Proceedings of the Sixth International Conference on Computing Methodologies and Communication (ICCMC 2022)

IEEE Xplore Part Number: CFP22K25-ART; ISBN: 978-1-6654-1028-1

Screening Covid-19 Infection from Chest CT Images using

Deep Learning Models based on Transfer Learning
Malliga Subramanian
Department of Computer Science Adhithiya G J Gowthamkrishnan S
2022 6th International Conference on Computing Methodologies and Communication (ICCMC) | 978-1-6654-1028-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICCMC53470.2022.9754041

and Engineering, Department of Computer Science Department of Computer Science

Kongu Engineering College, Erode, and Engineering, and Engineering,
Tamil Nadu, India Kongu Engineering College, Erode, Kongu Engineering College, Erode,
Tamil Nadu, India Tamil Nadu, India

Deepti R
Department of Computer Science
and Engineering,
Kongu Engineering College, Erode,
Tamil Nadu, India
for RT-PCR results. Chest radiography (X-rays) and chest
computed tomography (CT) scans, according to a Chinese
Abstract— As the global epidemic of Covid19
progresses, accurate diagnosis of Covid19 patients becomes study, can help in diagnosis. Both, notably Covid-19,
important. The biggest problem in diagnosing test-positive could suggest lung disease abnormalities. Chest CT is
people is the lack or lack of test kits due to the rapid spread of more effective than chest X-ray in detecting early Covid-
Covid19 in the community. As an alternative rapid diagnostic 19 illness. A CT chest scan is a sophisticated imaging
method, an automated detection system is needed to prevent procedure that uses X-rays to create 3D chest images. For
Covid 19 from spreading to humans. This article proposes to identifying lung-related disorders, chest CT scans are a
use a convolutional neural network (CNN) to detect patients more effective imaging tool. A chest CT scan, which is
infected with coronavirus using computer tomography (CT)
used to diagnose pneumonia, may be done fast and easily.
images. In addition, the transfer learning of the deep CNN
CT has a sensitivity of 98 percent for Covid-19 infection,
model VGG16 is investigated to detect infections on CT
scans.The pretrained VGG16 classifier is used as a classifier, compared to 71 percent for RT-PCR testing, according to
feature extractor, and fine tuner in three different sets of tests. new research. As a result, CT has emerged as one of the
Image augmentation is used to boost the model's generalization most important tools for early detection and diagnosis of
capacity, while Bayesian optimization is used to pick optimum Covid-19 [1, 2].
values for hyperparameters. In order to fine-tune the models
and reduce training time, transfer learning is being researched. Due to the rapid expansion of artificial intelligence,
S urprisingly, all of the proposed models scored greater than computer vision algorithms, which were originally used
93% accuracy, which is on par with or better than previous for recognizing images have been used in medical
deep learning models. The results show that optimization imaging, particularly CT images [3]. For collecting visual
improved generalization in all models and highlight the efficacy
information, including spatial correlations, artificial
of the proposed strategies.
intelligence has been developed utilizing medical
Keywords—Covid-19, CT Images, Convolution Neural imaging-based deep learning algorithms. The CNN, in
Networks, VGG16, Transfer learning, Bayesian Optimization particular, performed well when it came to feature
extraction. In a number of computer vision tasks, the
I. INT RODUCT ION CNN performed brilliantly [4, 5]. [6] Several CNN-based
Corona viruses are a type of virus. There are a number of algorithms for Covid-19 identification have been
different types, some of which are pathogenic. The proposed, and these strategies use a small number of
Covid-19 respiratory sickness pandemic was caused by training samples from medical image datasets. CNNs
SARS-CoV-2, a coronavirus detected in 2019. Symptoms have been constructed for big datasets using a variety of
appear two to fourteen days after being exposed to the methodologies, including parameter optimization, data
virus. The Covid-19 pandemic is now sweeping the augmentation, and transfer learning, to achieve
globe. People who test positive for Covid-19 should unprecedented efficiency. The main goal of this study is
discover out if they are infected as soon as possible so to develop a reliable method for identifying patients with
that they can avoid contact with others, keep away from Covid 19 from CT images of CNN.
them, and notify close friends and family members. The
presence of Covid-19 is being detected through a In this study, a set of models are developed using a pre-
laboratory test of nasal and throat samples for a definitive trained VGG16 architecture and automatically classify
diagnosis. Medical practitioners can diagnose people with and detect Covid19 infections using CT images
Covid-19 symptoms using chest imaging while waiting downloaded from the Internet. Three strategies are

978-1-6654-1028-1/22/$31.00 ©2022 IEEE 1358

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.
 employed: full retraining, feature extraction, and fine- as sensitivity, specificity, and AUC, as well as testing on
tuning to optimize the pre-trained model for diversion. a variety of data sets.
Freeze the convolution base of the pre-trained model for
For comparison and modeling, chest CT scans are
feature extraction and train some convolution layers
collected from 88 patients diagnosed with Covid19 in
while freezing other convolution layers for fine tuning. hospitals in two regions of China, 100 patients infected
Another feature of this study is to use Bayesian with pneumonia, and 86 healthy people [10]. Based on
optimization to find the optimal hyperparameter the data, a deep CT diagnostic method was developed to
configuration for training the VGG16 architecture. The identify Covid19 patients. The authors argue that their
contributions listed below have made a significant impact model can tell the difference between Covid19 and
on this project.: patients with bacterial pneumonia. [11] proposed a
collection of five convolutional neuron network-based
models (ResNet50, ResNet101, ResNet152, InceptionV3,
(i) investigating the performance of VGG16 through
and InceptionResNetV2) to identify patients with
transfer learning and fine-tuning. coronavirus lung infection. Using 5-validation, three
different binary classifications are created, each with four
(ii) a comparison of the accuracy, precision, recall and classes (Covid19, normal, viral pneumonia, and bacterial
F1-score of proposed models. pneumonia). The pre-trained ResNet 50 model has the
highest classification accuracy depending on performance
(iii) Applying Bayesian optimization to hyper-parameter results. [12] employs a CNN to determine if individuals
optimization. are infected or not. In addition, multi-objective
differential evolution is employed to fine-tune CNN's
initial parameters. Extensive studies are conducted out on
The study's key uniqueness is that it employs transfer
chest CT scans utilizing the proposed and competing for
learning with two fine-tuning procedures for pre-trained machine learning techniques. A thorough examination
models. VGG16's performance is evaluated by adapting demonstrates that the suggested model can correctly
previously trained models developed on the ImageNet categorize chest CT images.
dataset and evaluating them on the CT image dataset via
transfer learning. To our knowledge, no other study has Deep learning architectures are increasingly being
applied in the detection of Covid-19 infection using CT
integrated transfer learning with two fine-tuning
imaging, as indicated by the papers listed above.
processes, as well as Bayesian Optimization, for However, various shortcomings in the application of deep
classifying CT images in the literature. learning architectures must be addressed, such as shorter
training times, proper hyper-parameter tuning, fewer
II. LIT ERAT URE SURVEY parameters, and so on. Transfer learning is utilized in this
paper to try to cut down on training time and determine
Mishra et al. [7] investigated various deep-CNN- the optimal set of hyper-parameters for the suggested pre-
based approaches to detect Covid19 on chest CT images. trained models.
It also describes a decision fusion technique that merges
predictions from many models to make the final III. PROPOSED W ORK
prediction. Experimental results show that the proposed
decision fusion-based approach can exceed all
performance criteria by 86% or more. The authors of [8] The primary goal of this study is to create and
have developed a deep learning-based approach for compare several models for classifying CT images using the
autonomous segmentation and measurement of infected VGG16 architecture. Evaluating the performance of the
areas and the entire lung using chest CT scans. The VGG 16 architecture and identifying the variants that
neural network "VBNet" is used in deep learning-based reliably detect Covid-19 infection have been proposed. This
segmentation techniques to segment Covid19 infected section covers all of the materials as well as the
lesions on CT scans. A new deep learning-based methodologies employed in this study.
segmentation method was trained on 249 Covid19 CT A. Dataset Description
scans and validated on 300 CT scans. Wang et al. [9]
From training to testing classification models, an
looked at Covid19 cases with confirmed pathogens and
previously diagnosed cases of normal viral pneumonia. acceptable dataset is necessary at all phases of object
To develop the model, they optimized the first transfer recognition. A total of 1426 CT pictures were collected from
learning model and then used internal and external various sources, including Kaggle and other websites, and
validation. The authors performed external and internal divided into two categories: Covid-19 and Non-Covid-19. A
verifications to ensure accuracy, and the results showed sample of each group of CT scans is shown in Figure 1. The
an overall accuracy of 89.5 percent with a specificity of blue circle in Figure 1(a) indicates the Covid-19 infection
0.88 and a sensitivity of 0.87. The external test dataset
showed an overall accuracy of 79.3%, a specificity of (a). There are 672 images with and 754 images without
0.83, and a sensitivity of 0.67. The authors of [6] used the Covid-19 infection.
DenseNet-121 CNN model for classification and
identification of patients with Covid-19, and the
experimental analysis looked at a variety of indices such

1359

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.
 the use of multiple 3x3 kernel-size filters rather than a single
large kernel filter. Figure 2 shows the layers of VGG16.

(a) Covid-19 (b) Non-Covid-19 Figure 2. VGG16 Architecture

Figure 1. Covid-19 and Non-Covid-19 CT Images (Source: https://www.geeksforgeeks.org/vgg-16-cnn-

model/)
B. Image Augmentation
A large amount of data is required to train a neural
network. The greater the amount of data available to the
network, the more features it will be able to learn. Image D. Transfer Learning
augmentation techniques are used because only a restricted It's a method for creating new artificial intelligence
number of images are available. Image augmentation is a models by fine-tuning previously trained neural networks
useful strategy when there is no sufficient data to train deep [21]. It is usually demonstrated in computer vision by using
learning models. Rotation, shifting, flipping, blurring, pre-trained models. Using the ImageNet dataset, Canziani et
scaling, cropping, padding, translation, and affine al. [22] evaluated the performance of pre-trained models on
transformation are some of the picture enhancement computer vision tasks. CNN consists of two parts: a
techniques used. In-place / on-the-fly data augmentation is convolutional base and a classifier. The main purpose of a
also used. The main advantage of using this augmentation is convolution base is to extract features from an image and the
that it allows for real-time augmentation. That is, while classifier is to categorize the image using the extracted
training a model, it generates augmented images on the fly features. The lower layers of a CNN's convolutional base
and ensures that the model receives new variations of the learn general characteristics, whereas the classifier part and
images at each epoch. To increase feature extraction and a few top layers of the convolutional base, learn specialized
consistency, the images in the dataset are pre-processed features. Transfer learning, even if based on ImageNet, is
before training. For VGG16 models, all of the images in this useful because the sample size is small, which is common in
study have been normalized and scaled to 224*224 pixels. medical imaging [23]. In this study, the pre-trained model's
After augmentation, the total number of images in the early layers are frozen, and only the remaining higher layers
dataset has increased to 4152. To test the various CNN are retrained. The upper layers would then need to be
architectures, the dataset was divided into 70% training, customized to the new images. In this scenario, the pre-
15% validation, and 15% test sets. trained model is employed as a classifier. Because the new
data set has no relation to ImageNet, the upper layers must
C. Methods be retrained and customized to account for the variations in
CNN has been programmed to derive visual patterns the new data.
directly from input images [13, 14], and it is at the cutting
In order to demonstrate the performance of fine-tuning, the
edge technology for image classification tasks. Lecun et al. VGG16 is trained from scratch (full retraining) and also use
[15] used a gradient-based algorithm with CNN to solve the it as a feature extractor. While developing from scratch, it is
problem of handwritten digit categorization. Several intended to retrain all the layers of the model. As a feature
baseline CNN architectures for image recognition extractor, the convolution base is kept in its original form
applications [17,18,19,20] have been successfully developed with the weights learned from ImageNet. In VGG16, the
classifier generates 1000 different output labels, but the
and applied to complex visual imagery tasks. The ability of
number of neurons in the output layer can be decided based
VGG16 models to detect infection in CT scans was
on the number of classes in the dataset. So, the convolution
examined in this work. The rationale for selecting VGG16 is base can be imported and its own classifier can be added.
presented below. The output from the convolution base is fed into a classifier.

The VGG16 object-recognition model has up to 16. On a

variety of tasks and datasets, VGG16, which is built as a E. Bayesian Optimization
deep CNN. VGG16 is still one of the most popular image- The hyperparameters include network structure,
recognition models on the market today. As a result, this such as the number of hidden units, dropout, activation
model is suggested to be employed. VGG16, a 16-layer function, and weight initialization, as well as how network
learning parameters like learning rate, momentum, burst
neural network presented in 2014 by Simonyan and
size, and epoch are used. The purpose of hyperparameter
Ziserman of Oxford University's Visual Geometry Group tuning is to determine the ideal hyperparameter setting to
Lab [16] and was found to be even deeper than AlexNet but reduce the loss function and improve accuracy. A variety of
with significantly lower computational complexity due to methods, including manual, grid, random searches, and
Bayesian optimization, can be used to fine-tune the hyper-

1360

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.
parameters. Bayesian optimization, unlike Grid and Random
search, makes use of prior iterations of the algorithm. In the
grid and random searches, each hyper-parameter guess is Table 1. Hyper-parameters with tuned values
independent.
Hyper- VGG16
With Bayesian techniques, on the other hand, with each parameters Full Feature Fine-
selection and testing of potential hyper-parameters, the Retrainin Extractor tuner
perfect solution is approached. To put it another way, g
Bayesian optimization assists in deciding which hyper- Optimizer Adam Adam Adam
parameter combination is ideal for evaluating a model [24, Learning rate 0.001 0.0001 0.0001
Activation Relu tanh tanh
25]. Due to the amount of data and computer density, deep
function
learning models can take a long time to train. Bayesian Number of 128 512 128
optimization can be quite effective in these situations. To neurons in
improve the hyper-parameters of the classifier layers, customized
Bayesian optimization is employed in conjunction with pre- layers
trained models.

C. Performance Metrics
IV. EXPERIMENT S The next step after developing the models is to
To examine the performance of models developed evaluate their effectiveness using metrics for the test
using VGG16, three sets of experiments are run: datasets. The various models built in this study were
constructing from scratch (full retraining), Feature Extractor, evaluated using performance metrics like accuracy,
and Fine-tuner. Graphical Processing Units (GPUs) are used precision, recall, and F1 Score. These metrics are calculated
to run the proposed models because they consume a lot of using the True Positive (TP), True Negative (TN), False
power and require high-performance hardware to work Positive (FP), and False Negative (FN) indices. The total
properly. The relevant Keras model architectures are number of successfully categorized photos in each class is
imported and used ImageNet weights to instantiate them. denoted by TP. The number of photos misclassified in
classes other than the right one is denoted by FP. The
A. Tuning of Hyper-parameters
number of misclassified photos in that class is represented
Hyper-parameters are significant in deep learning by FN. TN stands for the total number of correctly identified
algorithms because they specify training information and photos in all classes excluding the correct one.
have a direct impact on model output. Applying Bayesian
optimization to find the best hyperparameter values that
provide high accuracy. Optimizers (Adam, RMSProp, SGD, V. RESULT S AND DISCUSSION
Nadam), learning rate (1e-3, 1e-4, 1e-5, 1e-6), activation This section presents the performance of the proposed
function (Relu, Elu, LeakyRelu, and Tanh), and the number models and compares their performance with each other.
of neurons (64,128, 56, 512,1024) are among the
hyperparameters tuned in this work. A. Results of Proposed Models
The performance of the model generated in this study
B. Bayesian Optimization
was evaluated. Experiments were performed with the
First, a model is developed using pre-trained adjusted hyperparameters listed in Table 1 that gave the best
VGG16 in which the weights of all the layers are randomly training results. The overall validation and testing accuracy
initialized at first and then retraining happens. Then, VGG16 for each of the models is shown in Table 2.
has been used as a feature extractor in the second set of
experiments, and the extracted features are then used to train Table 2. Performance of the proposed models
the newly added classifier. The weights learned from the
ImageNet dataset have been used in the convolution base. In Experiment VGG16
the third set of experiments, a few top layers of the scenario Validation Testing Accuracy
convolution base are retained. The initial layers of pre- Accuracy (%)
trained models can be avoided from retaining because they (%)
learn lower-level features in the input images, resulting in a Full Retraining 94.24 95.18
significant reduction in training effort. So, simply training Feature Extractor 92.99 93.58
the classifier and the convolutional base layers is sufficient. Fine-tuner 95.18 96.79

The best values for the hyper-parameters are determined

As can be seen in Table 2, fully retraining the models and
through the application of Bayesian optimization. The
using the models as feature extractors result in lower
Bayesian optimization process has been repeated 50 times.
validation and testing accuracy than retraining a few top
The number of epochs is fixed to 100 for each iteration of
layers. The fine-tuning strategy keeps the weights of
Bayesian optimization. The accuracy and loss are recorded
previously trained models from the initial layers and retrains
after each iteration. The hyper-parameter values that resulted
them on top layers. The weights of the initial layers are
in the maximum accuracy in each experiment are listed in
usually retained because the features obtained from these
Table 1. Because additional rounds did not yield substantial
layers are generic and applicable to a wide range of tasks.
results, the hyper-parameters found after 50 iterations were
The latter layers, on the other hand, provide more specific
deemed the best in this study.
features that, because they have been tailored to the dataset,
can benefit from the fine-tuning approach. In Strategy 1, the

1361

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.
entire model is fully retrained and predefined weights have For each of the proposed models, accuracy, precision, recall,
not been used. In Strategy 2, however, the pre-trained model and F1-score are calculated and displayed in Table 3 using
is simply used as feature extractors without being fine- the values of TP, FP, TN, and FN.
tuned, and the top layers of the models are not retrained
specifically for the dataset. As a result, the learned features Table 3. Performance of the pre-trained VGG16 models
are unique to ImageNet, and the accuracy is lower than in (in percentage)
strategy 3.
Models Accuracy Precision Recall F1-S core
The performance of the proposed model is evaluated Full
with respect to each class. TP, FP, TN, and FN indices are Retraining 95.18 96.07 95.53 95.8
used to calculate the values for various performance Feature
indicators such as Precision, Accuracy, Recall, and F1 Extractor 93.58 94.69 93.59 94.13
Score. To find the values of these exponents, the confusion Fine-tuner 96.79 97.46 96.92 97.19
matrix obtained while training and testing the model is used.
As known, confusion matrices are a visualization tool used
in supervised learning and can be used to check how well Table 3 compares the results based on performance metrics
the prediction results fit the real data. on figs. Figure 3 like accuracy, precision, etc. It is found that, like overall
shows the confusion matrix obtained by training all models accuracy, the fine-tuning strategy outperforms the feature
in the two selected scenarios. The diagonal elements of the extractor strategy and fully retrained model. It is believed
errata table indicate the correct classification. The predicted that higher accuracy is achieved because the top layers of
class is shown on the x-axis and the actual class is shown on the convolution base are fine-tuned with the data set. Also,
the y-axis. For example, in Figure 3(a), it can be understood the performance of all versions of VGG16 in all metrics is
that the model classified 14 non-Covid images as Covid. found to be greater than 93%. Furthermore, among all pre-
trained models developed in this work, fine-tuning VGG
performed the best in terms of metrics. The pre-trained
networks can be used to recognize classes they weren't
originally trained on by using fine-tuning. Furthermore, this
strategy has the potential to produce higher accuracy than
feature extraction-based transfer learning. From these
results, it is understood that transfer Learning, even if it is
ImageNet based, is beneficial when the sample size is small,
as it is in most cases in medical imaging and when the
convergence rates of the different models are compared, the
pre-trained models converge faster.
(a) VGG16 ( Full Retraining)

VI . CONCLUSION

With the recent spread of COVID-19, CT has become the

simplest and fastest diagnostic method. Problems arise,
however, from the lack of quantifiable diagnostic criteria
and experienced radiologists. Therefore, automatic
classification of CT images using a computer system is
important for medical image analysis. In this study, pre-
trained VGG16 was used to predict Covid19 infection.
Transfer learning was chosen for this study because of its
(b) VGG16 ( Feature Extractor) advantages that it does not require very large training data
sets and requires low computational power by retraining
only the weights of the top few layers. Bayesian
optimization is also used to select the best hyperparameter
values used during training. According to the findings of this
work, using pre-trained models powered by Bayesian
hyperparameter optimization and transfer learning for CT
image categorization for Covid19 infection diagnosis are
viable options.

REFERENCES
(b) VGG16 ( Fine Tuner)
[1] J. Lei, J. Li, X. Li, and X. Qi, (2020) "CT imaging of the 2019
Figure 3. Visualization results novel coronavirus (2019-nCoV) pneumonia", Radiology, vol.
295, pp. 18-18.

1362

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.
[2] H. Shi, X. Han, and C. Zheng, (2020) "Evolution of CT vision", in Proceedings of the IEEE conference on computer
manifestations in a patient recovered from 2019 novel vision and pattern recognition, pp. 2818-2826.
coronavirus (2019-nCoV) pneumonia in Wuhan, China",
Radiology, vol. 295, pp. 20-20. [20] F. Chollet, (2017) "Xception: Deep learning with depthwise
separable convolutions", in Proceedings of the IEEE conference
[3] Z. Yang, T . Luo, D. Wang, Z. Hu, J. Gao, and L. Wang, (2018) on computer vision and pattern recognition, 2017, pp. 1251-
"Learning to navigate for fine-grained classification", in 1258.
Proceedings of the European Conference on Computer Vision
(ECCV), 2018, pp. 420-435. [21] J. Chen, J. Chen, D. Zhang, Y. Sun, and Y. A. Nanehkaran,
(2020) "Using deep transfer learning for image-based plant
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, (2012) "Imagenet disease identification”, Computers and Electronics in
classification with deep convolutional neural networks", Agriculture, vol. 173, p. 105393.
Advances in neural information processing systems, vol. 25, pp.
1097-1105. [22] A. Canziani, A. Paszke, and E. Culurciello, (2016) "An analysis
of deep neural network models for practical applications," arXiv
[5] J. Qin, W. Pan, X. Xiang, Y. T an, and G. Hou, (2020) "A preprint arXiv:1605.07678.
biological image classification method based on improved
CNN", Ecological Informatics, vol. 58, p. 101093. [23] M. Raghu, C. Zhang, J. Kleinberg, and S. Bengio, (2019)
"T ransfusion: Understanding transfer learning for medical
[6] N. Hasan, Y. Bao, A. Shawon, and Y. Huang, (2021) "DenseNet imaging," arXiv preprint arXiv:1902.07208.
convolutional neural networks application for predicting
COVID-19 using CT image", SN Computer Science, vol. 2, pp. [24] M. I. Sameen, B. Pradhan, and S. Lee, (2020) "Application of
1-11. convolutional neural networks featuring Bayesian optimization
for landslide susceptibility assessment", Catena, vol. 186, p.
[7] A. K. Mishra, S. K. Das, P. Roy, and S. Bandyopadhyay, (202) 104249, 2020.
"Identifying COVID19 from chest CT images: a deep
convolutional neural networks based approach", Journal of [25] M. Pelikan, D. E. Goldberg, and S. T sutsui, (2003) "Hierarchical
Healthcare Engineering, vol. 2020. Bayesian optimization algorithm: toward a new generation of
evolutionary algorithms," in SICE 2003 Annual Conference
[8] F. Shan, Y. Gao, J. Wang, W. Shi, N. Shi, M. Han, et al., (2021) (IEEE Cat. No. 03TH8734), pp. 2738-2743.
"Abnormal lung quantification in chest CT images of COVID‐
19 patients with deep learning and its application to severity
prediction", Medical physics, vol. 48, pp. 1633-1645.
[9] S. Wang, B. Kang, J. Ma, X. Zeng, M. Xiao, J. Guo, et al.,
(2021) "A deep learning algorithm using CT images to screen
for Corona Virus Disease (COVID-19)", European radiology,
pp.1-9.
[10] Y. Song, S. Zheng, L. Li, X. Zhang, X. Zhang, Z. Huang, et al.,
(2021) "Deep learning enables accurate diagnosis of novel
coronavirus (COVID-19) with CT images", IEEE/ACM
Transactions on Computational Biology and Bioinformatics.
[11] A. Narin, C. Kaya, and Z. Pamuk, (2021) "Automatic detection
of coronavirus disease (covid-19) using x-ray images and deep
convolutional neural networks", Pattern Analysis and
Applications, pp. 1-14.
[12] D. Singh, V. Kumar, and M. Kaur, (202) "Classification of
COVID-19 patients from chest CT images using multi-objective
differential evolution–based convolutional neural networks",
European Journal of Clinical Microbiology & Infectious
Diseases, vol. 39, pp. 1379-1389.
[13] Y. LeCun, Y. Bengio, and G. Hinton, (2015) "Deep learning,”,
Nature, vol. 521, pp. 436-444.
[14] Y. LeCun, F. J. Huang, and L. Bottou, (2004) "Learning
methods for generic object recognition with invariance to pose
and lighting", in Proceedings of the 2004 IEEE Computer
Society Conference on Computer Vision and Pattern
Recognition, CVPR 2004, pp. II-104.
[15] Y. LeCun, L. D. Jackel, L. Bottou, C. Cortes, J. S. Denker, H.
Drucker, et al., (1995) "Learning algorithms for classification: A
comparison on handwritten digit recognition", Neural networks:
the statistical mechanics perspective, vol. 261, p. 2.
[16] K. Simonyan and A. Zisserman, (2014) "Very deep
convolutional networks for large-scale image recognition",
arXiv preprint arXiv:1409.1556.
[17] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, and D.
Anguelov, (2015) "Going deeper with convolution”, In:
Proceedings of the IEEE conference on computer vision and
pattern recognition; 2015, pp. 1-9.
[18] S. Ioffe and C. Szegedy, (2015) "Batch normalization:
Accelerating deep network training by reducing internal
covariate shift”, International conference on machine learning,
pp. 448-456.
[19] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna,
(2016) "Rethinking the inception architecture for computer

1363

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 18,2022 at 04:43:40 UTC from IEEE Xplore. Restrictions apply.