disease,” Neurocomputing, vol.
Nichols, S. MacIver, and P. K. Gupta, “Dry Eye Disease: 361, pp. 185–195, Oct. 2019.
Consideration for Women’s Health,” Journal of Women’s [14] K. Hu et al., “Deep supervised learning using self-
Health, vol. 28, no. 4, pp. 502–514, Apr. 2019. adaptive auxiliary loss for COVID-19 diagnosis from
[3] R. Sarki, K. Ahmed, H. Wang, and Y. Zhang, “Automatic imbalanced CT images,” Neurocomputing, vol. 458, pp.
Detection of Diabetic Eye Disease Through Deep Learn- 232–245, Oct. 2021.
ing Using Fundus Images: A Survey,” IEEE Access, vol. [15] Y. Guo and C. Chen, “An Orthopedic Auxiliary Diag-
8, pp. 151133–151149, 2020. nosis System Based on Image Recognition Technology,”
[4] T. Nazir, A. Irtaza, A. Javed, H. Malik, D. Hussain, and Journal of Healthcare Engineering, vol. 2021, pp. 1–14,
R. A. Naqvi, “Retinal Image Analysis for Diabetes-Based Nov. 2021.
Eye Disease Detection Using Deep Learning,” Applied [16] D. S. W. Ting et al., “Development and Validation
Sciences, vol. 10, no. 18, p. 6185, Sep. 2020. of a Deep Learning System for Diabetic Retinopathy
[5] P. Harasymowycz et al., “Medical Management of Glau- and Related Eye Diseases Using Retinal Images From
coma in the 21st Century from a Canadian Perspective,” Multiethnic Populations With Diabetes,” JAMA, vol. 318,
Journal of Ophthalmology, vol. 2016, pp. 1–22, 2016. no. 22, p. 2211, Dec. 2017.
[6] P.-F. Lin, H. Naveed, M. Eleftheriadou, R. Purbrick, M. [17] Y. Peng et al., “DeepSeeNet: A Deep Learning Model for
Zarei Ghanavati, and C. Liu, “Cataract service redesign Automated Classification of Patient-based Age-related
in the post-COVID-19 era,” Br J Ophthalmol, vol. 105, Macular Degeneration Severity from Color Fundus Pho-
no. 6, pp. 745–750, Jun. 2021. tographs,” Ophthalmology, vol. 126, no. 4, pp. 565–575,
[7] W. L. Wong et al., “Global prevalence of age-related Apr. 2019.
macular degeneration and disease burden projection for [18] F. A. Medeiros, A. A. Jammal, and A. C. Thompson,
2020 and 2040: a systematic review and meta-analysis,” “From Machine to Machine,” Ophthalmology, vol. 126,
The Lancet Global Health, vol. 2, no. 2, pp. e106–e116, no. 4, pp. 513–521, Apr. 2019.
Feb. 2014. [19] L. S.K., S. N. Mohanty, S. K., A. N., and G. Ramirez,
[8] T. J. Heesterbeek, L. Lorés-Motta, C. B. Hoyng, Y. T. “Optimal deep learning model for classification of lung
E. Lechanteur, and A. I. den Hollander, “Risk factors for cancer on CT images,” Future Generation Computer
progression of age-related macular degeneration,” Oph- Systems, vol. 92, pp. 374–382, Mar. 2019.
thalmic Physiologic Optic, vol. 40, no. 2, pp. 140–170, [20] H. Ali Khan et al., “Brain tumor classification in MRI
Mar. 2020. image using convolutional neural network,” Mathemat-
[9] S. E. Coupland, “The pathologist’s perspective on vitre- ical Biosciences and Engineering, vol. 17, no. 5, pp.
ous opacities,” Eye, vol. 22, no. 10, pp. 1318–1329, Oct. 6203–6216, 2020.
2008. [21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet
[10] N. G. Ghazi and W. R. Green, “Pathology and patho- classification with deep convolutional neural networks,”
genesis of retinal detachment,” Eye, vol. 16, no. 4, pp. Commun. ACM, vol. 60, no. 6, pp. 84–90, May 2017.
411–421, Jul. 2002.
[11] P. F. Kador and M. Wyman, “Asteroid hyalosis: patho-
genesis and prospects for prevention,” Eye, vol. 22, no.
10, pp. 1278–1285, Oct. 2008.

6
Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 07,2023 at 06:42:12 UTC from IEEE Xplore. Restrictions apply.