Classification of COVID–19 and Pneumonia X–ray

Images Using a Transfer Learning Approach

Sai Kishore H R and M S Bhargavi Pavan Kumar C
Department of Computer Science and Engineering Department of Computer Science and Engineering
Bangalore Institute of Technology Indian Institute of Information Technology Dharwad
Bangalore, India India
2021 IEEE Region 10 Symposium (TENSYMP) | 978-1-6654-0026-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/TENSYMP52854.2021.9550878

[email protected], [email protected] ORCID: 0000-0002-4907-5412

Abstract—The Coronavirus disease is a respiratory infection RT-PCR test. In days 1 through 7 after onset of illness, 11%
caused by the Severe Acute Respiratory Syndrome CoronaVirus 2 of sputum, 27% of nasal, and 40% of throat samples were
(SARS-CoV-2), also known colloquially as COVID-19 virus. Non- deemed falsely negative.
Covid Viral Pneumonia is also a respiratory disease which affects
the lungs of the patients. It becomes difficult for radiologists and These studies raise concerns about the efficacy of the RT-
pulmonologists to differentiate between these two respiratory PCR tests, added to fact the amount of time that is lost
diseases. Chest X-ray images of the patients can be used to in getting the test results. In emergency cases, the amount
efficiently diagnose between these two respiratory diseases. Deep of time spent on the RT-PCR tests becomes too costly for
learning models can be efficiently used to detect subtle differences the patients. Faster methods must be developed to identify
between these X-ray images. This method can be used to gain
faster results in COVID-19 cases thereby reducing the time taken COVID-19 patients by the use of medical imaging [5].
for identifying a COVID-19 patient. X- ray images are cheaper The use of deep learning for classifying diseases has become
and faster than the currently existing methods. A Transfer quite popular in recent times. Diseases like Pulmonary Tuber-
Learning approach is adopted to classify chest X-rays into three culosis can be detected using Convolutional Neural Networks
categories such as COVID-19, Pneumonia and Normal. Popular
(CNNs) [6].
ImageNet Architectures: VGG-19, MobileNet and ResNet-50
are used to classify X-ray images of the patients. From the Lee et al. [7] examine the use of deep learning in identifying
experimental results, it is evident that the MobileNet is able to diseases using Computed Tomography (CT) scans and Chest
achieve a validation accuracy of 0.9777 in 40 epochs. Radiography. The use of X-rays instead of CT scans is a
Index Terms—COVID–19, Convolutional Neural Networks, conscious choice as X-rays are quite cheap and the results
Deep Learning, Transfer Learning.
can be obtained easily. The use of X-ray images for detection
of SARS-CoV-2 is proven to be a “better diagnostic tool than
I. I NTRODUCTION
CT scans in emergency situations” [8].
Severe Acute Respiratory Syndrome Coronavirus-2 popu- In this paper, we propose a method to use X–ray images
larly called as SARS-CoV-2 is a disease which has struck and classify these images using transfer learning approach.
the world and made it come to a standstill. COVID-19 as it This paper is organized as follows: preliminaries, related work
is generally called is quite similar to other acute respiratory and our contribution are discussed in section II, dataset and
diseases in its clinical symptoms [1]. The symptoms of this data augmentation process is discussed in section III, model
disease are Fever, Cough, Shortness of breath and fatigue. The architecture is discussed in section IV, experimental setup and
curative treatment for this disease is not yet discovered and the results are discussed in section V. Section VI concludes the
main strategy is to provide care so that symptoms alleviate paper.
gradually [1], [2].
The early detection of the COVID-19 infection becomes an II. P RELIMINARIES AND R ELATED W ORK
important task in tackling this pandemic. The use of nucleic
A. CNNs and Image Recognition
acid detection-based approach is a widely used technology for
viral detection. The Polymerase Chain Reaction (PCR) test is Due to the advent and advances in CNNs, image processing
considered the ‘gold-standard’ among the nucleic acid tests has taken a huge leap forward. Convolutional Networks can
for the diagnosis of certain viruses and is known for its high be effectively used to classify images.
sensitivity and specificity [3]. Convolutional Neural Networks are a special type of artifi-
However, there have been reports from all over the world cial neural networks which can recognize images by assigning
about incorrect diagnosis of SARS-CoV-2 using Reverse Tran- weights and biases to the pixels in the image [9], [10]. The
scription - Polymerase chain reaction (RT-PCR) tests [3]. Yang advantage of CNN’s are that they have very few parameters
et al. [4] described 213 patients hospitalized with COVID-19 and the filters are common. The concept of CNN’s are derived
of which 37 were critically ill. The team collected 205 throat from the connectivity of neurons within the brain where
swabs, 490 nasal swabs, and 142 sputum samples and used an individual neurons respond to a particular stimulus. Unlike

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.
other image recognition techniques where the weights have to Gozes et al. [18] proposed an AI model which could
be manually coded CNNs can efficiently learn these weights. classify COVID-19 cases from the non-COVID cases. The
The LeNet architecture was the first proposed architecture results obtained were 0.996 AUC (95% CI: 0.989-1.00). U-
which could recognize the handwritten digits using a 5 layer Net architecture is used for Lung Segmentation.
deep CNN [11]. This architecture provided excellent results Fei Shan et al. [19] developed a deep learning model
compared to the other methods. which auto-contours the infection areas in the Chest CT-scans.
Alongside auto-contouring, the system also predicts the shape,
B. CNNs in Medical Imaging
volume and percentage of Infection.
Convolutional Neural Networks has become a go-to tool Song Ying et al. [20] developed DeepPneumonia, a deep
for medical imaging. Pulmonary Tuberculosis is a respiratory learning based CT diagnosis system to help doctors in quick
disease which can be detected using CNN’s by using CT scans diagnosis of the COVID-19 infection. In this work, researchers
for classification [6]. Convolutional Neural Networks can be extracted the main regions of lungs from the CT images
used in areas of medical imaging which involves the use of and extracted top-k features from them to give image level
CT scans and Chest Radiography. Lee et al. [7] studies the predictions by designing a Details Relation Extraction neural
current state of the art in deep learning with respect to CT network (DRE-Net). Image level predictions were combined
scans and Chest Radiography. to provide patient level diagnosis. For training the network
researchers have used bacterial pneumonia infected, healthy
C. Transfer Learning
people and the COVID-19 infected people’s chest X-rays. This
Transfer Learning is a technique where the network is model achieved an accuracy of 0.86 and AUC of 0.95.
trained on one data, and the results are obtained for a Shuai Wang et al. [21] proposed a deep learning algorithm
completely new Data [12]. Research has proved that transfer which can predict COVID-19 using CT scan images. In this
learning can be used on models which have a small training model, initially features and areas of interest are identified and
dataset or some missing data points [12]–[14]. The idea was used transfer learning neural networks based on the Inception
to use an already trained network which would have learned network to classify CT images as typical viral pneumonia in-
certain key features on another dataset. These features prove fection or the COVID-19 infection. They achieved an accuracy
to be helpful and the validation accuracies attest to that. of 82.9%.
This approach is exploited in this research work where the Ghoshal and Tucker [22] investigated the uncertainty in
ImageNet dataset and an ImageNet architecture for training the deep learning based classification. They used Bayesian CNNs
model up to the last layer is used [15]. The learned weights by for estimating uncertainty in classification of COVID-19 using
these layers are frozen and the softmax layer which predicts X-ray images. They found that the uncertainty is ”strongly
1000 categories is replaced with a layer which predicts three correlated with the accuracy of the prediction.”
categories.
Khoshbakhtian, Ashraf and Khan [23] proposed a Convolu-
We experiment with three ImageNet Architectures:
tional Autoencoder framework which is trained by the normal
• MobileNet healthy adults and other non COVID-19 pneumonia patients’
• ResNet X-ray images such that whenever actual COVID-19 infected
• VGG–16/19 patient’s X-ray is detected then the model will identify it as
D. Related Work an anomaly. In two stages the model was tested on X-ray
images - first considering only the features of healthy adults
Xu et al. [16] developed a deep learning model which classi- and later considering healthy adults as well as non COVID-
fied COVID-19 from influenza-A viral pneumonia (IAVP) and 19 pneumonia type patients data. This has resulted in ROC -
healthy cases. They collected 618 CT samples in total with 219 AUC of 0.7652 and 0.6902 respectively.
samples of COVID-19 patients, 224 samples of IAVP patients Toraman, Alakus and Turkoglu [24] proposed Convolutional
and 175 samples of healthy cases. The 3D deep learning CapsNet for the detection of COVID-19 disease by using
model was used for the segmentation of infected regions chest X-ray images with capsule networks as backbone. In
and categorized into three groups using a location-attention this approach, detection of COVID-19 infection from X-ray
classification model. They obtained an overall accuracy rate images is carried out as a binary classification problem either
of 86.7% in terms of all the CT cases taken together. as COVID-19 infected or not infected as well as multi class
Brunese et al. [17] proposed a deep learning model to detect classification either as COVID-19 infected, pneumonia or not
COVID-19 infection from X-ray images. This model identifies infected problem. The binary classification method achieved
the presence of COVID-19 in three phases. Initially it separates an accuracy of 97.24% whereas multiclass classification ap-
the pneumonia infected chest X-ray images from the normal proach achieved an accuracy of 84.22%.
ones. In the second phase, it classifies the COVID-19 infected
patients’ X-rays from pneumonia infected ones. In the last
E. Our Contribution
phase, the model identifies the region of interest in X-rays
that is symptomatic of the COVID-19 disease. This model In the present research work, a transfer learning approach is
obtained an accuracy of 97%. proposed which provides an accuracy of 0.9777 over 40 epochs

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.
 Fig. 2: X–ray image of a pneumonia infected patient

Fig. 1: X–ray image of a COVID-19 infected patient

for multi-class classification. This is the highest accuracy

obtained for COVID-19/Pneumonia classification.

III. DATA
A. Dataset
The images in the dataset have been divided into 3 cate-
gories: COVID-19, Pneumonia and Normal X-ray images.
The data for COVID-19 X-ray images are obtained from Fig. 3: X–ray image of a normal patient
University of Montreal [25]. The images for the normal and
pneumonia infected X-ray images are sourced from a Kaggle
Dataset which in turn has been sourced from Guangzhou handwritten single digits identification and note the significant
Women and Children’s Medical Center [26]. The COVID-19 change in performance.
dataset consisted of Anteroposterior (AP), Posteroanterior (PA) For certain images in this research work, random horizontal
and AP supine views. Posteroanterior views are obtained when flips are performed, images are zoomed by a scale of 0.2 and
the detector is against the chest of the patient and the rays pass have a shear range of 0.2. Fig. 4 gives the sequence of steps
from the back to the front. Anteroposterior views are preferred which will be followed for the Data Augmentation process.
when the patient cannot stand and the detector is put behind IV. M ODEL A RCHITECTURE
the back of the patient. The rays pass from front of the chest
to the back [27]. The model architecture in Fig. 5 shows the flow of the
approach. The images are fed to the image transformer which
The images with AP views are obtained when the patient
does Data Augmentation as shown in Fig. 4 The augmented
is relatively immobile [27]. We have sourced 142 AP view
images are fed into three ImageNet architectures which are
images from the dataset which will be split into 112 and 30
modified according to the problem. The results of these models
images for the train and test data.
are then plotted in a graph.
Pneumonia dataset consists of normal X-ray images and
pneumonia infected X-ray images. The dataset has AP views A. MobileNet
only, and consists of 5,863 images with 2 categories. 142 MobileNet uses depth wise separable convolutions to reduce
images from each category are considered and the images are computations and build smaller networks for visual based
selected at random. Fig. 1 is a sample COVID-19 X-ray image
from the dataset. Fig. 2 and Fig. 3 are the Pneumonia infected
and Normal X-ray images respectively.

B. Data Augmentation
Data Augmentation is the process of augmenting the already
available training data by transforming certain aspects of
the image so that the Neural Network gets to see a lot of
images which reduces the chance of over fitting in training.
Wong et al. [28] investigate the use of Data Augmentation in Fig. 4: Data Augmentation Process

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.
 C. ResNet
The ResNet architecture was developed by He et al. in
2015 [31]. This paper showed that deep residual networks can
be trained to provide as good accuracy as shallow networks
of small depth. The ResNet architecture comes with several
variants such as ResNet-50, ResNet-101, ResNet-152 etc.
ResNet-50 is used for our model as we are looking at training
these sets of images with not so deep networks.
V. E XPERIMENTATION AND R ESULTS
A. Experimental Setup
We use Keras library as the deep learning Library and
Google Colab GPU for implementation. The specifications of
the computer used for this experiment are as follows–
• Intel i5 processor
• 20 GB RAM
• Nvidia 2GB Graphics Card
The images in the dataset have been divided into 3 cat-
egories. COVID-19 , Pneumonia and Normal X-ray images.
The total number of images considered are 426, with each
category having an equal number of 142 images. The images
are equally divided among all the 3 classes to avoid class
imbalance. 79% of the images (336) in random are selected
as training set and 21% of the images (90) as test set. We vary
the epoch numbers for the experiment from 30 - 80 epochs.
This is done to find the best epoch number within a given
range.
B. Metrics
Fig. 5: Model Architecture We use Accuracy, Precision, recall and F1 scores in terms
of True Positives (TP), True Negatives (TN), False Positives
(FP) and False Negatives (FN) for evaluating the model
performance.
tasks [29]. MobileNet uses Batch Normalization and ReLU
1) Accuracy: Accuracy informally, is defined as the ratio
non-linearity activation function. The Depth of this archi-
of the Number of correct predictions to the total number of
tecture is 28 layers considering depth wise and point wise
predictions.
convolution as separate layers. For this research work, the last
Accuracy = Number of Correct Predictions / Total number
layer of Softmax is replaced to classify three categories.
of Predictions
TP + TN
B. VGG Accuracy =
TP + TN + FP + FN
VGG architecture was introduced in 2014 by Simonyan and 2) Precision:: calculated as the ratio of number of true
Zisserman [30]. An unique characteristic of this architecture positives to the total number of true positives and false
is that they used 3 × 3 filters in all the layers [30]. This positives. Higher the value of precision, lower is the false
architecture did not use Batch Normalization and used ReLU positives.
as an activation function. We use this architecture to analyze TP
Precision =
the close spatial properties of the images, i.e., whether the 3 TP + FP
× 3 filter captures the most important features of the image. 3) Recall:: calculated as the ratio of number of true positive
There are two versions of VGG: VGG-16 containing 16 to the total number of true positive and false negative. Higher
layers and VGG-19 containing 19 Layers. The performance the value of recall, lower is the false negatives.
of these two versions are generally comparable [30]. TP
Recall =
We use VGG-19 in our model as it’s depth is comparable TP + FN
to the other networks being used. The results don’t change a 4) F1 – score::
lot when we replace VGG-19 with VGG-16. The core idea for precision × recall
using VGG architecture is to have a model with a fixed kernel F1–score = 2 ×
precision + recall
size.

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.
 TABLE I: Validation Accuracy for Different Architectures

VGG-19 MobileNet RestNet-50 VGG-19 MobileNet RestNet-50

No of epochs
Best Accuracy Best Accuracy Best Accuracy Mean Accuracy Mean Accuracy Mean Accuracy
30 0.9555 0.9555 0.8555 0.9061 0.9144 0.7304
40 0.9444 0.9777 0.8555 0.9061 0.9241 0.7321
50 0.9555 0.9666 0.8777 0.8993 0.9086 0.7366
60 0.9555 0.9667 0.8777 0.8993 0.9098 0.7408
70 0.9555 0.9555 0.8666 0.8993 0.9219 0.7622
80 0.9555 0.9667 0.8666 0.9021 0.9291 0.8121

(a) VGG-19 (b) MobileNet (c) ResNet-50

Fig. 6: Training and Validation – Accuracy and Loss Graphs

5) Loss Function: The Loss function we use is the Cross architecture takes 13 seconds on an average. The time taken
Entropy Loss. Cross entropy loss is the standard loss function by the ResNet architecture is 13 seconds on an average. The
used for multiclass classification. The cross entropy loss time taken by these architecture are comparable because of
accurately captures the loss made by the model and this loss the small number of images.
function can be used for back-propagation. Cross entropy loss The accuracies of VGG-19 and MobileNet are comparable.
is formally defined as, The MobileNet architecture is 44 times smaller than VGG-
n
X 19 in terms of the parameters but still requires less time in
LCE = − ti log(pi ) processing input. This results in much faster calculations and
i=1 very less processing time. The use of MobileNet over VGG-19
is naturally preferred.
where ti is the truth label and pi is the softmax probability of
The obtained results were further validated by additional
the ith class.
metrics: precision, recall and F1–scores. From Table II, it is
C. Reproducible result evident that MobileNet has a better average precision, recall
and F1-Scores compared to VGG-19 and ResNet–50.
The source code for this work is freely available online for
the readers/researchers to use it for research1 . TABLE II: Precision, Recall and F1-score of MobileNet,
RestNet–50 and VGG–19
D. Result Analysis
Precision Recall F1-score
From Table I, it is evident that the VGG-19 architecture is Mobilenet 0.9728 1.0 0.9862
immune to the change in the number of epochs. The ResNet- ResNet–50 0.7486 0.7820 0.7674
50 architecture is quite unstable to the change in the number VGG–19 0.8952 0.901 0.899
of epochs. There is a 0.818 difference in the mean accuracy.
Fig. 6 depicts the accuracy and loss graph for the different VI. C ONCLUSION
architectures with respect to the epoch numbers.
Need for the quick diagnosis of the diseases has increased in
The MobileNet architecture performs the best amongst all
recent times. In order to cater that need of quick diagnosis, in
the other models giving the highest accuracy of 0.9777 within
this paper popular imagenet architectures such as MobileNet,
40 epochs. This architecture is quite stable to the change in the
VGG-19 and ResNet-50 are compared. Transfer learning ap-
number of epochs. The average time taken in the MobileNet
proach is adopted for the quick classification of COVID-19
architecture is 10 seconds per epoch whereas the VGG-19
and Pneumonia X-ray images. The accuracy achieved by the
1 https://github.com/saik2121/Transfer-Learning-for-Covid-Pneumonia- model indicates efficacy of this approach. Even though this
Prediction approach is faster than the standard RT-PCR method, it can

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.
be further improved by introducing human intervention and [17] L. Brunese, F. Mercaldo, A. Reginelli, and A. Santone, “Explainable
making the predictions more reliable. deep learning for pulmonary disease and coronavirus covid-19 detection
from x-rays,” Computer Methods and Programs in Biomedicine, vol.
R EFERENCES 196, p. 105608, 2020.
[18] O. Gozes, M. Frid-Adar, H. Greenspan, P. D. Browning, H. Zhang,
[1] D. Tang, P. Comish, and R. Kang, “The hallmarks of covid-19 disease,” W. Ji, A. Bernheim, and E. Siegel, “Rapid ai development cycle for the
PLoS pathogens, vol. 16, no. 5, p. e1008536, 2020. coronavirus (covid-19) pandemic: Initial results for automated detection
[2] I. Ali and O. M. Alharbi, “Covid-19: Disease, management, treatment, & patient monitoring using deep learning ct image analysis,” arXiv
and social impact,” Science of the total Environment, vol. 728, p. 138861, preprint arXiv:2003.05037, 2020.
2020. [19] F. Shan, Y. Gao, J. Wang, W. Shi, N. Shi, M. Han, Z. Xue, D. Shen,
[3] A. Tahamtan and A. Ardebili, “Real-time rt-pcr in covid-19 detection: and Y. Shi, “Abnormal lung quantification in chest ct images of covid-19
issues affecting the results,” Expert review of molecular diagnostics, patients with deep learning and its application to severity prediction,”
vol. 20, no. 5, pp. 453–454, 2020. Medical physics, 2020.
[4] Y. Yang, M. Yang, C. Shen, F. Wang, J. Yuan, J. Li, M. Zhang, Z. Wang, [20] Y. Song, S. Zheng, L. Li, X. Zhang, X. Zhang, Z. Huang, J. Chen,
L. Xing, J. Wei et al., “Laboratory diagnosis and monitoring the viral H. Zhao, Y. Jie, R. Wang et al., “Deep learning enables accurate
shedding of 2019-ncov infections,” MedRxiv, 2020. diagnosis of novel coronavirus (covid-19) with ct images,” medRxiv,
[5] E. A. Akl, I. Blažić, S. Yaacoub, G. Frija, R. Chou, J. A. Appiah, 2020.
M. Fatehi, N. Flor, E. Hitti, H. Jafri et al., “Use of chest imaging in [21] S. Wang, B. Kang, J. Ma, X. Zeng, M. Xiao, J. Guo, M. Cai, J. Yang,
the diagnosis and management of covid-19: a who rapid advice guide,” Y. Li, X. Meng et al., “A deep learning algorithm using ct images to
Radiology, vol. 298, no. 2, pp. E63–E69, 2021. screen for corona virus disease (covid-19),” MedRxiv, 2020.
[6] P. Lakhani and B. Sundaram, “Deep learning at chest radiography: au- [22] B. Ghoshal and A. Tucker, “Estimating uncertainty and interpretability
tomated classification of pulmonary tuberculosis by using convolutional in deep learning for coronavirus (covid-19) detection,” arXiv preprint
neural networks,” Radiology, vol. 284, no. 2, pp. 574–582, 2017. arXiv:2003.10769, 2020.
[7] S. M. Lee, J. B. Seo, J. Yun, Y.-H. Cho, J. Vogel-Claussen, M. L. [23] F. Khoshbakhtian, A. B. Ashraf, and S. S. Khan, “Covidomaly: A deep
Schiebler, W. B. Gefter, E. J. Van Beek, J. M. Goo, K. S. Lee convolutional autoencoder approach for detecting early cases of covid-
et al., “Deep learning applications in chest radiography and computed 19,” arXiv preprint arXiv:2010.02814, 2020.
tomography,” Journal of thoracic imaging, vol. 34, no. 2, pp. 75–85, [24] S. Toraman, T. B. Alakus, and I. Turkoglu, “Convolutional capsnet: A
2019. novel artificial neural network approach to detect covid-19 disease from
[8] D. Ippolito, C. Maino, A. Pecorelli, P. Allegranza, C. Cangiotti, x-ray images using capsule networks,” Chaos, Solitons & Fractals, vol.
C. Capodaglio, I. Mariani, T. Giandola, D. Gandola, I. Bianco et al., 140, p. 110122, 2020.
“Chest x-ray features of sars-cov-2 in the emergency department: a [25] J. P. Cohen, P. Morrison, L. Dao, K. Roth, T. Q. Duong, and M. Ghas-
multicenter experience from northern italian hospitals,” Respiratory semi, “Covid-19 image data collection: Prospective predictions are the
medicine, vol. 170, p. 106036, 2020. future,” arXiv preprint arXiv:2006.11988, 2020.
[9] K. O’Shea and R. Nash, “An introduction to convolutional neural [26] D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L.
networks,” arXiv preprint arXiv:1511.08458, 2015. Baxter, A. McKeown, G. Yang, X. Wu, F. Yan et al., “Identifying
[10] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, medical diagnoses and treatable diseases by image-based deep learning,”
X. Wang, G. Wang, J. Cai et al., “Recent advances in convolutional Cell, vol. 172, no. 5, pp. 1122–1131, 2018.
neural networks,” Pattern Recognition, vol. 77, pp. 354–377, 2018. [27] D. Kermany, K. Zhang, M. Goldbaum et al., “Labeled optical coherence
[11] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning tomography (oct) and chest x-ray images for classification,” Mendeley
applied to document recognition,” Proceedings of the IEEE, vol. 86, data, vol. 2, no. 2, 2018.
no. 11, pp. 2278–2324, 1998. [28] S. C. Wong, A. Gatt, V. Stamatescu, and M. D. McDonnell, “Under-
[12] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans- standing data augmentation for classification: when to warp?” in 2016
actions on knowledge and data engineering, vol. 22, no. 10, pp. 1345– international conference on digital image computing: techniques and
1359, 2009. applications (DICTA). IEEE, 2016, pp. 1–6.
[13] C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu, “A survey on [29] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang,
deep transfer learning,” in International conference on artificial neural T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convo-
networks. Springer, 2018, pp. 270–279. lutional neural networks for mobile vision applications,” arXiv preprint
[14] B. Rezaeianjouybari and Y. Shang, “Deep learning for prognostics and arXiv:1704.04861, 2017.
health management: State of the art, challenges, and opportunities,” [30] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
Measurement, vol. 163, p. 107929, 2020. large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[15] S. Bianco, R. Cadene, L. Celona, and P. Napoletano, “Benchmark [31] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
analysis of representative deep neural network architectures,” IEEE recognition,” in Proceedings of the IEEE conference on computer vision
Access, vol. 6, pp. 64 270–64 277, 2018. and pattern recognition, 2016, pp. 770–778.
[16] X. Xu, X. Jiang, C. Ma, P. Du, X. Li, S. Lv, L. Yu, Q. Ni, Y. Chen,
J. Su et al., “A deep learning system to screen novel coronavirus disease
2019 pneumonia,” Engineering, 2020.

Authorized licensed use limited to: VIT University. Downloaded on October 12,2021 at 10:29:28 UTC from IEEE Xplore. Restrictions apply.