Content uploaded by Shadman Q. Salih
Author content
All content in this area was uploaded by Shadman Q. Salih on Jun 03, 2021
Content may be subject to copyright.
Kurdistan Journal of Applied Research (KJAR)
Print-ISSN: 2411-7684 | Electronic-ISSN: 2411-7706
Website: Kjar.spu.edu.iq | Email: kjar@spu.edu.iq
Modified AlexNet Convolution
Neural Network For Covid-19
Detection Using Chest X-ray
Images
Shadman Q. Salih
Hawre Kh. Abdulla
Database Technology Department
Database Technology Department
Technical College of Informatics
Technical College of Informatics
Sulaimani Polytechnic University
Sulaimani Polytechnic University
Sulaimani, Iraq
Sulaimani, Iraq
shadman.salih@spu.edu.iq
hawre.abdulla@spu.edu.iq
Zanear Sh. Ahmed
Nigar M. Shafiq Surameery
Information Technology Department
Building and Construction Engineering Department
Erbil Technical institute
College of Engineering
Erbil Polytechnic University
University of Garmian
Erbil, Iraq
Kalar, Sulaimani, Iraq
zanear.ahmad@epu.edu.iq
nigar.mahmoud@garmian.edu.krd
Rasber Dh. Rashid
Software Engineering Department
Faculty of Engineering
Koya University
Koya, Erbil, Iraq
rasber.rashid@koyauniversity.org
Article Info
ABSTRACT
Special Issue on
Coronavirus (COVID-19)
DOI:
10.24017/covid.14
Article history:
Received: 29 May 2020
Accepted 03 June 2020
First outbreak of COVID-19 was in the city of Wuhan in China
in Dec.2019 and then it becomes a pandemic disease all around
the world. World Health Organization (WHO) confirmed more
than 5.5 million cases and 341,155 deaths from the disease till
the time of writing this paper. This new worldwide disease forced
researchers to make more precise way to diagnose COVID-19. In
the last decade, medical imaging techniques show its efficiency
in helping radiologists to detect and diagnose the diseases. Deep
learning and transfer learning algorithms are good techniques to
detect disease from different image source types such as X-Ray
and CT scan images. In this work we used a deep learning
technique based on Convolution Neural Network (CNN) to detect
and diagnose COVID-19 disease using Chest X-ray images.
Moreover, the modified AlexNet architecture is proposed in
different scenarios were differing from each other in terms of the
type of the pooling layers and/or the number of the neurons that
have used in the second fully connected layer. The used chest X-
ray images are gathered from two COVID-19 X-ray image
datasets and one dataset includes large number of normal and
pneumonia X-ray images. With the proposed models we obtained
Keywords:
COVID-19,
Chest X-Ray Images,
CNN,
AlexNet,
Deep Learning.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 120
the same or even better result than the original AlexNet with
having a smaller number of neurons in the second fully
connected layer.
Copyright © 2020 Kurdistan Journal of Applied Research.
All rights reserved.
1. INTRODUCTION
At the end of 2019 an emerged novel Corona Virus has declared in Wuhan, Hubei
Province, China after investigating the unidentified pneumonia cases which caused
by a disease officially is called COVID-19 [1, 2]. The virus is rapidly spread in other
countries all over the world and millions of people have infected by the disease.
Consequently, World Health Organization announced COVID-19 as a global health
emergency [3]. In addition, a report categorizes the cases of COVID-19 into two
parts which are active and closed, under the active cases 98 percent infected patients
in mild condition [3]. The prodrome of the virus includes high temperature, cough,
dyspnea also it causes middle east respiratory syndrome (MERS-COV) and sever
acute respiratory syndrome (SARS) that is known as (SARS-COV) [4, 5].
As the virus is transmitted human-to-human, accordingly, it is strongly
recommended to halt direct contact with others and put social distancing to prevent
the distribution the disease from one person to another [6]. Hence, identifying those
who are infected by the virus in early stage is very important to save their lives. The
diagnoses of the disease become an issue for the hospitals due to the shortage of the
test kits, limited finding rate and it takes time as well as using reverse transcription
polymerase chain reaction (RT-PCR) to determine the infected patient is not accurate
as per the rate is from 30 to 60 percent, which means a large number of COVID-19
patients have not detected and will not take proper treatment and isolation.
Moreover, inaccurate test helps the disease to speared faster as it transmits person to
person [7]. Furthermore, the X-ray and (CT) scan of chest are also used to establish
the MERS- COV and SARS-COV in early stage of the disease. In this technology
era, enhancing the CT and X-ray based on artificial intelligence (AI) tools may help
clinicians to detect and monitor the virus faster and more efficiently [7].
Additionally, there are a bunch of researches [1, 7, 8, 9] used deep learning-based
techniques to detect the infected patients which help to reduce the load on doctors
and hospitals to identify the disease and make the process faster in compare with
RT-PCR.
Accordingly, we have studied available Convolution Neural Network (CNN) in the
literature named AlexNet. Based on this CNN architecture, we proposed a modified
version of AlexNet to train and test the available X-ray images for the study. The
aim is to determine and identify the COVID-19 cases from other two class cases
named Normal and Abnormal cases.
2. RELATED WORK
Nowadays, due to rapidly spreading the COVID-19 disease, more attention goes to
the field of detecting the disease by the researchers. In [10] the researchers proposed
two different techniques to evaluate lung pneumonia and cancer namely modified
AlexNet (MAN), moreover, in the study, the principle component
analysis (PCA) used by combination with MAN. In the first proposed technique, the
classification result improved by using five blocks from 1-5 with two fully
connected layers, so, each initial block consists of Convolution, max pooling, ReLU
and normalization. At the end, the SoftMax layer has been replaced by Support
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 121
Vector Machin (SVM). The proposed technique was tested on two different chest
image types X-ray and lung CT scans with including only two types of classes
Normal and Abnormal for each of them. The accuracy obtained was 96% and
97.27% for X-ray and CT scans respectively. According to [11] the new CNN
proposed including 16 layers only, the method designed to distinguish the Normal
and COVID-19 classes only. The algorithm tested on both X-ray and CT scan
images. It is worth to mention that the proposed technique was obtained good
accuracy performance with declaring that they used a very small number of cases.
In [12] three different scenarios have been tested with using three different available
CNN architecture that are: AlexNet, googlenet, and resnet18. The idea of
transferring learning (TL) is used for all of the three scenarios. The scenarios are
different from each other in the number of classes that have used. In the first one,
four different classes (Normal, COVID-19, Viral, and Bacteria) have been used with
googlenet that resulting in having the accuracy rate of 80.6 %. In the second
scenario, three different classes have been used (Normal, COVID -19, and
Pneumonia) and obtaining the accuracy rate about 85.2 with using AlexNet CNN
architecture, while in the third scenario, only two classes have been used (Normal
and COVID-19) and the algorithm guaranteed 100% of accuracy with using
googlenet architecture. Jianping Zhang et al. [13] presents a new deep learning
model for detecting COVID-19 that contain three components, they used 100 X-Ray
images as a Covid-19 from 70 patients and 1431 X-Ray images as non covid-19. The
result shows 96% for detecting covid-19 and 70.65% for detecting non-covid-19.
Wang et al. [14] proposed a new CNN architecture named as COVID-Net. They
generate the COVIDx dataset by combining and modifying two different datasets,
“covid-19 image data collection” and “RSNA Pneumonia Detection Challenge”
dataset, and it contain 16,756 chest radiography across 13,645 patient cases from the
two different dataset that have been mentioned above. The accuracy of testing
COVIDx is 92.4%.
From the proposed works available in the literature, it is clear that a good accuracy
performance can be obtained when they tested and applied over only two classes
(Normal and COVID-19). The challenge is to have more than two classes, therefore
in this work, we aimed to design and test a new CNN architecture to be suitable for
more than two classes. Moreover, we aim to design an architecture to distinguish
three different classes (Normal, COVID-19, and Pneumonia (Bacteria and other
(Non COVID-19) Viral).
3. AlexNet Architecture Analysis
The history of deep CNN starts with the LeNet which originally designed for the
purpose of handwritten digit recognition. The LeNet not performed well due to its
limitation to the handwritten digit recognition tasks [15]. Thus, the AlexNet is
supposed to be the first deep CNN architecture because of its’ outstanding results for
the classifications and recognition tasks performed on images [16]. In early 2000, the
learning capacity of the deep CNN architectures was restricted to small size because
of the hardware limitations. Thus, in order to overcome the limitations of the
hardware and getting the whole capacity of deep CNN, AlexNet was trained on two
“NVIDIA GTX 580” GPUs in parallel.
In order to make the CNN applicable for different types of images, in AlexNet, the
CNN depth was extended from only (5) layers in the LetNet CNN to (8) layers. The
layers are: five convolutional layers, two fully connected hidden layers, and one
fully connected output layer, as it shown in Fig1.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 122
Figure 1: AlexNet Architecture Design
Although increasing the depth of the architecture will improve the generalization for
different image resolutions, on the other hand, it results in the overfitting issue. To
fix this issue, the dropout algorithm has been used, in which the algorithm randomly
removes some transformational units during the training process to motivate the
model to learn more better features [17, 18]. In addition, the local response
normalization and overlapping subsampling have been applied to reduce the
overfitting in order to improve the generalization. Moreover, in AlexNet, the sigmoid
activation function has been replaced by the ReLU function in order to improve the
convergence rate by reducing the problem of vanishing gradient [19, 20].
As it is clear in Fig1, in AlexNet, the input image must be augmented to a fix size of
227x227x3, in the first layer the 96 convolution filter with window shape of size 11
× 11 applied, whereas in the second layer 256 convolution filter with window shape
size of 5×5 applied, followed by 384, 384, and 256 convolution filter with window
size of 3×3 for the rest of the other three layers applied. Moreover, after the first,
second, and the last convolutional layers, the network contains 3 × 3 maximum
pooling layers with stride of 2. In addition to this five convolution layers, there are
two fully-connected layers with 4096 neuron outputs after the fifth convolutional
layer. This is followed by one fully connected output layer placed at the end of the
network which is originally having 1000 output classes. Finally, Dropout, ReLU and
preprocessing represent the important key steps in achieving excellent performance
in computer vision tasks.
In this paper we used the original AlexNet as a transfer learning mode by modifying
the last output layer to be suitable to the number of the X-ray classes that have been
used in these experiments. We replaced the 1000 classes that the original AlexNet
have, with only 3 classes because we tested only three classes (COVID-19, Normal,
and Pneumonia ) in this study.
4. Proposed Models (Modified AlexNet)
In this research, we proposed a new CNN model by modifying the architecture
layers of the AlexNet Convolution Neural Network to see how the classification
result will be affected and improved. In fact, we selected the AlexNet architecture
because the structure of the AlexNet is the simplest among other CNN systems and
need suitable time and efficiency for learning phase especially we are using the
normal personal computer with normal CPU based systems. The proposed model(s)
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 123
has been developed using software package in MATLAB program. Another reason
of using this architecture is high performance and fast training with very few epochs.
We propose four different modified models based on the original AlexNet
architecture. In all of the four models, the structure of the original AlexNet remain
with having modifications on some layers. The four modified models can be
described as follows:
Model 1: All of the 3x3 Max Pooling layers that are available in the original AlexNet
have been replaced with corresponding 3x3 Average Pooling layers with remaining
the stride value as its which is 2. See figure 2.
Figure 2: Model 1 of Modified AlexNet Architecture
Model 2: The 3x3 Max Pooling layers that are available after the first and second
convolution layers are replaced with corresponding 3x3 Average Pooling layers with
remaining the Max Pooling after the fifth convolution without any changes (i.e Max
Pooling), again the stride value remain as it is which is 2 in all of the three Pooling
layers. See figure 3.
Figure 3: Model 2 of Modified AlexNet Architecture
Model 3: As shown in Figure 4, the number of the neurons in the second fully
connected layer has been reduced form 4096 neurons to 2096 neurons. All other
available layers are remained as the original AlexNet. This model designed to see
how reducing the number of neurons affects the results besides increasing the
efficiency.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 124
Figure 4: Model 3 of Modified AlexNet Architecture
Model 4: In this model, we combined the two previously proposed models’ number 2
and 3 as shown in figure 5. This means that we replaced the Max Pooling layers after
the first and second convolution layers with Average Pooling layers and the Max
Pooling after the fifth convolution remain as it is, in addition, we reduced the number
of the neurons in the second fully connected layer form 4096 neurons to 2096
neurons.
Figure 5: Model 4 of Modified AlexNet Architecture
5. Experimental Data and Setup
Recently, Covid-19 virus has affected a global population. The virus continues to have a huge
influence on the welfare and health of the people around the world. The essential step in
confrontation this risky virus is radiology examination. This could help the doctors to
diagnose the disease and gave the rapid treatment to the patient in a precise way [21]. These
radiographs are then stored as X-ray or CT-scan images in various public available sources
such as COVIDx dataset by Cohen [22], a set of 6290 chest X-ray images of non-Covid-19
viral cases, bacterial infections and confirmed normal images by Kermany et al authors [23].
Another source of dataset is a collection of 79 Covid-19 chest X-ray confirmed radiographs
that was recently made public by some websites such as Radiopedia, “SIRM (Italian Society
of Medical and Interventional Radiology)” and “RSNA (Radiological Society of North
America)” [24].
The main reason of these huge-data collection and made them publicly available
sources is to provide a free access mechanism for the researchers to be able to use
them in their research papers as well as helping them to contribute and share their
de-identified COVID-19 chest X-ray data to the project. Hence, in this section, we
will explain the details of the datasets that have been used for the train and test
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 125
process in this paper for purpose of Covid-19 and other cases detections such as
Normal, Bacterial and Viral.
5.1 First Dataset (COVIDx)
The COVIDx dataset recently made available at [22] by the authors of Covid-Net,
which is one of the first open source network. It is a deep CNN, which is made
publicly available to detect corona virus from CT-Scan or X-ray images. The dataset
is consists of 13,975 X-ray and CT scans images taken from 13,870 patient cases. In
the recent version of the COVIDx dataset, there was a total of 111 Covid-19 X-ray
images from approved corona virus patients. In this paper work, we used all 111
chest X-ray positive Covid-19 images.
5.2 Second Dataset
This dataset was collected and made publicly available by Kermany and et. al. It
contains 6290 chest X-ray images from numerous patients with normal infections,
bacterial and other viral (non-Covid-19) cases [23]. In this huge data set
images, only 1108 chest radiography images has been chosen to train, validation and
test. More especially, we used 554 images from normal cases, 277 images from
bacterial cases and 277 images from viral (non-Covid-19) cases.
5.3 Third Dataset
Third dataset contains only 79 positive chest X-ray images. It was recently composed
from Radiological Society of North America (RSNA), Radiopedia, and
Italian Society of Medical and Interventional Radiology (SIRM) websites. This
collection of confirmed positive images are publicly available in [24]. In this dataset
only 43 chest X-ray positive Covid-19 images are nominated to train, validation and
test sets. Figure 6 shows random samples of the images for all classes.
Figure 6: Chest X-ray samples
As a result, from the three above mentioned databases, we used in total 154 (100 for
train and 54 for test) X-ray images of confirmed Covid-19 cases, 554 (500 for train
and 54 for test) X-ray images of Normal cases, and 554 (500 for train and 54 for test)
X-ray images of Abnormal (pneumonia) cases. The 554 cases of Abnormal
(pneumonia) cases originally taken from 277 cases of bacteria and 277 cases of viral
(Non-Covid-19) cases.
It is valuable to be mentioned that the number of Covid-19 X-ray images is
considerably lower than other classes that lead to imbalance classification issue.
However, to reduce the impact of imbalance classification we tried to not use all
available cases of normal and pneumonia and only the above mentioned number of
cases have been used.
To investigate and test the mentioned proposed models in the previous section as
well as the original AlexNet architecture, we used the above-classified dataset. The
same train and test sets have been used with the original AlexNet architecture as well
as the proposed modified models. See the next section for the experimental results
obtained.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 126
6. Experimental Results
The process of utilizing learning weights or using obtained data from determining
and solving one issue to another correlated issue is known as Transfer Learning
(TL). Weights enhanced from the AlexNet architecture on X-ray image dataset used
in TL model. In this paper, TL approach has been adopted to investigate the
performance discussion of original AlexNet CNN architecture and then compared
with Modified AlexNet architectures. Accordingly, in all of the layers, the weights
are re-trained on our chest X-ray images. The resizing process to the appropriate
dimensions of the original AlexNet and Modified AlexNet architectures has been
done on the entire set of the images from the training and the testing sets. According
to the former researches in literature review, no preprocessing applied to input
images, hence the same norm has been followed.
For all CNN architectures, the training parameters in TL are as follows: Mini Batch
Size =15, Number of Epochs =10, Initial Learn Rate = 3e-4. The software package
MATAB is used for conducting all experimental results. Two metrics were
recorded in order to measure the performance of CNN classification, which
are specificity and sensitivity. For calculating the mentioned metrics, we recorded
the following classifications test measurements:
TP – True Positive: represents the number of the correctly recognized positive cases.
TN – True Negative: represents the number of the correctly recognized negative
cases.
FP – False Positive: represents the number of the incorrectly identified negative
cases.
FN – False Negative: represents the number of the incorrect classification of disease.
The amount of the unhealthy cases appropriately identified by CNN models is known
as Sensitivity while the Specificity measures the amount of negative cases properly
detected as healthy by CNNs. For the purpose of comparisons as well as testing the
performance, we calculate the sensitivity and specificity of the original and all of the
four modified proposed models. Table 1 shows the obtained results from all
experiments conducted. Moreover, the table shows the sensitivity and specificity of
all of the three individual classes Covid-19, Normal, and Abnormal (pneumonia). To
see individual classified cases for every single experiment, we calculated the
confusion matrix among the classes, figure 7 shows the confusion matrix of the
original AlexNet while figure 8 shows the confusion matrix of obtained results from
all of the four proposed models.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 127
Table 1: Testing result for Original and Modified AlexNet architecture
Figure 7: Confusion matrix of original AlexNet
CNN Architecture
Class
Sensitivity
Specificity
Original AlexNet
Covid-19
98.14
99.04
Normal
72.22
87.60
Abnormal
90.74
94.89
Proposed (Model 1)
Covid-19
79.62
90.67
Normal
83.33
91.42
Abnormal
94.44
97.02
Proposed (Model 2)
Covid-19
90.74
95.57
Normal
81.48
90.90
Abnormal
94.44
97.02
Proposed (Model 3)
Covid-19
90.74
95.32
Normal
83.33
91.74
Abnormal
90.74
95.37
Proposed (Model 4)
Covid-19
94.44
97.27
Normal
81.48
90.99
Abnormal
90.74
95.14
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 128
(a) Model 1
(b) Model 2
(c) Model 3
(d) Model 4
Figure 8: Confusion matrix of Proposed Models
From table1 and figure7, we can notice that the obtained results from the
original AlexNet for recognising Covid-19 cases is promising but there is
a problem with the results of detecting the normal case. There are 15 cases out of 54
cases are misclassified. Although our aim is to detect the Covid-19 cases accurately,
there is a problem when the normal cases are classified as affected cases especially in
the medical side. Therefore, the proposed models are designed to decrease
incorrectly classified cases. For example, when we compare the obtained results
from model1 with the original AlexNet results, we can see that the sensitivity and the
specificity of both of the normal and abnormal classes are increased with decreasing
the sensitivity and the specificity of the Covid-19 class. In model2, after we
mixed the average and max pooling, we obtained higher detection degree for Covid-
19 classes compared to model1. Again having better detection degree for both of the
normal and abnormal classes comparing to the original AlexNet. Although reducing
the number of neurons of the original AlexNet such as in model3 increased the
efficiency of executing the training phase, the same conclusion as model2 can be
obtained when the results are compared with the original AlexNet. Finally when we
combine both of the ideas in model2 and model3, we can guaranteed the efficiency
as well as the better detection rate of Covid-19 classes if compared with all other
three proposed models and losing a small degree of detection if compared with
the original AlexNet, taken in account that model4 have better detection for both
other classes normal and abnormal when compared with the original AlexNet.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 129
If we take overall, the detection rate and counting the corrected number of cases for all of the
three classes in each architecture, we can say that out of 162 cases the
original AlexNet classified 141 cases correctly while the correctly classified cases are 139,
144, 143, and 144 for our four proposed model1, model2, model 3, and model 4 respectively.
7. Conclusion and Future Work
Many researchers have exhibited that the already existing convolutional neural network
architectures have a significant ability of detecting COVID-19 from X-ray or CT-Scan images
in an accurate manner and provide excellent results [11, 25, 26]. However, their results can be
improved by modifying their structure in one or more of their layers or their other features. In
this paper we used original AlexNet architecture for Covid-19 detection, Pneumonia, and
Normal cases. This architecture consists of eight layers: five convolutional layers and three
fully-connected layers with some of the features that are new approaches to convolutional
neural networks and have made the AlexNet special [27].
The aim of this paper was to introduce four effectives Modified AlexNet CNN models that
can outperform pre-trained AlexNet for the purpose of helping radiologists to recognize and
detect correctly the classes from chest X-ray images. These models were differing from each
other in terms of the type of the pooling layers and/or the number of the neurons that have
used in the second fully connected layer. Moreover, the images that have been used in this
study are collected from three different sources. So, we collectively make the use of these X-
ray images to be publicly available for the researchers. From the obtained results, it became
obvious that in case of a small dataset and unbalanced number of the X-ray images, we can
conclude that with replacing all of the max pooling layers with average Pooling layers, the
result has been enhanced even though we have a noticeable reduction in the sensitivity of the
Covid-19 detection result. This issue was reduced by replacing just the first two max pooling
layers with average Pooling layers, which results in the improvement of the Covid-19 case
sensitivity from 79.6 to 81.46. On the other hand, with reducing the number of the neurons by
2000, we have got the same or even better result than the original AlexNet with a little bit
reduction in the Covid-19 case sensitivity. Finally, one open door for any experiments further
is to apply the same models with a large-scale dataset where its images can be collected from
different countries. In addition, more than three classes can be tested with keeping in mind the
balancing issue. Moreover, trying to apply the same modifications on deeper transfer networks
such as Densenet, and Inceptionresnet can be another future research direction.
REFERENCE
[1] Y. Song, S. Zheng, X. Zhang, X. Zhang, Z. Huang,J. et al., "Deep learning Enables Accurate Diagnosis of
Novel Coronavirus (COVID-19) with CT images," p. 10, 25 February 2020.
[2] W. H. Organization, Health, [Online]. Available: https://www.who.int/dg/speeches/detail/who-director-
general-s-opening-remarks-at-the-media-briefing-on-covid-19---11-may-2020. [Accessed 12 May 2020].
[3] "Worldometers," [Online]. Available: https://www.worldometers.info/coronavirus/. [Accessed 27 May
2020].
[4] Q. Li, M. Med, et al, "Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus–Infected
Pneumonia," The new england journal of medicine, vol. 382, p. 13, 26-March-2020.
[5] S. Stoecklin, P. Rolland, et al, "First cases of coronavirus disease 2019 (COVID-19) in France: surveillance,
investigations and control measures," Euro surveillance, vol. 2000094, January 2020.
[6] "How to Protect Yourself & Others," Centers for Disease Control and Prevention (CDC), [Online].
Available: https://www.cdc.gov/coronavirus/2019-ncov/prevent-getting-
sick/prevention.html?CDC_AA_refVal=https%3A%2F%2Fwww.cdc.gov%2Fcoronavirus%2F2019-
ncov%2Fprepare%2Fprevention.html. [Accessed 13 May 2020].
[7] S. Khobahi, Ch. Agarwal and M. Soltanalian, "CoroNet: A Deep Network Architecture for Semi-Supervised
Task-Based Identification of COVID-19 from Chest X-ray Images," medRxiv , 2020.
[8] A. Narin, C. Kaya and Z. Pamuk, "Automatic Detection of Coronavirus Disease (COVID-19) Using X-ray
Images and Deep Convolutional Neural Networks," eprint arXiv:2003.10849, p. 17, March 2020.
[9] X. Xu, X. Jiang, et al., "Deep Learning System to Screen Coronavirus Disease 2019 Pneumonia," arXiv, vol.
2002.09334, p. 29, 2020.
[10] A. Bhandary, G. Prabhu, et al, "Deep-Learning Framework to Detect Lung Abnormality – A study with
Chest X-Ray and Lung CT Scan Images," elsevier Pattern Recogn Lett , vol. 129, pp. 271-278, 2020.
Kurdistan Journal of Applied Research | Special Issue on Coronavirus (COVID-19)| 130
[11] H. S. Maghdid, A. T. Asaad, et al., "Diagnosing COVID-19 Pneumonia from X-Ray and CT Images using
Deep Learning and Transfer Learning Algorithms," arxiv, vol. 2004.00038, p. 8, 2020.
[12] M. Loey, F. Smarandache, et al, "Within the Lack of Chest COVID-19 X-ray Dataset: A Novel Detection
Model Based on GAN and Deep Transfer Learning," Symmetry, vol. 12, no. 651, p. 19, 2020.
[13] J. Zhang, Y. Xie, et al., "COVID-19 Screening on Chest X-ray Images Using Deep Learning based Anomaly
Detection," arXiv, vol. 2003.12338, p. 6, 2020.
[14] K. Hammoudi, H. Benhabiles, et al., "Deep Learning on Chest X-ray Images to Detect and Evaluate
Pneumonia Cases at the Era of COVID-19," arXiv, vol. 2004.03399, p. 6, 2020.
[15] Y. LeCun, L. D. Jackel, et al, "Learning algorithms for classification: A comparison on handwritten digit
recognition," Neural networks: the statistical mechanics perspective, vol. 261, p. 16, 1995.
[16] A. Krizhevsky, I. Sutskever, et al., "Imagenet classification with deep convolutional neural networks,"
in Advances in Neural Information Processing Systems 25 (NIPS 2012), 2012.
[17] N. Srivastava, G. Hinton, et al., "Dropout: A Simple Way to Prevent Neural Networks from
Overfitting," Journal of Machine Learning Research 15, Vols. 1929-1958, p. 30, 2014.
[18] G. E. Dahl, T. N. Sainath, G. E. Hinton, "IMPROVING DEEP NEURAL NETWORKS FOR LVCSR
USING RECTIFIED LINEAR UNITS AND DROPOUT," in 2013 IEEE international conference on
acoustics, speech and signal processing, 2013.
[19] S. Hochreiter, "The vanishing gradient problem during learning recurrent neural nets and problem
solutions," International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 06, no. 02,
pp. 107-116, 1998.
[20] V. Nair, G. E. Hinton, "Rectified Linear Units Improve Restricted Boltzmann Machines," in Proceedings of
the 27th international conference on machine learning (ICML-10), 2010.
[21] L. Wang, Z. Q. Lin, A. Wong, "COVID-Net: A Tailored Deep Convolutional Neural Network Design for
Detection of COVID-19 Cases from Chest X-Ray Images," arXiv, vol. 2003.09871, p. 12, 2020.
[22] J. P.Cohen, P. Morrison, L. Dao, "An open database of COVID-19 cases with chest X-ray or CT
images.," arXiv, vol. 2003.11597, p. 4, Mar 2020.
[23] D. S.Kermany, M. Goldbaum, W. Cai , et al., "Identifying Medical Diagnoses and Treatable Diseases by
Image-Based Deep Learning," Cellpress, vol. 172, no. 5, pp. 1122-1131, 2018.
[24] Larxel, "COVID-19 X rays," Kaggle, [Online]. Available: https://www.kaggle.com/andrewmvd/convid19-
X-rays. [Accessed 27 May 2020].
[25] M. Farooq, A. Hafeez, "COVID-ResNet: A Deep Learning Framework for Screening of COVID19 from
Radiographs," arXiv:, vol. 2003.14395 , p. 6, 2020.
[26] T. Majeed, R. Rashid, et al., "Covid-19 Detection using CNN Transfer Learning from X-ray
Images," medRxiv, vol. 20098954, p. 10, 2020.
[27] J. Wei, "AlexNet: The Architecture that Challenged CNNs," Towards Data Science, 3 Jul 2019. [Online].
Available: https://towardsdatascience.com/alexnet-the-architecture-that-challenged-cnns-e406d5297951.
[Accessed 25 May 2020].
[28] L. Liao, Y. Zhao, et al., "Finding the Secret of CNN Parameter Layout under Strict Size Constraint,"
in Proceedings of the 25th ACM international conference on Multimedia, 2017.