https://doi.or g/10.31449/inf.v48i5.5424 Informatica 48 (2024) 55–62 55
Efficient COVID-19 Pr ediction by Merging V arious Deep Learning
Ar chitectur es
Zakariya A. Oraibi
1, ∗ , Safaa Albasri
2
1
Department of Computer Science, College of Education for Pure Sciences, University of Basrah, Basrah, Iraq
2
Electrical Engineering Department, College of Engineering, Mustansiriyah University , Baghdad, Iraq
E-mail: zakaria_au@uobasrah.edu.iq, asafaa@uomustansiriyah.edu.iq
∗ Corresponding author
Keywords: COVID-19, deep learning, coronavirus, hybrid models, feature maps
Received: November 15, 2023
In late 2019, COVID-19 virus emer ged as a danger ous disease that led to millions of fatalities and changed
how human beings interact with each other and for ced people to wear masks with mandatory lockdown.
The ability to diagnose and detect this novel disease can help in isolating the infected patients and curb
the spr ead of the virus. Artificial intelligence techniques including machine learning and deep learning
showed huge potential in accurately classifying COVID-19 chest X-ray images. In this paper , we pr opose
to combine the featur e maps of multiple powerful CNN models (Xception, VGG-16, VGG-19) using the
rule of sum. Each of these models is trained fr om scratch and tested on the given test images. The dataset
was collected fr om a lar ge public r epository of COVID images with thr ee classes: COVID, Normal, and
Pneumonia. During experiments, data augmentation is also applied to pr o vide mor e training samples.
Experimental r esults show that combining multiple models impr ove the classification accuracy and achieve
better performance than standalone models. An accuracy of 97.91% was achieved using a combination of
thr ee models which outperforms state-of-the-art techniques.
Povzetek: Pr edlagana je združitev modelov CNN (Xception, VGG-16, VGG-19) za izboljšanje klasifikacije
COVID-19 na osnovi slik prsnega koša.
1 Intr oduction
The pandemic that started more than three years ago caused
by coronavirus (SARS-COV -2) led to changing the world
drastically from forcing people to wear masks in public ar -
eas to driving nations to impose a complete lockdown to
prevent its spread [1, 2]. Infected people c an easily spread
the virus through sneeze, cough, or even by speaking. The
major concern related to COVID-19 is that it af fects human
lung by damaging the tissues. Symptoms at early stages of
infection could be similar to the usual flu infection and may
include fever , throat pain, and headache [3]. It is estimated
that the virus infected hundreds of millions of people and
caused millions of fatalities until a vaccine was developed
in 2021 which limited the danger of the infection and re-
stored life to normality [4].
Since COVID-19 virus is very contagious, it should be
detected early to help isolating and curing the patient. The
traditional technique of diagnosing coronavirus involves
using Polymerase Chain Reaction (PCR) [1 1]. However ,
using PCR to perform such test is also challenging due its
low sensitivity leading to dif ficulty in detecting positive
COVID-19 cases. In addition, PCR needs more time to
obtain the results with limited availability of kits in clin-
ics and hospitals [12]. As a result, chest X-ray images
can be used for screening by employing machine learning
techniques to automate the process of detecting COVID-19
cases. These techniques range from traditional image clas-
sifiers like SVM, kNN, and RFs to deep learning models
like VGG-16, VGG-19, Xception, Inception, MobileNet,
and ResNet [13, 14].
In literature, Kaur et al. [5] presented a hybrid deep
learning model to classify COVID-19 images into three
and four classes. InceptionV4 model is applied to extract
image features then, SVM is applied to classify COVID-
19 samples. Their method achieved a high accuracy of
95.51% on three classes COVID-19 dataset. Hossain et
al. [6] applied transfer learning with fine-tuning on dif ferent
state-of-the-art deep learning models including ResNet50,
VGG-16, InceptionV3, and MobileNet-V2. A high ac-
curacy of 99.17% was achieved on a public COVID-19
dataset with two classes. Monshi et al. [7] proposed to
optimize data augmentation and training hyperparameters
using Ef ficientNetB0 model. The new model is named
CovidXrayNet which achieved a high classification accu-
racy of 95.82%. Kaya et al. [8] proposed using MobileNet
model with new fine-tuning technique to predict COVID-
19 images. They applied their approach on a lar ge dataset
with three classes and achieved high accuracy . Kausar et
al. [9] designed an end-to-end model called Style Distribu-

56 Informatica 48 (2024) 55–62 Z. A. Oraibi et al.
T able 1: Summary of related work.
T echnique Y ear Dataset Performance
Features extraction using Incep-
tionV4 + SVM Classifier [5]
2021 1900 Chest XR Detection accuracy = 95.51%
T ransfer learning + Fine-tuning
CNN architectures [6]
2022 COVID-19 Radiog-
raphy dataset
Classification accuracy = 99.17 (two
classes)%
Ensemble of pre-trained models
[18]
2021 2326 X-ray images Sensitivity = 90.5% , Specificity = 90.0%
Covidxraynet [7] 2021 COVIDx dataset Classification accuracy = 95.82%
MobileNet + Novel Fine-
tuning [8]
2023 9457 COVID XR Classification accuracy = 97.61%
SD-GAN [9] 2023 [A-D] datasets Accuracy = [98.7%, 99.3%, 99.1%, 98.6%]
New CNN [10] 2023 COVID-Xray-5k Sensitivity = 95% , Specificity = 99.32%
tion transfer Generative Adversarial Network (SD-GAN)
and managed to achieve high classification accuracy on X-
ray datasets. Ensemble learning techniques were used by
Aldhahi et al. [15] by introducing a new technique called
Uncertain-CAM, which provides better classification accu-
racy for COVID-19 images. In Our recent work [16, 10],
we proposed to train a new CNN architecture from scratch
to classify two COVID-19 classes. W e achieved 95.0% and
99.32% in both sensitivity and specificity rates on a stan-
dard COVID-19 dataset. In general, many methods have
been proposed in literature and the majority of them rely
on deep learning models and how to exploit multiple ar -
chitectures to improve the prediction performance. T able 1
summarizes the work in literature.
In this paper , we exploit the power of pre-trained stat-of-
the-art deep learning architectures by mer ging them to im-
prove the COVID-19 classification accuracy . These mod-
els include VGG-16, VGG-19, and Xception and were com-
bined using the rule of sum. In addition, COVID-19 images
were collected from two public image repositories where
two classes (COVID and Normal) where collected from a
huge set of images available online for research purposes.
The third class, Pneumonia, was collected from the work
of [17]. In total we collected 964 images. The proposed
technique is simple yet it improves the classification accu-
racy and provides a good baseline for future developments.
The structure of the paper is or ganized as follows: Sec-
tion 2 describes in details the methodology of hybrid CNN
models. Section 3 describes in details the dataset used in
the experiments. Section 4 lists the results of the proposed
approach. Section 5 discusses the results and finally , we
present conclusions and future work in section 6.
2 Hybrid CNN models
The methodology used in our work relies on mer ging var -
ious state-of-the-art CNN architectures. These models in-
clude Xception, VGG-16, and VGG-19 [18, 19]. The intu-
ition behind this technique is that each model proved to be
very powerful and capable of extracting the required pat-
terns from the image and perform very well in computer
vision tasks. As a result, combining the feature maps of
these models will result in an ef ficient hybrid model that
can be used to obtain high classification accuracy . Xcep-
tion model produces7× 7× 2048 feature maps on its last
layer of feature extractor . Both VGG-16 and VGG-19 pro-
duce7× 7× 512 features maps. The concatenation of these
models will result in a feature map of7× 7× 3072 in size.
Finally , a flatten layer is added with three dense layers. Fig-
ure (1) shows the pipeline of the proposed approach.
In order to apply our approach, the COVID-19 dataset of
three classes is divided into two subsets, training and test-
ing. During the training phase, images are further divided
into training and validation. In order to train the model ef-
ficiently , data augmentation has been used by modifying
the training set of images using techniques like shearing,
zooming, image flipping, and shifting. Then, each model is
imported and the trainable layers are turned of f. After that,
the concatenated model is used during training. When the
model is finished training, the accuracy is computed using
the subset of test images to report the classification results.
More details about each model used in our approach are
given in the following subsections.
2.1 Xception model
This powerful model was introduced by Google in 2017
which is inspired by Inception model [18]. In Xcep-
tion, authors proposed decoupling the correlations of cross-
channels and spatial correlations in the feature maps of
CNN. The model has 36 convolutional layers or ganized into
14 modules. Linear residual connections are used in every
module except for the first and last modules. The structure
of this model makes it very easy to modify since it is im-
plemented as a linear stack of depth-wise separable convo-
lution layers. In our work, we used the standard Xception
model and trained it from scratch. Figure (2) shows the lay-
ers of the model.

Ef ficient COVID-19 Prediction by Mer ging V arious Deep Learning… Informatica 48 (2024) 55–62 57
Figure 1: COVID-19 classification pipeline. T raining is performed after the feature maps of the three models (VGG-16,
VGG-19, and Xception) are concatenated to form a bigger input layer . Then, weights are used to predict the testing images
into three classes.
Figure 2: Architectures of VGG-16, VGG-19, and Xception Models.

58 Informatica 48 (2024) 55–62 Z. A. Oraibi et al.
2.2 VGG models
Both VGG-16 and VGG-19 CNN models were introduced
in 2014 by simonyan et al. [19]. The architecture of these
models uses a fixed size of input image (224× 224× 3 ). A
very small receptive field filter is used (3× 3 ) with a convo-
lutional stride fixed to 1. Five max-pooling layers are em-
ployed which are performed over a (2× 2 ) pixel window ,
having a stride of 2. Three fully connected layers are added
at the end of the stack of convoluitonal layers. Finally soft-
max layer is used for classification. The dif ference between
VGG-16 and VGG-19 is that the first one has 16 convolu-
tional layers while the latter has 19 convolutional layers. In
the experiments, VGG-19 always generates better results
than VGG-16 especially in the classification tasks. In our
work, we used both models and trained them from scratch.
W e don not rely on the weights trained on ImageNet dataset
or transfer learning. Figure (2) shows the structure of both
models.
3 Materials
3.1 Dataset description
The hybrid CNN model proposed in this paper is applied
on a dataset collected from a public repository in the link
below
1
. The dataset consists of three classes: COVID,
Normal, and Pneumonia. Both COVID and Normal classes
were collected from the aforementioned repository . The
third class, Pneumonia, was collected from the work of
Shastri et al. [17].
In total we have 964 images across the three classes. Im-
ages of the dataset vary in size. As a result we had to resize
all images to (224× 224× 3 ) to meet the VGG-16, VGG-
19, and Xception models needs. The number of images per
class are listed in T able 2. Sample images of each class are
shown in Figure (3).
T able 2: Number of images per class of the COVID dataset.
Class No. of Images
COVID 357
Normal 365
Pneumonia 242
4 Experimental r esults
The results of applying the hybrid model proposed in this
paper are presented in this section. Five-fold cross valida-
tion was used in the experiments. This means, 80% of the
original data were used for training and 20% were used for
1
https://www .kaggle.com/datasets/pranavraikokte/
covid19-image-dataset
T able 3: T otal number of COVID, Normal, and Pneumonia
classes training images after augmentation.
Class Original Augmented
COVID 285 1460
Normal 292 1425
Pneumonia 193 965
T able 4: Hyperparameters used in the experiments.
Hyperparameter V alue
Epochs 5
Learning Rate 0.001
Optimizer Adam
Batch Size 64
Loss Function Categorical Crossentropy
Pooling Size 2× 2
testing. The final accuracy was generated by finding the
average of the five-fold results. As we mentioned earlier ,
augmentation was used to increase the number of training
images. T able 3 shows the number of images used for train-
ing in fold 1 and the corresponding number of images re-
sulted from augmenting the training set samples.
T o evaluate the performance of our approach, five met-
rics were used: accuracy , sensitivity , specificity , precision,
and F1 score. The corresponding equations to compute
each metric are given in the equations below:
Accuracy =
TP +TN
TP +FP +FN +TN
(1)
Sensitivity =
∑ TN
∑ FP +
∑ TN
(2)
Specificity =
∑ TP
∑ TP +
∑ FN
(3)
F1 Score =
2
∑ TP
2
∑ TP +FP +FN
(4)
Precision =
∑ TP
∑ TP +
∑ FP
(5)
The combined model of Xception, VGG-16, and VGG-
19 proposed in this paper is implemented using Keras soft-
ware [20]. Google Colaboratory was used to benefit from
GPU. In the training stage, Adam optimizer is used. The
batch size used during training is fixed to 64. This is to pre-
vent any session crash. The number of epochs applied in
the training stage was fixed to 5 for all experiments. The
reason to choose such a small number of epochs is because
the network conver ges very fast. Learning rate metric was
fixed to 0.001. For the loss function, we employed cross-
entropy in the implementation. T able 4 summarizes the hy-
perparameters used in the implementation.

Ef ficient COVID-19 Prediction by Mer ging V arious Deep Learning… Informatica 48 (2024) 55–62 59
Figure 3: Sample images of the COVID dataset. images in row 1 represent COVID class. Images in row 2 represent
Normal class. Finally , images in row represent Pneumonia class.
T raining accuracy and validation accuracy conver ged
very fast after only 5 epochs as shown in Figure (4). This
is because we are using a combination of three powerful
models during training: Xception, VGG-16, and VGG-19.
In addition, the graph on the right of Figure (4) shows that
training loss and validation loss reach to 0 after only 5
epochs.
T able 5 shows the results of the experiments conducted
using our approach of combining multiple CNN models.
Xception and VGG-16 models achieved decent results of
94.31% and 93.32%. On the other hand, VGG-19 out-
performed them by achieving a high accuracy of 96.91%.
However , combining Xception and VGG-16 together out-
performed each model alone by scoring 95.85%. Finally ,
combining the three models outperformed all the previous
results with 97.91% accuracy . Other metrics including s en-
sitivity , specificity , F1 score and precision are also very
high for the triple combined model. Figure (5) shows the
confusion matrix generated from applying our hybrid ap-
proach using Adam optimizer with 0.001 learning rate. For
the COVID and Normal classes, only 2 samples were mis-
classified as Pneumonia for each. For the Pneumonia class,
all test samples were classified correctly .
In the previous work related to COVID classification, au-
thors proposed several approaches for COVID-19 predic-
tion and applied them on various datasets including stan-
dard ones and ones that were collected from public reposi-
tories. T able 6 provides a comparison in accuracy between
our proposed methodology and other state-of-the-art ap-
proaches applied on three-classes COVID datasets. Ioan-
nis et al. [21] proposed using VGG-19 with transfer learn-
ing and achieved 93.48% accuracy . Ali et al. [22] sug-
gested a new deep learning architecture called DRE-Net
for COVID-19 automatic detection and achieved 95.51%
accuracy . COVID-Net was proposed by W ang et al. [23]
and scored 92.4%. T ulin et al. [24] proposed DarkCovid-
Net and achieved a low accuracy of 87.02%. A work that
is similar to our proposed technique is Xu et al. [25] which
used ResNet architecture with location attention achieved
an accuracy of 86.7%. Our approach of mer ging multiple
CNN models proved to outperform these models. In addi-
tion, the standalone models used in our work were trained
from scratch without any need for transfer learning.

60 Informatica 48 (2024) 55–62 Z. A. Oraibi et al.
Figure 4: T raining accuracy vs validation accuracy legends are on the left. T raining loss vs validation loss legends are
on the right. These graphs are the results of training our hybrid CNN model of Xception, VGG-16, and VGG-19 using 5
epochs.
T able 5: Performance of our hybrid model using five accuracy metrics.
Model Accuracy Sensitivity Specificity F1 scor e Pr ecision
Xception 94.31% 94.0% 93.31% 93.31 % 94.31%
VGG-16 93.32% 94.0% 93.67% 93.67% 93.32%
VGG-19 96.91% % 96.32 97.29% 97.63% 96.88%
Xception + VGG-16 95.85% 95.63% 95.31% 95.32% 95.81%
Xception + VGG-16 + VGG-19 97.91% 97.33% 98.0% 97.96% 97.9%
5 Discussion
COVID prediction approach employed in this paper re-
lies on pre-trained state-of-the-art CNN architectures. W e
trained three well-known models from scratch: Xception,
VGG-16, and VGG-19 and combined them using the rule
of sum applied on the feature maps. Every model accepts
images of size224× 224× 3 and images were collected from
various resources and grouped into three classes: COVID,
Normal, and Pneumonia. Results of applying this approach
showed huge potential resulted from mer ging these mod-
els. The combination of the first two models, Xception +
VGG-16, improved the accuracy by almost 1% from the
standalone Xception model. Furthermore, the combination
of the three models, Xception + VGG-16 + VGG-19, fur -
ther improved the best accuracy by almost 1% achieving
97.91%. It is worthy to mention that training parameters
for all models were selected to achieve best training accu-
racy . W e used fewer number of epochs, 5, and showed that
the best hybrid model conver ges quickly with that number .
The learning rate was selected to be 0.001 and Adam opti-
mizer was used across all experiments. Data augmentation
was also employed to increase the number of training im-
ages for the three classes.
The techniques summarized in T able 1 relied on trans-
fer learning, ensemble learning, and deep features mer g-
ing. Some of these methods were applied on two classes
COVID-19 datasets while others were applied on three and
four classes. W e showed that using a combination of deep
learning models with data augmentation and batch size of
64 can produce superb results and can outperform the stan-
dalone models performance. The mer ge of Xception, VGG-
16, and VGG-16 improved the accuracy by 1% than the best
standalone model (VGG-19). The potential of combining
various CNN architectures is huge and we intend to apply
more models and mer ge them to further improve the accu-
racy .
6 Conclusion
The methodology proposed in this paper is simple yet it is
very ef fective in achieving high accuracy multi-class pre-
diction. State-of-the-art deep learning models were used
and trained from scratch using the same hyperparameters
in each experiment. These models include Xception, VGG-
16, and VGG-19. Then, we combined Xception and VGG-

Ef ficient COVID-19 Prediction by Mer ging V arious Deep Learning… Informatica 48 (2024) 55–62 61
T able 6: Evaluating the proposed hybrid model with the previous work applied on COVID datasets of three classes.
Methodology Accuracy (%)
VGG-19 [21] 93.49
DRE-Net [22] 95.51
Covid-Net [23] 92.4
DarkCovidNet [24] 87.02
ResNet + Location Attention [25] 86.7
Proposed (Xception + VGG-16) 95.85
Proposed (Xception + VGG-16 + VGG-19) 97.91
Figure 5: Confusion matrix generated for the hybrid model.
16 and showed that this binary hybrid model is better than
each model alone. After that, we combined the three
models and showed the final mer ged model is better that
the three models alone and better than the combination of
Xception and VGG-16. W e set the number of epochs to 5
and learning rate to 0.001 with a batch size of 64. Adam op-
timizer was employed and all experiments were performed
on Google Colaboratory with GPU. The dataset used in
the experiments was collected from multiple resources with
three classes: COVID, Normal, and Pneumonia. Five-fold
cross validation was used in the experiments with 80% of
samples in each class were used for training and the remain-
ing 20% were used for testing. Augmentation techniques
were applied to the subset of training images to enlar ge
it to perform well during training. Five accuracy metrics
were reported: Accuracy , precision, F1-score, sensitivity ,
and specificity . The best accuracy achieved from mer ging
the three models was 97.91%.
For the future work, since physicians need a robust
COVID classification system, we will focus on improving
the accuracy by applying ensemble methods with various
voting mechanisms. In addition, we will explore training
more state-of-the-art models to check their performances on
binary and multi-class COVID datasets. Furthermore, we
will apply our method on bigger COVID datasets with var -
ious hyperparameters settings. Our main goal will remain
to achieve the highest classification accuracy since this sys-
tem is related to human lives.
Ethical considerations and practical
applicability
The confidentiality of patient’ s COVID-19 data is very
crucial and must adhere to specific protocols in regard to
privacy and security of these information. In addition,
COVID-19 detection using machine learning models can
easily be biased because of the nature of datasets used dur -
ing training the models. Hence, careful evaluation and miti-
gation of model bias is very important. Moreover , explain-
ing to patients how the decision was made by a machine
learning model is very necessary since these models are
complex. In terms of practical applicability , COVID-19
prediction systems that depend on AI must adhere to reg-
ulatory scrutiny to make sure they are ef ficient, safe, and
amenable with healthcare system regulations.
Refer ences
[1] N. Zhu, D. Zhang, W . W ang, X. Li, B. Y ang, J. Song,
X. Zhao, B. Huang, W . Shi, R. Lu, et al. , “A novel
coronavirus from patients with pneumonia in china,
2019,” New England journal of medicine , vol. 382,
no. 8, pp. 727–733, 2020.
[2] F . W u, S. Zhao, B. Y u, Y . -M. Chen, W . W ang, Z.-G.
Song, Y . Hu, Z.-W . T ao, J.-H. T ian, Y .-Y . Pei, et al. ,
“A new coronavirus associated with human respira-
tory disease in china,” Natur e , vol. 579, no. 7798,
pp. 265–269, 2020.
[3] C. W ang, P . W . Horby , F . G. Hayden, and G. F . Gao,
“A novel coronavirus outbreak of global health con-

62 Informatica 48 (2024) 55–62 Z. A. Oraibi et al.
cern,” The lancet , vol. 395, no. 10223, pp. 470–473,
2020.
[4] T . T . Le, Z. Andreadakis, A. Kumar , R. G. Román,
S. T ollefsen, M. Saville, S. Mayhew , et al. , “The
covid-19 vaccine development landscape,” Nat Rev
Drug Discov , vol. 19, no. 5, pp. 305–306, 2020.
[5] P . Kaur , S. Harnal, R. T iwari, F . S. Alharithi, A. H.
Almulihi, I. D. Noya, and N. Goyal, “A hybrid convo-
lutional neural network model for diagnosis of covid-
19 using chest x-ray images,” International Journal of
Envir onmental Resear ch and Public Health , vol. 18,
no. 22, p. 12191, 2021.
[6] M. B. Hossain, S. H. S. Iqbal, M. M. Islam, M. N.
Akhtar , and I. H. Sarker , “T ransfer learning with fine-
tuned deep cnn resnet50 model for classifying covid-
19 from chest x-ray images,” Informatics in Medicine
Unlocked , vol. 30, p. 100916, 2022.
[7] M. M. A. Monshi, J. Poon, V . Chung, and F . M.
Monshi, “Covidxraynet: Optimizing data augmenta-
tion and cnn hyperparameters for improved covid-
19 detection from cxr ,” Computers in biology and
medicine , vol. 133, p. 104375, 2021.
[8] Y . Kaya and E. Gürsoy , “A mobilenet-based cnn
model with a novel fine-tuning mechanism for covid-
19 infection detection,” Soft Computing , vol. 27,
no. 9, pp. 5521–5535, 2023.
[9] T . Kausar , Y . Lu, A. Kausar , M. Ali, and A. Y ousaf,
“Sd-gan: A style distribution transfer generative ad-
versarial network for covid-19 detection through x-
ray images,” IEEE Access , vol. 1 1, pp. 24545–24560,
2023.
[10] Z. A. Oraibi and S. Albasri, “A robust end-to-end
cnn architecture for ef ficient covid-19 prediction form
x-ray images with imbalanced data,” Informatica ,
vol. 47, no. 7, 2023.
[1 1] E. E. W alsh, A. R. Falsey , I. A. Swinburne, and M. A.
Formica, “Reverse transcription polymerase chain re-
action (rt-pcr) for diagnosis of respiratory syncytial
virus infection in adults: Use of a single-tube “hang-
ing droplet” nested pcr ,” Journal of medical vir ology ,
vol. 63, no. 3, pp. 259–263, 2001.
[12] A. R. Khan, T . Saba, T . Sadad, et al. , “Computer
vision in co-clinical medical imaging for precision
medicine,” 2022.
[13] H. S. Basavegowda and G. Dagnew , “Deep learning
approach for microarray cancer data classification,”
CAAI T ransactions on Intelligence T echnology , vol. 5,
no. 1, pp. 22–33, 2020.
[14] A. Ahmad, D. Saraswat, V . Aggarwal, A. Etienne,
and B. Hancock, “Performance of deep learning mod-
els for classifying and detecting common weeds in
corn and soybean production systems,” Computers
and Electr onics in Agricultur e , vol. 184, p. 106081,
2021.
[15] W . Aldhahi and S. Sull, “Uncertain-cam: Uncertainty-
based ensemble machine voting for improved covid-
19 cxr classification and explainability ,” Diagnostics ,
vol. 13, no. 3, p. 441, 2023.
[16] Z. A. Oraibi and S. Albasri, “Predicting covid-19 from
chest x-ray images using a new deep learning architec-
ture,” in 2022 IEEE Applied Imagery Pattern Recog-
nition W orkshop (AIPR) , pp. 1–6, IEEE, 2022.
[17] S. Shastri, I. Kansal, S. Kumar , K. Singh, R. Popli,
and V . Mansotra, “Cheximagenet: a novel architec-
ture for accurate classification of covid-19 with chest
x-ray digital images using deep convolutional neu-
ral networks,” Health and T echnology , vol. 12, no. 1,
pp. 193–204, 2022.
[18] F . Chollet, “Xception: Deep learning with depthwise
separable convolutions,” in Pr oceedings of the IEEE
confer ence on computer vision and pattern r ecogni-
tion , pp. 1251–1258, 2017.
[19] K. Simonyan and A. Zisserman, “V ery deep convo-
lutional networks for lar ge-scale image recognition,”
arXiv pr eprint arXiv:1409.1556 , 2014.
[20] F . Chollet et al. , “Keras.”
https://github.com/fchollet/keras, 2015.
[21] I. D. Apostolopoulos and T . A. Mpesiana, “Covid-
19: automatic detection from x-ray images uti-
lizing transfer learning with convolutional neural
networks,” Physical and engineering sciences in
medicine , vol. 43, pp. 635–640, 2020.
[22] A. Al-Bawi, K. Al-Kaabi, M. Jeryo, and A. Al-
Fatlawi, “Ccblock: an ef fective use of deep learning
for automatic diagnosis of covid-19 using x-ray im-
ages,” Resear ch on Biomedical Engineering , pp. 1–
10, 2020.
[23] L. W ang, Z. Q. Lin, and A. W ong, “Covid-net: A tai-
lored deep convolutional neural network design for
detection of covid-19 cases from chest x-ray images,”
Scientific r eports , vol. 10, no. 1, p. 19549, 2020.
[24] S. Asif, Y . W enhui, H. Jin, Y . T ao, and S. Jinhai,
“Automatic detection of covid-19 using x-ray images
with deep convolutional neural networks and machine
learning,” 2020.
[25] X. Xu, X. Jiang, C. Ma, P . Du, X. Li, S. Lv , L. Y u,
Q. Ni, Y . Chen, J. Su, et al. , “A deep learning system
to screen novel coronavirus disease 2019 pneumonia,”
Engineering , vol. 6, no. 10, pp. 1 122–1 129, 2020.