Classiﬁcation of Images of Childhood Pneumonia using Convolutional

Neural Networks

A. A. Saraiva

1 a

, N. M. Fonseca Ferreira

2,3,4 b

, Luciano Lopes de Sousa

5 c

Nator Junior C. Costa

5 d

, Jos

e Vigno Moura Sousa

5,6 e

, D. B. S. Santos

5 f

, Antonio Valente

2,7

and Salviano Soares

7 h

UTAD University, Coimbra, Portugal

INESC-TEC Technology and Science, Campus da FEUP, Rua Dr. Roberto Frias 378, 4200 - 465 Porto, Portugal

Knowledge Engineering and Decision-Support Research Center (GECAD) of the Institute of Engineering,

Polytechnic Institute of Porto, Portugal

Department of Electrical Engineering, Institute of Engineering of Coimbra, Polytechnic Institute,

Rua Pedro Nunes, 3031-601 Coimbra, Portugal

State University of Piau

ı, Piau

ı, Brazil

University Brazil, S

ao Paulo, Brazil

IEETA-UA and School of Science and Technology, University of Tr

as-os-Montes and Alto Douro, Vila Real, Portugal

Keywords:

Pneumonia, X-Ray, CNN, K-Fold.

Abstract:

In this paper we describe a comparative classiﬁcation of Pneumonia using Convolution Neural Network. The

database used was the dataset Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for

Classiﬁcation made available by (Kermany, 2018) with a total of 5863 images, with 2 classes: normal and

pneumonia. To evaluate the generalization capacity of the models, cross-validation of k-fold was used. The

classiﬁcation models proved to be efﬁcient compared to the work of (Kermany et al., 2018) which obtained

92.8 % and the present work had an average accuracy of 95.30 %.

1 INTRODUCTION

Pneumonia is one of the most common causes of

death in children worldwide, accounting for 15 % of

all deaths of children under 5 years of age (Mathur

et al., 2018). Identifying and treating pneumonia has a

substantial effect on infant mortality. Although meth-

ods utilizing chest X-ray are promising modalities for

radiologic diagnosis, their role in clinical manage-

ment and their impact on outcomes need to be im-

proved (Zar et al., 2017).

https://orcid.org/0000-0002-3960-697X

https://orcid.org/0000-0002-2204-6339

https://orcid.org/0000-0003-0551-4804

https://orcid.org/0000-0001-5636-424X

https://orcid.org/0000-0002-5164-360X

https://orcid.org/0000-0003-4018-242X

https://orcid.org/0000-0002-5798-1298

https://orcid.org/0000-0001-5862-5706

For an adequate treatment, therefore, an early di-

agnosis of pneumonia is necessary, but it is not always

clear (Saraiva et al., 2018a), (Saraiva et al., 2018e). In

a review of the medical records of patients admitted

with pneumonia, 22 % of the patients presented some

reason for the uncertainty of the diagnosis, which

could result in delays in the delivery of antibiotics.

Furthermore, chest computed tomography is a gold

standard for diagnosis, but it is not always available

and is loaded with a high dose of radiation and high

cost, preventing its use in the routine diagnostic pro-

cess of patients with suspected pneumonia (Cortellaro

et al., 2012) (Saraiva et al., 2018f), (Marques et al.,

2018), (Saraiva et al., 2018c).

Several researchers use imaging for the detection

of pneumonia. Most methods use radiographic imag-

ing as a tool. For example, in (Sharma et al., 2017),

to detect clouds of pneumonia, the Otsu threshold

was used, which will separate the healthy part of the

lung from the nebulous regions infected by pneumo-

112

Saraiva, A., Ferreira, N., Lopes de Sousa, L., Costa, N., Sousa, J., Santos, D., Valente, A. and Soares, S.

Classiﬁcation of Images of Childhood Pneumonia using Convolutional Neural Networks.

DOI: 10.5220/0007404301120119

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 112-119

ISBN: 978-989-758-353-7

nia, and a detection method based on neural networks

using tomography models computerized, was used in

(Rajaraman et al., 2018). In addition to these meth-

ods, there are projects that extract information from

the cough sound analysis to diagnose cases of pneu-

monia (Amrulloh et al., 2018).

The doctor, when interpreting the chest x-ray, will

look for white patches in the lungs to detect pneu-

monia. However, such hazy patterns would also be

observed tuberculosis and severe cases of bronchitis.

For the purpose of conclusive diagnosis, the purpose

of this study is to identify pneumonia, through analy-

sis of chest X-ray images, to recognize patterns of the

disease using a neural network.

The classiﬁcation stage consists of two sub-steps,

where the ﬁrst one is the training of a Convolutional

Neural Networks (CNN), the second one is the vali-

dation of the model, that is, the test with images not

known by CNN. The method covered ensures a robust

coverage in image recognition (Saraiva et al., 2018d),

(Saraiva et al., 2018g) (Saraiva et al., 2018b), under

certain assumptions that will be clariﬁed throughout

the text. A classiﬁcation model (CNN) for the pneu-

monia and normal classes is proposed. The results

are compared with the work of (Kermany et al., 2018)

and obtained a slightly higher result, and the present

work presents cross-validation of k-fold. In order to

strengthen the results.

This article is divided into 5 sections, and section

2 is intended for related works. In section 3, it is com-

posed by the methodology, where will be explained

the steps taken for the development of the algorithm.

In section 4, it is assigned to the results obtained and,

following in the next section, in 5, where the conclu-

sion of the work is presented.

2 RELATED WORK

For the detection of pulmonary diseases chest radiog-

raphy is always required to identify pulmonary prob-

lems. Diseases such as tuberculosis, pneumonia and

lung cancer are a major threat to human health. Thus,

in (Khobragade et al., 2016) proposes pulmonary seg-

mentation, extraction of characteristics and its classi-

ﬁcation using an artiﬁcial neural network for the de-

tection of pulmonary diseases. (Khobragade et al.,

2016) a simple image processing technique with in-

tensity based method was used, and a method based

on the discontinuity to detect pulmonary limits, in this

way, statistical and geometric characteristics were ex-

tracted. Neural networks were used feed forward and

back propagation to detect diseases.

Pneumonia is one of the leading causes of infant

mortality. In developing countries there is little in-

frastructure and doctors in rural areas to provide the

necessary diagnosis. Therefore, in (Barrientos et al.,

2016) proposes a method for automatic diagnosis us-

ing ultrasonography of the lungs. The approach pre-

sented is based on the analysis of patterns in rectangu-

lar segments in the image of the ultrasonography. The

speciﬁc characteristics and characteristic vectors were

obtained and classiﬁed by a standard neural network.

In (Barrientos et al., 2016) I obtained a sensitivity of

91.5 % and speciﬁcity of 100 % but were extracted

from a single patient and only included in the test or

in the training set.

Many researchers have developed several algo-

rithms for the diagnosis of lung diseases through

sound. One of the parameters used for the detection of

pulmonary sound is entropy, so there are differences

in the sound of a normal respiratory system and a sys-

tem with pathologies. In the article (Rizal et al., 2017)

discurses several measures of entropy for a classiﬁca-

tion of pulmonary sounds. The result in (Rizal et al.,

2017) shows that the use of a single entropy could not

achieve high accuracy, so 7 entropies were used and

guaranteed 94.95 % accuracy using multilayer per-

ceptron.

In paper (Rodrigues et al., 2018) suggests a

Structural Co-occurrence Matrix (SCM) approach to

classify malignant nodules or benign nodules and

also their level of malignancy. The structural co-

occurrence matrix technique was applied to extract

characteristics of the nodule images and classify

them. The SCM was applied in gray scale and images

of the Hounsﬁeld unit with four ﬁlters, creating eight

different conﬁgurations. The classiﬁcation stage used

classiﬁers known as the multilayer perceptron, sup-

port vector machine, k-Nearest Neighbors algorithm

and were applied in two tasks: (i) to classify the nod-

ule images as malignant or benign, (ii) to classify the

nodules pulmonary lesions at the level of malignancy

(1 to 5). O (Rodrigues et al., 2018) had a result of

96.7 % for precision and F-score measurements in the

ﬁrst task and 74.5 % accuracy and 53.2 % F-score in

the second task.

The (Santosh and Antani, 2018) has proposed an

idea that takes into account the alterations of the right

and left lung region in terms of symmetry and auto-

mated the chest X-ray system for the evidence of tu-

berculosis. The proposed method is the observation of

radiological exams leading to bilateral comparisons

in the lung ﬁeld. In (Santosh and Antani, 2018) an-

alyzed the symmetric lung region using a multiple-

scale shape feature, as well as border texture char-

acteristics. Three different types of classiﬁers were

used: Bayesian network, multilayer network percep-

Classiﬁcation of Images of Childhood Pneumonia using Convolutional Neural Networks

113

tron and random forest. The results obtained with an

abnormality detection accuracy of 91 % and area un-

der the ROC curve of 0.96.

Many researchers use various methods for detect-

ing diseases based on lung sound, for example the use

of entropy measurement. Sound of pulmonary snor-

ing is a sound that is discontinuous, of short duration

and appears in the inspiratory, expiratory or in both

cases. Thus, in (Rizal et al., 2016) the Tsallis entropy

was used as the characteristic extraction method to

classify lung sounds. The results were achieved us-

ing at least three Tsallis entropy values with q = 2, 3,

and 4 with MLP as a classiﬁer and three-fold cross

validation at an accuracy of 95.35 %, sensitivity of

90.48 % and 100 % speciﬁcity, were achieved.

Lung cancer accounts for 26 % of all cancer

deaths in 2017, accounting for more than 1.5 mil-

lion deaths. Thus, in (d. N

obrega et al., 2018) is

proposed to explore the performance of deep transfer

learning to classify pulmonary nodule malignancies.

In this work, we have used a convolutional neural

network such as VGG16, VGG19, MobileNet, Xcep-

tion, InceptionV3, ResNet50, Inception-ResNet-V2,

DenseNet169, DenseNet201, NASNetMobile and

NASNetLarge, where they were used to extract pa-

rameters from an image database of the lung. The

characteristics were classiﬁed using Naive Bayes,

Perceptron multilayer, support vector machine, Near

Neighbors and Random Forest. The results obtained

in (d. N

obrega et al., 2018) were 88.41 % of ACC and

93.19 % of AUC.

(Paing and Choomchuay, 2017) has the objective

of detecting pulmonary nodules from a series of dig-

itized CT images. The threshold and morphological

operations of Otsu are applied for the segmentation

of nodules. In (Paing and Choomchuay, 2017), after

segmentation, objects that can not be nodes are dis-

carded. In view of this, multilayer Perceptron was

used for the classiﬁcation and 95 % accuracy was

achieved.

In the work (Kermany et al., 2018) a method was

proposed using Convolutional Neural Networks with

a transfer learning technique. Transfer learning has

proven to be a highly effective technique, particularly

when confronted in domains with limited data. The

main application of the transfer learning algorithm

was in the diagnosis of OCT images of the retina, but

it was also tested in pediatric thoracic radiographs. In

case of retinal OCT, in a multiclass comparison be-

tween choroidal neovascularization, diabetic macular

edema, drusen and normal, obtained a precision of

96.6 %, with a sensitivity of 97.8 %, a speciﬁcity of

97, 4 %. In the comparison of chest X-rays with pneu-

monia versus normal, it obtained an accuracy of 92.8

%, with sensitivity of 93.2 % and speciﬁcity of 90.1

The convolutional neural networks offer great

help in detecting diseases such as (Lisowska et al.,

2017), (Ponzio et al., 2018) and (Mabaso et al., 2018).

Therefore, a 3D Convolutional Neural Network for

the detection of subtle signs of a Stroke has been pro-

posed in (Lisowska et al., 2017), thus, the architecture

was developed to explore contralateral resources and

anatomical information from the cerebral atlas. In the

work (Ponzio et al., 2018), a deep learning technique

based on Convolutional Neural Networks was pro-

posed to differentiate adenocarcinomas from healthy

tissues and benign lesions. In the article (Mabaso

et al., 2018), it presents an automated way of detect-

ing and counting points in microscopic images based

on a Convolutional Neural Network based on a slid-

ing window for detection of multiple points in micro-

scopic images.

Before these works, in this one, the objective is the

construction of an algorithm that recognizes the pneu-

monia, from x-ray images of the thoracic region. In

the algorithm, a Convolutional Neural Network was

used to recognize patterns in the images of people in-

fected by pneumonia, either by virus or bacteria.

3 METODOLOGY

3.1 Description of the Dataset

In the dataset (Kermany, 2018) used for training and

validation, it contained 5863 X-ray (JPEG) images

and two categories: Normal and Pneumonia. A total

of 5863 patients (Junge and Dettori, 2018)

The radiographic images were from pediatric pa-

tients one to ﬁve years old from the Medical Center in

Guangzhou. In this way, radiographs were performed

as part of clinical care.

All images in dataset (Kermany, 2018) underwent

a treatment in order to remove all low-quality scans,

as well as being classiﬁed by two specialist physicians

and by a third party specialist, in order to prevent any

misclassiﬁcation.

3.1.1 Pneumonia (Bacterial and Viral) and

Normal

For the diagnosis of pneumonia, the alveoli become

ﬁlled with secretion and appear as a white spot on

the chest radiograph. Pulmonary consolidation means

that the pulmonary alveoli are ﬁlled with inﬂamma-

tory ﬂuid. In radiography, pulmonary consolidation

corresponds to an opacity, that is, the whitish area. As

BIOIMAGING 2019 - 6th International Conference on Bioimaging

114

shown in the Figure 1 (II), the X-ray of the thoracic re-

gion of patients presenting with pneumonia and X-ray

of patients under normal conditions, shown in Figure

1 (I).

Figure 1: Exemple normal and pneumonia.

3.2 Input Structure for Learning Model

In Figure 2 the classiﬁcation structure is represented,

which consists of the image entry in the network, re-

sponsible for classiﬁcation in pneumonia or normal.

Images are converted to grayscale and normalized so

that they are between 0 and 1, making them a 32-bit

ﬂoating point.

Figure 2: Exemple input image.

3.3 Training Structure

The training and validation structure has three steps,

represented in Figure 3, the ﬁrst one is the division

and normalization of the images. In the second one

the training of the images is realized. Network vali-

dation is performed using the test data.

To validate the model, cross-validation of k-fold

is used, which consists of dividing the images into

two sets: test and training. The process is repeated 5

times, changing the test and training images after the

calculation of k.

3.4 Proposed Model Convolution

Neural Network (CNN)

In the ﬁrst stage we have the convolutional layer,

which is one of the main layers, where it is the extrac-

tion of characteristics of the image through a series of

Figure 3: Training structure and validation.

ﬁlters (kernels). A ﬁlter is an array of values, called

weights, trained to detect speciﬁc features. The ﬁlter

moves over each part of the image to see if the feature

that it should detect is present. When the feature is

present in part of an image, the convolution operation

between the ﬁlter and that part of the image results in

a real number with a high value. If the resource is not

present, the resulting value will be low. To provide a

value that represents the conﬁdence that a particular

resource is present, the ﬁlter performs a convolution

operation, which is an elementary product and sums

between two arrays.

Given a two-dimensional image, I, and a small ar-

ray, K of size h x w (kernel), the convoked image,

I ∗K, is calculated by superimposing the kernel at the

top of the image all possible forms, and recording the

sum of the elementary products between the image

and the kernel equation 1.

(I ∗ K)

∑

i=1

∑

j=1

i j

x+i−1,y+ j−1

(1)

After the convolution layer is the pooling layer,

which is responsible for reducing the spatial size of

the feature map, preserving the resources detected

in a smaller representation. There are several pool-

ing layer options, with maxpooling being the most

popular. Basically, maxpooling operates by locating

the locations in the image that show the strongest

correlation for each resource (the maximum value)

are preserved and those maximum values combine to

Figure 4: Example max-pooling with a 8x8 image.

Classiﬁcation of Images of Childhood Pneumonia using Convolutional Neural Networks

115

form a smaller space as shown in Figure 4.In general,

they are used after the convolutional layers, both with

the objective of progressively reducing computational

costs in the network, as well as minimizing the prob-

ability of overﬁtting.

After the pooling layer, we ﬁnd the fully con-

nected layers (FCN) that are used to make ﬁnal pre-

dictions. A FCN layer obtains ”resources” in a vector

form from a previous resource extraction layer, mul-

tiplies a weight matrix, and generates a new resource

vector whose computation pattern is a dense matrix-

vector multiplication (Zhang et al., 2018).Some FCNs

are used in a cascade mode that ultimately produce

the CNN classiﬁcation result that generates a proba-

bility (a number ranging from 0 to 1) for each of the

classiﬁcation labels that the model is trying to pre-

dict. Sometimes multiple input vectors are processed

simultaneously in a single batch to increase the over-

all throughput as shown in the following expression

when the batch size h is greater than 1. Note that the

FCN layers are also the major components of neu-

ral networks (DNNs) that are widely used in speech

recognition (Zhang et al., 2018).

Out[m][h] =

∑

n=0

W [m][n] ∗ In[n][h] (2)

3.4.1 Categorical Cross-entropy

The cross-entropy loss measures the performance of

a classiﬁcation model, with the output being a proba-

bility value ranging from 0 to 1 (Zhang and Sabuncu,

2018).

H(p, q) = −

∑

p(x) log q(x). (3)

3.4.2 CNN Architecture

In Figure5 is the CNN architecture used, it has 10 lay-

ers, the ﬁrst seven convolutionals and the last three

without convolution with the softmax activation func-

tion (Peng et al., 2017) equation 4, the network in-

put receives a 300x300 pixel image, each convolution

layer has the ReLUs enable function. For the convo-

lution kernel, the 3x3 size was adopted, because this

way it is possible to have a greater precision in the

time to go through the entire image.

After two convolutional layers a Max-pooling

layer is used, this reduces the size of the matrices re-

sulting from the convolution. With this layer it is pos-

sible to reduce the amount of parameters that will be

learned by the network, this way it is done overﬁtting

control.

In the last layer the softmax activation function is

used (Peng et al., 2017) equation 4, this function is

responsible for making the probabilistic distribution

of the input image belong to each of the classes in

which the network was trained. To reduce the train-

ing time and to avoid overﬁtting is used dropout in the

layer, ie it is randomly removed at each training inter-

action, a certain percentage of the neurons of a layer,

re-adding them in the following iteration. The loss

function used was equation 3 and the optimization

function used was ADAM (Kingma and Ba, 2014)

σ(z)

∑

k=1

f or j = 1, . . . , K. (4)

4 METRICS OF THE

EVALUATION

4.1 Cross Validation

Cross-validation is one of the fundamental methods

in machine learning for method evaluation and pa-

rameter selection in a machine prediction or learning

task. Thus, K-fold cross validation was used to eval-

uate this model.

In K-fold cross-validation, the sample is randomly

divided into k sets of equal sizes. In each of the k

shares, a single set is separated as the validation data

to test the model, and the remaining sets of k − 1

are used as training data (Fan and Hauser, 2018).

The cross-validation process is then repeated k times,

with each of the k sets used exactly once for valida-

tion (Fan and Hauser, 2018).The mean performance is

then used as the evaluation index of the method stud-

ied. This approach can be computationally expensive,

but it takes full advantage of the entire set of data,

especially when the number of samples is very small

(Men et al., 2018). This approach can also demon-

strate how the trained model is generalizable for un-

seen data in order to avoid deliberate choice of data

with superior test results (Men et al., 2018).

4.2 Confusion Matrix

The confusion matrix is an array that contains correct

and incorrect predictions of the algorithm and the ac-

tual situation. As shown in table 1.

• True Positive: Number of people who actually

have pneumonia according to the algorithm.

• False Negative: Number of people who are ac-

tually with pneumonia but categorized as healthy

according to the algorithm.

BIOIMAGING 2019 - 6th International Conference on Bioimaging

116

Figure 5: Construction of the CNN training mode.

Table 1: Confusion Matrix.

• False Positive: Number of people who are actu-

ally healthy, but categorized as pneumonia, ac-

cording to the algorithm.

• True Negative: Number of people who are really

healthy and categorized as healthy according to

the algorithm.

4.3 ROC Curve

The ROC curve is a graphical representation that vi-

sualizes the relationship between the true positive rate

and the false positive rate for a classiﬁer under vari-

ous decision thresholds. Thus, a ROC curve is con-

ceptually equivalent to a curve that shows the rela-

tionship between the power of the test and the prob-

ability of error, with the variation of the cutoff value

of a statistical test. Therefore, the ROC curve com-

pares classiﬁer performance across the full range of

class distributions, offering an assessment, covering

a wide range of operating conditions (Brzezinski and

Stefanowski, 2017).

ROC curves are an important tool in assessing the

uncertainty format, and are a valuable method in char-

acterizing the strengths and weaknesses of diagnostic

tests (Junge and Dettori, 2018).

5 RESULTS

In this section will be presented the results obtained

by the classiﬁcation model described above, The met-

rics used reinforce the results, in the table 2 CNN per-

formance results compared to the work of (Kermany

et al., 2018). In Figure 7 you can view a graph with

the ROC curve. In the ﬁgure 8 the confusion matrix

for iteration 1 of the table 2.

Figure 6: ROC curve interaction 1 reference table 2.

Figure 7: Precision Recall curve interaction 1 reference ta-

ble 2.

6 CONCLUSIONS

In this paper we have demonstrated a comparison with

the work of (Kermany et al., 2018) in the detection

and classiﬁcation of images for the detection of pneu-

monia from the chest X-ray of patients. The Convo-

Classiﬁcation of Images of Childhood Pneumonia using Convolutional Neural Networks

117

lutional Neural Network was used to train the neural

network and, for the validation of the model, Cross

validation was used. The classiﬁcation model pre-

sented was efﬁcient in the classiﬁcation, obtaining an

average accuracy of 95.30 % in the tests against 92.8

% of the work of (Kermany et al., 2018).

Figure 8: Confusion matrix reference interaction 1 table 2.

Table 2: Accuracy of the interactions the model learning.

** Normal Pneumonia ACC

1 Normal 120 24

1 Pneumonia 5 435

1 95.03 %

2 Normal 131 33

2 Pneumonia 2 418

2 94.00 %

3 Normal 117 25

3 Pneumonia 3 439

3 95.20 %

4 Normal 142 17

4 Pneumonia 5 420

4 96.23 %

5 Normal 132 19

5 Pneumonia 2 431

5 96.04 %

** (Kermany et al., 2018) 92.8 %

** Average accuracy this paper 95.30 %

ACKNOWLEDGMENTS

This work is ﬁnanced by National Funds through the

FCT - Fundac¸

ao para a Ci

encia e a Tecnologia (Por-

tuguese Foundation for Science and Technology) as

part of project UID/EEA/00760/2019.

REFERENCES

Amrulloh, Y. A., Abeyratne, U. R., Swarnkar, V., Herath,

D., Triasih, R., and Setyati, A. (2018). Hmm

based cough sound analysis for classifying pneumo-

nia and asthma in pediatric population. In Eskola,

H., V

ais

anen, O., Viik, J., and Hyttinen, J., editors,

EMBEC & NBC 2017, pages 852–855, Singapore.

Springer Singapore.

Barrientos, R., Roman-Gonzalez, A., Barrientos, F., Solis,

L., Correa, M., Pajuelo, M., Anticona, C., Lavarello,

R., Casta

neda, B., Oberhelman, R., Checkley, W.,

Gilman, R. H., and Zimic, M. (2016). Automatic

detection of pneumonia analyzing ultrasound digital

images. In 2016 IEEE 36th Central American and

Panama Convention (CONCAPAN XXXVI), pages 1–

Brzezinski, D. and Stefanowski, J. (2017). Prequential auc:

properties of the area under the roc curve for data

streams with concept drift. Knowledge and Informa-

tion Systems, 52(2):531–562.

Cortellaro, F., Colombo, S., Coen, D., and Duca, P. G.

(2012). Lung ultrasound is an accurate diagnostic tool

for the diagnosis of pneumonia in the emergency de-

partment. Emergency Medicine Journal, 29(1):19–23.

d. N

obrega, R. V. M., Peixoto, S. A., d. Silva, S. P. P., and

Filho, P. P. R. (2018). Lung nodule classiﬁcation via

deep transfer learning in ct lung images. In 2018

IEEE 31st International Symposium on Computer-

Based Medical Systems (CBMS), pages 244–249.

Fan, C. and Hauser, H. (2018). Fast and accurate cnn-based

brushing in scatterplots. In Computer Graphics Fo-

rum, volume 37, pages 111–120. Wiley Online Li-

brary.

Junge, M. R. J. and Dettori, J. R. (2018). Roc solid: Re-

ceiver operator characteristic (roc) curves as a founda-

tion for better diagnostic tests. Global Spine Journal,

8(4):424–429.

Kermany, D. S., Goldbaum, M., Cai, W., Valentim, C. C.,

Liang, H., Baxter, S. L., McKeown, A., Yang, G.,

Wu, X., Yan, F., et al. (2018). Identifying medical

diagnoses and treatable diseases by image-based deep

learning. Cell, 172(5):1122–1131.

Kermany, Daniel; Zhang, K. G. M. (2018). Labeled optical

coherence tomography (oct) and chest x-ray images

for classiﬁcation. Mendeley Data.

Khobragade, S., Tiwari, A., Patil, C. Y., and Narke, V.

(2016). Automatic detection of major lung diseases

using chest radiographs and classiﬁcation by feed-

forward artiﬁcial neural network. In 2016 IEEE 1st In-

ternational Conference on Power Electronics, Intelli-

gent Control and Energy Systems (ICPEICES), pages

1–5.

Kingma, D. P. and Ba, J. (2014). Adam: A method for

stochastic optimization. CoRR, abs/1412.6980.

Lisowska, A., Beveridge, E., Muir, K., and Poole, I. (2017).

Thrombus detection in ct brain scans using a convolu-

tional neural network. In BIOIMAGING, pages 24–33.

Mabaso, M. A., Withey, D. J., and Twala, B. (2018). Spot

BIOIMAGING 2019 - 6th International Conference on Bioimaging

118

detection in microscopy images using convolutional

neural network with sliding-window approach.

Marques, J. F., das Chagas, Fontenele, A. A., Costa, J.

V. M., De Araujo, N. F., and Valente, A. (2018). Ma-

nipulation of bioinspiration robot with gesture recog-

nition through fractional calculus. IEEE LARS 2018 –

15th Latin American Robotics Symposium.

Mathur, S., Fuchs, A., Bielicki, J., Anker, J. V. D., and

Sharland, M. (2018). Antibiotic use for community-

acquired pneumonia in neonates and children: Who

evidence review. Paediatrics and International Child

Health, 38(sup1):S66–S75. PMID: 29790844.

Men, K., Geng, H., Cheng, C., Zhong, H., Huang, M.,

Fan, Y., Plastaras, J. P., Lin, A., and Xiao, Y. (2018).

More accurate and efﬁcient segmentation of organs-

at-risk in radiotherapy with convolutional neural net-

works cascades. Medical physics.

Paing, M. P. and Choomchuay, S. (2017). A computer aided

diagnosis system for detection of lung nodules from

series of ct slices. In 2017 14th International Con-

ference on Electrical Engineering/Electronics, Com-

puter, Telecommunications and Information Technol-

ogy (ECTI-CON), pages 302–305.

Peng, H., Li, J., Song, Y., and Liu, Y. (2017). Incrementally

learning the hierarchical softmax function for neural

language models. In AAAI, pages 3267–3273.

Ponzio, F., Macii, E., Ficarra, E., and Di Cataldo, S. (2018).

Colorectal cancer classiﬁcation using deep convolu-

tional networks.

Rajaraman, S., Candemir, S., Kim, I., Thoma, G., and An-

tani, S. (2018). Visualization and interpretation of

convolutional neural network predictions in detecting

pneumonia in pediatric chest radiographs. Applied

Sciences, 8(10).

Rizal, A., Hidayat, R., and Nugroho, H. A. (2016). Pul-

monary crackle feature extraction using tsallis entropy

for automatic lung sound classiﬁcation. In 2016 1st

International Conference on Biomedical Engineering

(IBIOMED), pages 1–4.

Rizal, A., Hidayat, R., and Nugroho, H. A. (2017). Entropy

measurement as features extraction in automatic lung

sound classiﬁcation. In 2017 International Confer-

ence on Control, Electronics, Renewable Energy and

Communications (ICCREC), pages 93–97.

Rodrigues, M. B., N

obrega, R. V. M. D., Alves, S. S. A.,

Filho, P. P. R., Duarte, J. B. F., Sangaiah, A. K., and

Albuquerque, V. H. C. D. (2018). Health of things

algorithms for malignancy level classiﬁcation of lung

nodules. IEEE Access, 6:18592–18601.

Santosh, K. C. and Antani, S. (2018). Automated chest

x-ray screening: Can lung region symmetry help de-

tect pulmonary abnormalities? IEEE Transactions on

Medical Imaging, 37(5):1168–1177.

Saraiva, A., Barros, M., Nogueira, A., Fonseca Ferreira, N.,

and Valente, A. (2018a). Virtual interactive environ-

ment for low-cost treatment of mechanical strabismus

and amblyopia. Information, 9(7):175.

Saraiva, A., Miranda de Jesus Castro, F., Ferreira, N.,

and Valente, A. (2018b). Compression of electro-

cardiographs comparative study between the walsh

hadamard transform and discrete cosine transform.

Saraiva, A. A., , N. F. F., Salviano F.S.P. Soares and, M. J.

C. S. R., and Antonio, V. (2018c). Filtering of cardiac

signals with mathematical morphology for qrs detec-

tion. Proceedings of ICAT’18, 7th International Con-

ference on Advanced Technologies.

Saraiva, A. A., Costa, N., Sousa, J. V. M., De Araujo, T. P.,

Ferreira, N. F., and Valente, A. (2018d). Scalable task

cleanup assignment for multi-agents. In Memorias de

Congresos UTP, volume 1, pages 439–446.

Saraiva, A. A., Ferreira, N. M. F., and Valente, A. (2018e).

New bioinspired ﬁlter of dicom images. In Proceed-

ings of the 11th International Joint Conference on

Biomedical Engineering Systems and Technologies -

Volume 1: BIODEVICES,, pages 258–265. INSTICC,

SciTePress.

Saraiva, A. A., Nogueira, A. T., Ferreira, N. F., and Valente,

A. (2018f). Application of virtual reality for the treat-

ment of strabismus and amblyopia. In 2018 IEEE 6th

International Conference on Serious Games and Ap-

plications for Health (SeGAH), pages 1–7. IEEE.

Saraiva, A. A., SANTOS, D. S., JUNIOR, F. M., SOUSA,

J. V. M., FERREIRA, N. F., and Valente, A.

(2018g). Navigation of quadruped multirobots by ges-

ture recognition using restricted boltzmann machines.

In Memorias de Congresos UTP, volume 1, pages

431–438.

Sharma, A., Raju, D., and Ranjan, S. (2017). Detection of

pneumonia clouds in chest x-ray using image process-

ing approach. In 2017 Nirma University International

Conference on Engineering (NUiCONE), pages 1–4.

Zar, H. J., Andronikou, S., and Nicol, M. P. (2017). Ad-

vances in the diagnosis of pneumonia in children.

BMJ, 358.

Zhang, C., Sun, G., Fang, Z., Zhou, P., Pan, P., and Cong,

J. (2018). Caffeine: Towards uniformed representa-

tion and acceleration for deep convolutional neural

networks. IEEE Transactions on Computer-Aided De-

sign of Integrated Circuits and Systems, pages 1–1.

Zhang, Z. and Sabuncu, M. R. (2018). Generalized cross

entropy loss for training deep neural networks with

noisy labels. CoRR, abs/1805.07836.

Classiﬁcation of Images of Childhood Pneumonia using Convolutional Neural Networks

119