Classiﬁcation of Chest X-ray Images to Diagnose Covid-19 using Deep

Learning Techniques

Isabel Helo

ıse Santos Silva

, Ramoni Reus Barros Negreiros

, Andr

e Luiz Firmino Alves

1,2

Dalton C

ezane Gomes Valadares

2,3,4 a

and Angelo Perkusich

2,3

Federal Institute of Para

ıba (IFPB), Picu

ı, PB, Brazil

Federal University of Campina Grande, Computer Science, Campina Grande, PB, Brazil

VIRTUS RDI Center, Campina Grande, PB, Brazil

Federal Institute of Pernambuco, Mechanical Engineering Department, Caruaru, PE, Brazil

Keywords:

Artiﬁcial Neural Networks, ANNs, Machine Learning, Image-based Diagnosis, Radiographic Images.

Abstract:

The new coronavirus pandemic has brought disruption to the world. One of the signiﬁcant dilemmas to be

solved by countries, especially in underdeveloped countries like Brazil, is the lack of mass testing for the

population. An alternative to these tests is detecting the disease through the analysis of radiographic images.

To process different types of images automatically, we employed deep learning algorithms to achieve success

in recognizing different diagnostics. This work aims to train a deep learning model capable of automatically

recognizing the Covid-19 diagnosis through radiographic images. Comparing images of coronavirus, healthy

lung, and bacterial and viral pneumonia, we obtained a result with 94% accuracy.

1 INTRODUCTION

The new coronavirus pandemic has plagued the world

since November 2019, with its appearance in the

province of Wuhan (China), causing from asymp-

tomatic infections to severe respiratory issues, leading

to death. Initially, the belief was that there were just

a few isolated pneumonia cases, which further aggra-

vated the local population’s situation. According to

the WHO

, the Wuhan Municipal Health Commission

has already reported an outbreak of pneumonia cases

on December 31, 2019. Eventually, the scientists dis-

covered that it was a new member of the Coronavirus

family.

The rapid spread of the disease has been one of

the main problems for the health area. According to

the ”Centers for Disease Control and Prevention”

the contamination occurs mainly by contact with in-

fected people, through droplets of saliva expelled by

them, which land in the mouths and noses of those

nearby. Besides, there may be contamination through

the sharing of objects that have had contact with these

https://orcid.org/0000-0003-1709-0404

https://www.who.int/news-room/detail/27-04-2020-

who-timeline—covid-19

https://www.cdc.gov/coronavirus/2019-ncov/faq.html

droplets.

To have an idea of the pandemic dimension, at the

end of June (2021), what appeared to be a simple dis-

ease claimed the lives of almost 4 million human be-

ings across the globe

. Brazil occupies the 3rd place

in the world ranking, with more than 500,000 deaths

At this moment, the country is behind only India and

the United States, which generates an emotional im-

passe, of people who have lost their loved ones, and

political, on the part of the political authorities trying

to solve the problem.

With the rapid contamination of Covid-19, there is

a lack of infrastructure and medical resources world-

wide. Furthermore, the diagnosis of COVID-19 is

typically associated with pneumonia symptoms that

can be revealed by genetic and imaging tests (Li et al.,

2020; Silva et al., 2021). Countries suffer from the

lack of hospital beds, respirators, exams, and, mainly,

from testing the population, becoming difﬁcult to

know the real proportions of the damage to public and

private health.

Due to this situation, some research areas be-

come good agents to solve or mitigate these prob-

lems. In Computer Science, the Artiﬁcial Neural Net-

https://www.worldometers.info/coronavirus/

https://www.worldometers.info/coronavirus/#countries

Silva, I., Negreiros, R., Alves, A., Valadares, D. and Perkusich, A.

Classiﬁcation of Chest X-ray Images to Diagnose Covid-19 using Deep Learning Techniques.

DOI: 10.5220/0011339700003286

In Proceedings of the 19th International Conference on Wireless Networks and Mobile Systems (WINSYS 2022), pages 93-100

ISBN: 978-989-758-592-0; ISSN: 2184-948X

works (ANNs) — characterized by the ability to learn

through examples and to generalize the information

learned (Sp

orl et al., 2011) — are an alternative to

help in different areas science (Os

oio and Bittencourt,

2000; Shi et al., 2020). Indeed, the computational so-

lutions, especially those from the Computational Vi-

sion’s of state-of-the-art, indicate changes in every-

day life (Cui et al., 2020), such as the unmanned car,

which recognizes routes and objects.

In the ﬁeld of medicine, Computer Vision has

made several signiﬁcant contributions, mainly with

the use of advanced ANN techniques (Chen, 1995).

In particular, these contributions can be applied to as-

sist in the Covid-19 diagnosis through the processing

of radiographic images, in which the presence of dis-

eases is visualized. Thus, these techniques allow the

analysis and screening of cases with coronavirus, en-

abling an alternative to mitigate the problem of spe-

cialized mass testing in a country’s population. Thus,

the rapid detection of the disease can contribute to

controlling its propagation (Abbas et al., 2020).

Given this problem, we decided to apply the con-

cepts of Machine Learning (ML) and Deep Learning

(DL) in a computational model, allowing the machine

to automatically diagnostic the Covid-19, through the

analysis of radiographic chest images, assisting in

hospital screening. We trained the model to identify

and catalog the images among the following classes:

Covid-19; healthy lung; viral pneumonia; and bacte-

rial pneumonia.

The main contributions of this work are listed be-

low:

• we gather groups of images from different

databases;

• we catalog and pre-process these images, mak-

ing them ready for use in machine learning tech-

niques;

• we train a model using a deep learning technique

to diagnose Covid-19, bacterial pneumonia, and

viral pneumonia in x-ray images;

• we make available the code and datasets regarding

the model training and test processes;

• we grant access to our model, which highlights

the Covid-19 class with good accuracy.

The structure of this article is organized as fol-

lows: in Section 2, we explain the general concepts

related to this work; in Section 3, we mention some

related works; in Section 4, we describe the steps of

the proposed methodology; in Section 5, we present

and discuss the model’s results; lastly, in Section 6,

the present our ﬁnal remarks.

2 BACKGROUND

In Machine Learning, learning can take place in two

ways: supervised learning and unsupervised learning

(dos Santos et al., 2017). In supervised learning, a

labelled data set with input patterns and their corre-

sponding output patterns is applied to train the model.

The model must learn from these examples, providing

the best responses as output, according to the acquired

knowledge from the original dataset. In unsupervised

learning, there is no external agent to accompany this

process, i.e., no type of labelled dataset is given to the

learning algorithm. In this case, the model learns to

identify patterns on its own and tries to classify them

automatically.

Additionally, among these concepts, there is semi-

supervised learning, which has been highly explored

in both machine learning and data mining ﬁelds. This

learning type can use available unlabelled data to im-

prove the supervised learning tasks when the datasets

are expensive or scarce/insufﬁcient (Zhu and Gold-

berg, 2009).

In Information Retrieval problems, this learning

can be achieved using Artiﬁcial Neural Networks

(ANNs) to predict unknown examples (Chen, 1995).

In this sense, ANNs are learning models that seek to

simulate human brain behavior, inspired by the Cen-

tral Nervous System (dos Santos et al., 2017). Thus,

the ANN operation is normally performed by inter-

connected “neurons” that process the input data and

return a series of outputs, identifying patterns in im-

ages, for example. Within this context, Convolutional

Neural Networks (CNN) are a class of ANNs that was

inspired by the human visual system for image pro-

cessing and recognition (Da Silva and Costa, 2019).

CNN is a class of neural networks known as deep

learning (DL). The main difference between CNN and

simple ANNs is in the depth of the learning methods

to ﬁnd f (x), the function that determines the model

training. The DL method learns through a series

of f (x) functions that are composed of each other,

functioning as layers (Kopiler et al., 2019): f (x) =

f n(...( f 2( f 1(x))...)). The set of layers f 1, f 2, ..., f n

receive an input value, x, from the input layers, which

”crosses” layer by layer during the learning process -

usually called hidden layers - and returns a value as

output.

Besides the CNNs, there are other ANN types,

such as the Feedback ANN and the Feed Forward

ANN

. The main characteristic of the Feed Forward

ANN is recognizing and evaluating input patterns,

while the Feedback ANN is commonly used due to

https://www.elprocus.com/artiﬁcial-neural-networks-

ann-and-their-types/

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

its capacity of solving optimization problems.

Figure 1: Simple and Deep Learning Neural Networks.

Fig. 1 shows that the approach in ANN is simpler

than in DL. A CNN comprises three main layers: (i)

convolutional layers, which apply the data augmen-

tation concept, an image regularization method that

avoids overﬁtting, in addition to increasing the data

batch (this occurs by reusing the images, but changing

their translation, rotation, ﬂattening, etc., to increase

the training dataset); (ii) pooling layers, which reduce

the spatial dimensions of the image; and (iii) fully

connected layers, which convert the 2D feature maps,

the output of a ﬁlter applied to the previous layer, into

a 1D feature vector, generating the ﬁnal classiﬁcation

(Voulodimos et al., 2018). Besides the CNNs, there

are other ANN types, such as the Feedback ANN and

the Feed Forward ANN

. The main characteristic of

the Feed Forward ANN is recognizing and evaluating

input patterns, while the Feedback ANN is commonly

used due to its capacity of solving optimization prob-

lems.

Another relevant concept is Transfer Learning,

a technique that adapts a pre-trained model, which

performs a general task, to perform a speciﬁc task

(Da Silva and Costa, 2019). This pre-trained model,

in our case, was the resnet34

. The resnet34 is a CNN

belonging to the Residual Networks family, which are

CNNs adapted to maintain good rates of train loss and

validation loss even with many layers of processing.

Before working on our dataset, the resnet34 is pre-

trained with the gigantic ImageNet

database, which

puts the Transfer Learning concept into practice.

When working with machine learning, we may en-

counter some problems linked to the dataset. If the

dataset is built with little or “dirty” data, the trained

model may present unsatisfactory results. For in-

stance, the model may suffer overﬁtting due to not

performing a pre-processing of the dataset before

training. Besides, the model may suffer from in-

sufﬁcient adjustments, such as low/high number of

epochs, loss function, batch size and optimization al-

https://www.elprocus.com/artiﬁcial-neural-networks-

ann-and-their-types/

https://www.kaggle.com/pytorch/resnet34

http://image-net.org/

gorithm. In this case, the model fails to generalize

correctly, that is, to learn the necessary classiﬁcation

standards.

2.1 Covid-19 Detection

Figure 2: Patient with a clinical picture of Covid-19.

Bai et al. (Bai et al., 2020) observed that pneumonia

caused by the Covid-19 had presented peripheral dis-

tribution with ground-glass opacities (GGO) and vas-

cular thickening. The medical diagnostics of Covid-

19 is achieved through the analyses of the lung opac-

ities. Normally, this opacities distribution is bilateral,

peripheral, and in the lower zone of the lung (Ro-

drigues et al., 2020; Wong et al., 2020). Despite this,

due to the similarities in the images, the diagnostic of

diseases by x-ray is easily confused, becoming impor-

tant to diagnose the viral tests, to validate the individ-

ual situation.

Fig. 2 shows the abnormalities (opacities) located

in the lung of a patient with Covid-19. Other exams

carried out by the Italian Society of Medical Radio-

graphy (SIRM)

conﬁrmed the Covid-19 diagnostic

in the patient. The diagnostic prediction using the

images happens similarly to the specialists’ predic-

tions. The machine ﬁnds opacity patterns and at-

tributes them to speciﬁc diseases, once it has already

learned these patterns previously.

3 RELATED WORKS

The Deep Learning and Machine Learning - sub-

areas of Artiﬁcial Intelligence (AI) - have been very

useful in the ﬁeld of Computer Vision (CV) (Shi

et al., 2020). The various AI methods allow mak-

ing automated predictions of different image cate-

gories, making the visual classiﬁcation process faster

and simpliﬁed, without the need of a human specialist

(de Oliveira et al., 2019). Deep learning has brought

signiﬁcant developments to image processing tasks

such as object detection, image classiﬁcation, and im-

age segmentation (Ohri and Kumar, 2021).

https://www.sirm.org/

Classiﬁcation of Chest X-ray Images to Diagnose Covid-19 using Deep Learning Techniques

CV has been applied in image recognition for sev-

eral purposes, such as ﬁsh detection (Cui et al., 2020);

problem solving in the electrical sector (Kopiler et al.,

2019); diagnosis of pneumonia and Alzheimer in the

health area (de Oliveira et al., 2019; Duarte et al.,

2020); and oil recognition on beaches (Negreiros

et al., 2020) in the environmental area.

With applications in the health area, espe-

cially with the pandemic caused by the coronavirus

(COVID-19), AI techniques have made social contri-

butions highlighted by several academic works with

several purposes to alleviate the global crisis: prevent

the spread of the COVID-19 with automatic detection

of face masks (Singh et al., 2021), to monitor people

wearing masks in public places, and Covid-19 diag-

nosis through images x-ray (Shi et al., 2020; Abbas

et al., 2020; Li et al., 2020).

Shi et al. (Shi et al., 2020) analyzes several works

on Covid-19 diagnosis, being possible to observe the

preference of using the U-Net CNN specialized in

biomedical images. Prioritizing the need for labeled

images of lungs, mainly in studies for automatic de-

tection of Covid-19, Zheng et al. [22] proposed an un-

supervised learning model to generate image labels.

Abbas et al. (Abbas et al., 2020) used a pre-

trained CNN architecture called DeTraC, which high-

lights its ability to focus on irregularities (overlap-

ping images) present in the data for detecting the dis-

ease, with a class decomposition mechanism. Li et

al. (Li et al., 2020) developed a DL model based

on neural network, named COVNET, to classify im-

ages in three classes: COVID 19, CAP (Community-

Acquired Pneumonia) and non-pneumonia.

Similar to the previous work, Hu et al. (Hu et al.,

2020) proposed a semi-supervised model capable of

improving the necessary time for manually labeling

images based on three classes: Non-Pneumonia (NP),

Community-acquired pneumonia (CAP), and Covid-

19. Their work differs in a binary classiﬁcation type,

which compares the introduced classes in pairs, ob-

taining more detailed and precise results. However,

the total quantity of used images (450 equally divided

among the three classes) is lower than the quantity

used by us in this work.

Cohen et al.(Cohen et al., 2020) measured the

Covid-19 gravity to the patients using a linear regres-

sion model. As parameters, they used the lung exten-

sion and its opacity degree. As advantages, the work

enables us to conﬁrm the treatment efﬁcacy and to in-

crease the training epochs. The major disadvantage

is the number of samples (153), which limited their

evaluations on a large scale.

Rajaraman et al. (Rajaraman et al., 2020) anal-

ysed the performance of a set of CNN models,

which contains one customized and eight pre-trained

CNNs (VGG-16, VGG-19, Inception-V3, Xception,

InceptionResNet-V2, MobileNet-V2, DenseNet-201

e NasNet-mobile). They used images from different

datasets, and the results demonstrated that the preci-

sion was greater than 90% for all models. A weighted

mean for the models with the “best performance”

showed that the Covid-19 detection by pulmonary x-

rays presented 99.01% precision.

Phankokkruad (Phankokkruad, 2020) carried out

a similar analysis, considering three pre-trained CNN

models to detect among Covid-19, varied types of

pneumonia, and healthy lungs. The major disadvan-

tage was the limited number of images for Covid-19,

which was only 323 x-ray images. Despite that, the

author used techniques to increase the data, obtaining

excellent results for all CNNs: the Xception model

achieved a precision of 97.19%, while the VGG-16

model achieved 95,42%, and the Inception-Resnet-

V2 model achieved 93,87%.

Hu et al. (Hu et al., 2021) aimed to bring opti-

mization and agility to Covid-19’s detection and train-

ing processes. The authors used deep CNN and Ex-

treme Learning Machines (ELMs), as well as an op-

timization algorithm, to perform real-time detection.

They obtained accuracies of 98.25% and 99.11% for

the two evaluated datasets. The authors highlight the

excellent training time of the network (0.9474 mil-

liseconds). However, the image classiﬁer was trained

to recognize three types of x-ray images: covid 19,

pneumonia, and healthy lung.

Sousa et al. (de Sousa et al., 2020) researched

other works that apply deep learning techniques to

identify the Covid-19 disease. They presented the

main techniques used and the obtained results. Be-

sides, they subdivided the relevant works into two

types: those that used conventional radiographs and

those that used computed tomographies. For the ﬁrst

type, Altan and Karasu (Altan and Karasu, 2020) pre-

sented the best results, achieving 99,7% of accuracy,

99,4% of sensibility, and 99,5% of F1-Score. For the

second type, Ko et al. (Ko et al., 2020) presented the

best results, with 99,9% of accuracy, 99,6% of sensi-

bility, and 100% of speciﬁcity.

In this context, the prior differential of our work

is using a big dataset and the model’s ability to dis-

tinguish four distinct types of categories: Covid-19,

healthy lung, bacterial pneumonia, and viral pneumo-

nia. The inclusion of pneumonia images is because

they are easily confused with some cases of Covid-

19. Therefore, the model acquires a speciﬁc skill in

differentiating Covid-19 and pneumonia images.

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

4 METHODOLOGY

We used the Google Colaboratory environment

(Google Colab), a free cloud tool provided by Google

and built based on Jupyter Notebook to encourage re-

search in the ﬁeld of AI. In addition to providing GPU

(Graphics Processing Unit) acceleration, responsible

for rendering graphics in real-time, the Google Colab

is free of charge, allowing better optimization of time

and processing.

Figure 3: Applied Methodology.

Figure 3 presents the adopted methodology for

this work, which is divided into the following three

stages: image collection and pre-processing, model

training, and model evaluation. These stages are de-

scribed below.

4.1 Image Collection and

Pre-processing

In the ﬁrst stage of this work, we gathered 7683 x-ray

images obtained from the following distinct sources:

Cohen’s database

, Augmented Covid-19 X-Ray Im-

ages

, Chest X-Ray Images

, and Italian Society of

Medical Radiography (SIRM)

. The Table 1 shows

the total number of images collected for the classes

and the dataset sources.

There are three image classes in the Cohen

dataset: Covid-19, healthy lung, and pneumonia. As

we intended to train the model to classify four types of

lung radiographs, we separated the pneumonia class

images into two others: bacterial pneumonia and viral

pneumonia. In addition, we removed some repeated

images, from the dataset, and the computed tomog-

raphy images of the lungs, as the main objective was

working with chest x-ray images.

We divided the images as follows: the testing

dataset contains 10% of the total number of images,

https://github.com/ieee8023/covid-chestxray-dataset

https://data.mendeley.com/datasets/2fxz4px6d8/4

https://www.kaggle.com/paultimothymooney/chest-

xray-pneumonia

https://www.sirm.org/en/category/articles/covid-19-

database/

corresponding to 768 images; the validation dataset

contains 20% of the 6915 remaining images, corre-

sponding to 1383 images; and the training dataset

contains the rest of the images, which corresponds to

5532 images.

Figure 4 presents some pulmonary tomography

images, extracted from the collected dataset, with

their respective diagnoses.

Figure 4: Examples of tomography images.

4.2 Model Training

To build the x-ray image classiﬁer, we used the Fas-

tai

library to apply supervised deep learning tech-

niques. We imported the module vision, which con-

tains all the functions necessary to train the model.

This module used the following two sub-modules: vi-

sion.data, which contains the ImageDataBunch ob-

ject, responsible for creating the training, validation,

and test datasets, and vision.learner that allows us to

perform the training or use a pre-trained model as a

base.

To train the model, we used the following four

classes of lung radiography images: healthy, with

Covid-19, with bacterial pneumonia, and with viral

pneumonia. To train and validate the model, respec-

tively, we used 6915 e 1383 images from the dataset.

To test the model, we used 768 images.

To train the classiﬁer, we applied transfer learning

using the cnn learner method, from the fastai.vision

module. The cnn learner assists in obtaining a pre-

trained model from a given architecture. We selected

the architecture resnet34 once it is a neural network

very applied by other researchers (Lau et al., 2020;

Lei et al., 2018; Canziani et al., 2016).

Figure 5: Neural Network Architecture.

https://docs.fast.ai/

Classiﬁcation of Chest X-ray Images to Diagnose Covid-19 using Deep Learning Techniques

Table 1: Number of images by dataset.

Datasets Covid Healthy Bacterial P. Viral P.

Cohen’s database 920 - - -

Augmented Covid-19 866 - - -

Chest X-Ray Images - 1583 2761 1493

SIRM 60 - - -

Total of Images 7683

Fig. 5 represents the neural network architecture

used. The image to be categorized goes through an

initial 7x7 convolution layer with a 64 resource map,

and is interpreted through another 34 layers of the

model. The later layers are 3x3 convolutions, with

a map dimension of ﬁxed features at, respectively, 64,

128, 256, 512. During the training of this model, im-

ages of size 224 were used for standardization

4.3 Model Evaluation

Through the confusion matrix, we can make a visual

analysis of our classiﬁer’s mistakes and successes. On

the main diagonal, the correct predictions are located,

while the other positions refer to the classiﬁer errors.

With this tool, we can also scrutinize these results in

the four different terms: True Positives (TP), True

Negatives (TN), False Positives (FP), and False Neg-

atives (FN). These terms are used in the calculations

of the main evaluation metrics by a micro-average.

To evaluate the model, we used accuracy (Acc),

precision, recall, and F1-score (or f-measure). The ac-

curacy is applied to calculate the percentage of correct

answers in the general model. The precision is asso-

ciated with the effective number of correct classiﬁca-

tions. On the other hand, the recall indicates how of-

ten the classiﬁer ﬁnds examples from the same class.

Finally, the F1-score represents the harmonic mean of

precision and recall.

Acc =

V P +V N

V P +V N + FP + FN

(1)

Recall =

V P

V P + FN

(2)

Precision =

V P

V P + FP

(3)

score

2 × Precision × Recall

Precision + Recall

(4)

We also considered the construction of a model

that does not present underﬁtting, when the model

does not adapt well to the training set, and overﬁt-

ting, when a model presents learning bias by mem-

orizing the data set patterns and does not learn in a

generalized way. Experimentally, we evaluated that

15 epochs in the model’s training process were sufﬁ-

cient to present good results during training, valida-

tion, and tests.

5 RESULTS

Table 2 represents the confusion matrix of our model,

which was used to evaluate the performance from

the separate data used to test the X-ray image classi-

ﬁer. The “Covid” class corresponds to lungs affected

by the Covid-19, and the “Healthy” class indicates

healthy and clean lungs. The “BP” and “VP” classes

indicate, respectively, bacterial and viral pneumonia.

We observed that the results of the viral pneu-

monia classiﬁcation produced the highest error num-

bers, what probably happened due to this class have

a smaller number of images (1493) when compared

with other classes.

Table 2: Confusion Matrix.

Predicted

Covid Healthy BP VP FN

Covid 153 1 1 3 5

Actual Class Healthy 12 248 0 16 28

BP 1 0 184 0 1

VP 8 46 2 93 56

FP 21 47 3 19 90

Table 3: Metrics.

Covid Healthy BP VP

Recall 0.97 0.90 0.99 0.62

Precision 0.88 0.84 0.98 0.83

F1 score 0.92 0.87 0.99 0.71

Accuracy 0.94

The precision of bacterial pneumonia or “BP”

classiﬁcation is the most striking, as shown in Table 3,

as it is low even with excellent recall. To understand

why this occurs, we go back to Table 2. We can notice

that the FP of the bacterial pneumonia class totalizes

3, indicating that the model rarely classiﬁes the im-

age as “BP” when, in fact, it should be another class,

as happened many times with the categories “VP” and

”Healthy”. To mitigate this problem, we can add more

images of viral pneumonia and healthy lung, aiming

at training the model more frequently with these im-

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

ages and, consequently, getting better values of recall,

precision, and F1 score.

The model presented the best performance for the

“BP” class, exceeding expectations, as we can see that

it obtained 97% for all three metrics (Tab. 3). Such

results indicate that the model can diagnose bacterial

pneumonia with conﬁdence, helping professionals in

hospitals.

Regarding the “Covid” category, we can evidence

the model’s ability to generalize new images, indicat-

ing an effective adaptation of the classiﬁer in iden-

tifying images of lungs affected by the Coronavirus.

Looking at the error number of the classiﬁer, we

can notice that the “Covid” class presents three times

fewer errors than the “VP” and “Healthy” classes. In

this sense, we were able to build a model that iden-

tiﬁes the COVID-19 in X-ray images, which was the

main objective of our work.

Looking at Table 3, we can see that the best

results belong to the “BP” class, followed by the

“Covid”, “Healthy” and “VP” classes. Therefore, the

chances of the model correctly classifying a lung with

Covid-19 as “Covid” are higher than that of classify-

ing viral pneumonia as “VP”, for example. We can

also notice that the class with lower Recall, Precision,

and F1 values is the VP, with 62%, 83%, and 71%,

respectively. Despite the lower numbers, the general

performance of the model was not affected.

In general, we can check the test results using the

accuracy, which achieved 94% (Table 3). Using the

same metric, but for validation, we obtained 91% ac-

curacy. So we decided to create a test directory, noting

that the validation results are very good. In this sense,

we conclude that when it is required to generalize the

standards for images never seen before by the classi-

ﬁer, it presents more imperfections, even though its

results are satisfactory in the set of tests.

6 CONCLUSION

Given the current pandemic and the need to obtain

an urgent and more accurate diagnosis of COVID-19

contamination, this work proposed to build a com-

putational model capable of automatically identifying

signs of the coronavirus infection in chest x-ray im-

ages. For that, we used a Convolutional Neural Net-

work based on the resnet34 architecture, which allows

excellent results in image identiﬁcation tasks. In this

sense, using learning transfer techniques, we built and

trained a model to classify x-ray images, identifying

healthy lungs, Covid-19, bacterial pneumonia, and vi-

ral pneumonia.

The model presented excellent classiﬁcation re-

sults, with an accuracy of 87%. Among the four

classes identiﬁed, the “Covid-19” class achieved the

best results: 97% recall, 94% accuracy, and 95% F1

score. We conclude that our machine learning-based

application can automatically identify the Covid-19

disease, using pulmonary radiographs, with a 95%

F1-score. Thus, our solution contributes to assist in

the screening of Covid-19 cases, as an alternative to

the lack of mass testing for the population.

Despite the satisfactory results, medical analysis

is always valuable and important. The diagnoses us-

ing the model work as a medical monitoring instru-

ment, not eliminating the conﬁrmation by a special-

ized physician. Only the radiography is not enough to

deﬁne if a patient has Covid-19. Thus, to avoid over-

ﬁtting and bias when implementing such an experi-

ment in a hospital, an expanded and optimized testing

is recommended (Wynants et al., 2020).

As future work, we want to improve the model

training with more images, mainly of other diseases,

to increase the database and optimize the accuracy,

precision, and other evaluation metrics for viral and

bacterial pneumonia. Furthermore, we intend to adapt

the model to detect and classify more types of lung

diseases, such as lung cancers and other pneumonia

types.

Sourcecode. We provide the project source

code publicly at the following address:

https://github.com/double-blind/review/

ACKNOWLEDGEMENTS

We would like to thank VIRTUS Research, Devel-

opment & Innovation Center, Federal University of

Campina Grande (Para

ıba/Brazil), for the technical

support provided.

REFERENCES

Abbas, A., Abdelsamea, M. M., and Gaber, M. M. (2020).

Classiﬁcation of covid-19 in chest x-ray images us-

ing detrac deep convolutional neural network. arXiv

preprint arXiv:2003.13815.

Altan, A. and Karasu, S. (2020). Recognition of covid-19

disease from x-ray images by hybrid model consist-

ing of 2d curvelet transform, chaotic salp swarm algo-

rithm and deep learning technique. Chaos, Solitons &

Fractals, 140:110071.

Bai, H. X., Hsieh, B., Xiong, Z., Halsey, K., Choi, J. W.,

Tran, T. M. L., Pan, I., Shi, L.-B., Wang, D.-C., Mei,

J., et al. (2020). Performance of radiologists in differ-

entiating covid-19 from viral pneumonia on chest ct.

Radiology.

Classiﬁcation of Chest X-ray Images to Diagnose Covid-19 using Deep Learning Techniques

Canziani, A., Paszke, A., and Culurciello, E. (2016). An

analysis of deep neural network models for practical

applications. arXiv preprint arXiv:1605.07678.

Chen, H. (1995). Machine learning for information re-

trieval: Neural networks, symbolic learning, and ge-

netic algorithms. Journal of the American Society for

Information Science, 46(3):194–216.

Cohen, J. P., Dao, L., Morrison, P., Roth, K., Bengio, Y.,

Shen, B., Abbasi, A., Hoshmand-Kochi, M., Ghas-

semi, M., Li, H., et al. (2020). Predicting covid-19

pneumonia severity on chest x-ray with deep learning.

Cureus.

Cui, S. et al. (2020). Fish detection using deep learning.

Applied Computational Intelligence and Soft Comput-

ing.

Da Silva, F. L. and Costa, A. H. R. (2019). A survey on

transfer learning for multiagent reinforcement learn-

ing systems. Journal of Artiﬁcial Intelligence Re-

search, 64:645–703.

de Oliveira, R. P. d. C., Sganderla, G. R., Maur

ıcio, C.

R. M., and Peres, F. F. F. (2019). Classiﬁcac¸ao de

imagens de raio-x de torax com reconhecimento vi-

sual da ibm cloud para diagnostico de pneumonia. In

Anais Estendidos da XXXII Conference on Graphics,

Patterns and Images, pages 203–206. SBC.

de Sousa, O. L., Magalh

aes, D. M., Vieira, P. d. A., and

Silva, R. (2020). Deep learning in image analysis for

covid-19 diagnosis: a survey. IEEE Latin America

Transactions, 100(1e).

dos Santos, Y. C. P., ESTABELECIDAS, C., and

DO NORTE, J. (2017). Desaﬁos e impacto da in-

telig

encia artiﬁcial na medicina.

Duarte, K. T. N., Gobbi, D. G., Frayne, R., and de Carvalho,

M. A. G. (2020). Detecting alzheimer’s disease based

on structural region analysis using a 3d shape descrip-

tor. In 2020 33rd SIBGRAPI Conference on Graphics,

Patterns and Images (SIBGRAPI), pages 180–187.

Hu, S., Gao, Y., Niu, Z., Jiang, Y., Li, L., Xiao, X., Wang,

M., Fang, E. F., Menpes-Smith, W., Xia, J., Ye, H.,

and Yang, G. (2020). Weakly supervised deep learn-

ing for covid-19 infection detection and classiﬁcation

from ct images. IEEE Access.

Hu, T., Khishe, M., Mohammadi, M., Parvizi, G.-R.,

Taher Karim, S. H., and Rashid, T. A. (2021).

Real-time covid-19 diagnosis from x-ray images us-

ing deep cnn and extreme learning machines stabilized

by chimp optimization algorithm. Biomedical Signal

Processing and Control, 68:102764.

Ko, H., Chung, H., Kang, W. S., Kim, K. W., Shin, Y.,

Kang, S. J., Lee, J. H., Kim, Y. J., Kim, N. Y., Jung,

H., et al. (2020). Covid-19 pneumonia diagnosis us-

ing a simple 2d deep learning framework with a single

chest ct image: Model development and validation.

Journal of Medical Internet Research, 22(6):e19569.

Kopiler, A. A. et al. (2019). Redes neurais artiﬁciais e suas

aplicac¸

oes no setor el

etrico. Revista de Engenharias

da Faculdade Salesiana, (9):27–33.

Lau, S. L. H., Wang, X., Yang, X., and Chong, E. K. P.

(2020). Automated pavement crack segmentation us-

ing fully convolutional u-net with a pretrained resnet-

34 encoder. IEEE Access.

Lei, L., Zhu, H., Gong, Y., and Cheng, Q. (2018). A

deep residual networks classiﬁcation algorithm of fe-

tal heart ct images. In Intl. Conference on Imaging

Systems and Techniques (IST), pages 1–4. IEEE.

Li, L., Qin, L., Xu, Z., Yin, Y., Wang, X., Kong, B., Bai,

J., Lu, Y., Fang, Z., Song, Q., et al. (2020). Artiﬁcial

intelligence distinguishes covid-19 from community

acquired pneumonia on chest ct. Radiology.

Negreiros, R. R. B., dos Santos, R. A., Alves, A. L. F., and

Firmino, A. A. (2020). Oil identiﬁcation on beaches

using deep learning techniques. In Anais Estendidos

do XXXIII Conference on Graphics, Patterns and Im-

ages, pages 167–170. SBC.

Ohri, K. and Kumar, M. (2021). Review on self-

supervised image recognition using deep neural net-

works. Knowledge-Based Systems, 224:107090.

oio, F. S. and Bittencourt, J. R. (2000). Sistemas in-

teligentes baseados em redes neurais artiﬁciais apli-

cados ao processamento de imagens. In I Workshop

de intelig

encia artiﬁcial.

Phankokkruad, M. (2020). Covid-19 pneumonia detection

in chest x-ray images using transfer learning of con-

volutional neural networks. In Proc. of the 3rd Intl.

Conf. on Data Science and Information Technology,

DSIT 2020, page 147–152, New York, NY, USA. As-

sociation for Computing Machinery.

Rajaraman, S., Siegelman, J., Alderson, P. O., Folio, L. S.,

Folio, L. R., and Antani, S. K. (2020). Iteratively

pruned deep learning ensembles for covid-19 detec-

tion in chest x-rays. IEEE Access.

Rodrigues, J. C. L. et al. (2020). Performance of radiolo-

gists in differentiating covid-19 from viral pneumonia

on chest ct. Public Health Emergency Collection.

Shi, F., Wang, J., et al. (2020). Review of artiﬁcial intel-

ligence techniques in imaging data acquisition, seg-

mentation and diagnosis for covid-19. IEEE reviews

in biomedical engineering.

Silva, I., Leoni, G., Sadok, D., and Endo, P. (2021). Clas-

sifying covid-19 positive x-ray using deep learning

models. IEEE Latin America Transactions, 19.

Singh, S., Ahuja, U., Kumar, M., Kumar, K., and Sachdeva,

M. (2021). Face mask detection using yolov3 and

faster r-cnn models: Covid-19 environment. Multi-

media Tools and Applications, 80(13):19753–19768.

orl, C., Castro, E., and Luchiari, A. (2011). Aplicac¸

de redes neurais artiﬁciais na construc¸

ao de modelos

de fragilidade ambiental. Revista do Departamento de

Geograﬁa, 21:113–135.

Voulodimos, A., Doulamis, N., Doulamis, A., and Protopa-

padakis, E. (2018). Deep learning for computer vi-

sion: A brief review. Computational Intelligence and

Neuroscience, 2018:7068349.

Wong, H. Y. F., Lam, H. Y. S., Fong, A. H., et al. (2020).

Frequency and distribution of chest radiographic ﬁnd-

ings in patients positive for covid-19. Radiology.

Wynants, L., Van Calster, B., Collins, G. S., Riley, R. D.,

Heinze, G., Schuit, E., Bonten, M. M. J., Dahly, D. L.,

et al. (2020). Prediction models for diagnosis and

prognosis of covid-19: systematic review and critical

appraisal. BMJ, 369.

Zhu, X. and Goldberg, A. B. (2009). Introduction to semi-

supervised learning. Synthesis lectures on artiﬁcial

intelligence and machine learning, 3(1):1–130.

WINSYS 2022 - 19th International Conference on Wireless Networks and Mobile Systems

100