technologies

Article
Multi-Classification of Lung Infections Using Improved
Stacking Convolution Neural Network
Usharani Bhimavarapu 1 , Nalini Chintalapudi 2 and Gopi Battineni 2, *

1 Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation,

Vaddeswaram 522302, Andhra Pradesh, India; [email protected]
2 Clinical Research Center, School of Medicinal and Health Products Sciences, University of Camerino,
62032 Camerino, Italy; [email protected]
* Correspondence: [email protected]

Abstract: Lung disease is a respiratory disease that poses a high risk to people worldwide and
includes pneumonia and COVID-19. As such, quick and precise identification of lung disease is
vital in medical treatment. Early detection and diagnosis can significantly reduce the life-threatening
nature of lung diseases and improve the quality of life of human beings. Chest X-ray and computed
tomography (CT) scan images are currently the best techniques to detect and diagnose lung infection.
The increase in the chest X-ray or CT scan images at the time of training addresses the overfitting
dilemma, and multi-class classification of lung diseases will deal with meaningful information and
overfitting. Overfitting deteriorates the performance of the model and gives inaccurate results. This
study reduces the overfitting issue and computational complexity by proposing a new enhanced
kernel convolution function. Alongside an enhanced kernel convolution function, this study used
convolution neural network (CNN) models to determine pneumonia and COVID-19. Each CNN
model was applied to the collected dataset to extract the features and later applied these features as
input to the classification models. This study shows that extracting deep features from the common
layers of the CNN models increased the performance of the classification procedure. The multi-class
classification improves the diagnostic performance, and the evaluation metrics improved significantly
Citation: Bhimavarapu, U.; with the improved support vector machine (SVM). The best results were obtained using the improved
Chintalapudi, N.; Battineni, G. SVM classifier fed with the features provided by CNN, and the success rate of the improved SVM
Multi-Classification of Lung was 99.8%.
Infections Using Improved Stacking
Convolution Neural Network. Keywords: CNN; deep learning; activation; lung disease; pneumonia; COVID-19
Technologies 2023, 11, 128.
https://doi.org/10.3390/
technologies11050128

Academic Editors: Yudong Zhang 1. Introduction

and Zhengchao Dong Lung disease is a leading cause of death across the globe, and pneumonia, one of the
Received: 23 August 2023
most common infectious diseases, is also one of the leading causes of mortality world-
Revised: 11 September 2023 wide [1]. Pneumonia is an infection that occurs in the lungs, and every year it affects
Accepted: 15 September 2023 approximately 7% of the global population, and four million patients face mortality risks [2].
Published: 17 September 2023 Using imaging tests can help doctors diagnose abnormalities. However, lung disease is on
the rise, making it difficult for doctors to diagnose it due to a shortage of radiologists. Early
detection and diagnosis reduce life-threatening conditions and improve quality of life [3].
Respiratory system symptoms are the first indicator of pneumonia [4]. An X-ray of
Copyright: © 2023 by the authors. the chest, a CT scan, a blood gas analysis, and a complete blood image are all methods
Licensee MDPI, Basel, Switzerland. used to diagnose pneumonia. Medical professionals can view the internal structures of the
This article is an open access article body using chest X-rays and computed tomography (CT) images. It is possible for medical
distributed under the terms and
professionals to fail to diagnose minor lesions [5]. It has been shown that chest X-ray images
conditions of the Creative Commons
and CT scans are sensitive to pneumonia and infectious diseases, like COVID-19 [6].
Attribution (CC BY) license (https://
By using computer-aided diagnosis (CAD), diagnosis times can be reduced, and
creativecommons.org/licenses/by/
diagnostic convenience can be increased. Diagnostic performance is improved with CAD
4.0/).

Technologies 2023, 11, 128. https://doi.org/10.3390/technologies11050128 https://www.mdpi.com/journal/technologies

Technologies 2023, 11, 128 2 of 17

systems, which minimize misreading and allow medical professionals to provide precise
diagnoses. Using CT scan images and chest X-ray images obtained from lung-infected
disease patients, deep learning models have provided more efficient results than traditional
models [7]. In this work, the proposed CAD system extracts the most efficient features of
pneumonia and contributes positively to classification results. Image processing-based
feature extraction is complex compared to deep learning models, and the performance of
these models surpasses the traditional models.
The Infnet model [8] was focused on lung infection segmentation and failed in terms
of clinical management as it frequently needed to classify and then segment the infected
area for further treatment. Unet [9] segments the COVID-19 infected regions but fails to con-
duct the semantic segmentation in order to estimate clinical management, the robustness,
and the performance. The transfer learning approach [10] segments the COVID-19 lung
infection but fails to reduce the computational complications. The coronavirus pandemic
affected many people across the world and created great challenges for international health
management [11,12]. As such, continual screening, earlier diagnosis, and timely actions are
significant for controlling the spread of disease, as well as for reducing mortality. An exact
segmentation of all kinds of lung infections could help the health professionals or doctors
to prescribe the correct specific treatment. CNN models evaluate many attributes of the
mammalian visual context that were not visible and identified by radiologists [13].
CNN models achieve high accuracy for image datasets by adjusting weights and
parameters [14,15]. In [16], the authors implemented the CNN model and achieved 70%
classification accuracy. Similarly, the CNN model was implemented to classify pneumonia
cases with 98% accuracy [17]. The Inception v3 [18,19] model also identified as a potential
model to classify pneumonia [20]. Other studies implemented nine different transfer
learning approaches to classify COVID-19 and pneumonia and achieved 95% accuracy [21].
Using EMCNet to extract the features and later applying the ensemble approach to classify
COVID-19 cases achieved an accuracy of 98.91% [22]. Furthermore, transfer learning
classifiers help to identify COVID-19 with an accuracy of 81% [23].
CNN models do have some restrictions, though. In order to attain observable perfor-
mance, CNN models require extensive network training. Additionally, the dataset being
used for training should include all the different sample variants so that the model can be
simply generalized to previously unknown data, although this is not always attainable.
Overfitting, which results in generalization mistakes, may happen if DL-based models are
trained on a smaller dataset. To overcome this drawback, we proposed an improved kernel
in the CNN, which reduces the overfitting.
Most state-of-the-art models focus on the binary classification of either pneumonia
or COVID-19, and limited research on the multi-class classification of COVID-19 and
pneumonia from a chest X-ray and CT scan images has been carried out. The major
limitation of the existing works is prioritizing the dataset before feeding it to the classifier.
This study proposed an improved kernel convolution function to improve accuracy and
reduce complexity. Later, the proposed improved SVM can better classify the lesions and
reduce the complexity, which increases accuracy.
Motivation
The detection of the infected lung regions from the chest X-ray and the CTscan images
provides high potential for increasing classical healthcare planning to tackle lung diseases,
like COVID-19. In the existing studies, the infected lung region is frequently unclear and,
thus, it is very complicated for doctors and healthcare professionals to differentiate the
normal images from the infected images. This motivated us to propose an approach for the
identification of the lung infected areas in CT scan and chest X-ray images.
Research Contributions
1. We propose an improved integrated image enhancement approach to enhance the
image contrast.
2. We propose an improved kernel in the stacking CNN to detect lung infection using
CT scans and chest X-rays.
 Technologies 2023, 11, x FOR PEER REVIEW 3 of 19

1. We propose an improved integrated image enhancement approach to enhance the

Technologies 2023, 11, 128
image contrast. 3 of 17
2. We propose an improved kernel in the stacking CNN to detect lung infection using
CT scans and chest X-rays.
3. The proposed model is compared to the existing models utilizing the CT scan and
3. The proposed model is compared to the existing models utilizing the CT scan and
chest X-ray images.
chest X-ray images.
4. A comparison of the proposed model with other models shows that the proposed
4. A comparison of the proposed model with other models shows that the proposed
models have the highest performance.
models have the highest performance.
The rest of the paper is organised as follows. Section 2 describes the dataset collection
The rest of the paper is organised as follows. Section 2 describes the dataset collection
and the proposed approach in detail. Results analysis and comparison with the existing
and the proposed approach in detail. Results analysis and comparison with the existing
models are presented in Section 3. Section 4 discusses the advantages of the proposed
models are presented in Section 3. Section 4 discusses the advantages of the proposed
work. Section 5 concludes the proposed work and the future work needed to overcome
work. Section 5 concludes the proposed work and the future work needed to overcome the
the limitations of the proposed work.
limitations of the proposed work.

2.2.Methods
Methods
Weaim
We aimto tocategorize
categorizeCTCTscans
scansandandchest
chestX-rays
X-raysinto
intodifferent
differentseverity
severitycategories
categoriesin
in
thecurrent
the currentstudy.
study.AnAnautomated
automatedsystem
systemforforassessing
assessinglung
lung disease
disease severity
severity was
was discussed.
discussed.
Accordingto
According tocurrent
currentresearch,
research,the
themodified
modifiedCNN CNNarchitecture
architectureimproved
improvedthe theaccuracy
accuracyofof
lungdisease
lung diseaseclassification.
classification.We
Wedivided
dividedthethe methodology
methodology into
into three
three phases:
phases: pre-processing,
pre-processing,
segmentation, and
segmentation, and classification.
classification. In the data
data pre-processing
pre-processing phase,
phase, we proposed
proposed an en-en-
hanced Gaussian distribution to improve the lesions’ visual contents. In the
hanced Gaussian distribution to improve the lesions’ visual contents. In the segmentationsegmentation
phase,we
phase, weproposed
proposedan anenhanced
enhancedCNN CNNmodel
modelto tosegment
segmentthethelesions
lesionsin
inthe
theCT
CTscan
scanand
and
chestX-ray
chest X-rayimages.
images. After
After segmentation,
segmentation, we we performed
performed multi-class
multi-class classification
classification by
by imple-
imple-
mentingthe
menting theimproved
improvedSSVM.
SSVM.An Anillustration
illustrationofofthe
theproposed
proposedapproach
approachflow
flowcancanbe
befound
found
inFigure
in Figure1.1.

Figure1.1.Methodology
Figure Methodologyflow.
flow.

2.1. Dataset
CT scan images were collected with three different sources [24–26]. We removed the
repeated images and unrelated body parts (e.g., head) that are not related to the lung. The
cleaned final dataset contains 1043 CT scan images, and each class consists of pneumonia
(n = 359), normal (n = 333), and COVID-19 (351) samples.
Additionally, 5935 chest X-ray images from two different source are collected [17,18].
We divided the training and testing data in the ratio of 85:15, which consist of 4747 for
training and 1188 for testing. This training dataset consists of 3 classes: 63 images of
 Technologies 2023, 11, 128 4 of 17

COVID-19, 1266 normal, and 3418 pneumonia chest X-ray images. This testing dataset
consists of 3 classes: 16 images of COVID-19, 317 of normal, and 855 of pneumonia chest
X-ray images. Table 1 tabulates the splitting of training and testing sets of chest X-ray and
CT scan datasets.

Table 1. Dataset image distribution.

Chest X-ray CT Scan

Class
Training Testing Training Testing
Normal 1266 317 266 67
Pneumonia 3418 855 287 72
COVID-19 63 16 281 70
Total 4747 1188 834 209

2.2. Data Pre-Processing

In this study we used the improved ensemble approach to enhance the images with
the integration of two approaches of top hat filtering and a Wiener filter. Given a database
of k images, it is represented as D k . Each image is defined as I k ( p, q) with dimension PXQ,
where P and Q have a value of 512. Figure 2 shows the unprocessed and the processed CT
scan images. The mathematical operation of the top hat filtering is as follows:
1, x FOR PEER REVIEW 5 of 19
I t ( p, q) = I k ( p, q) − ( I k ( p, q)os) (1)
where s represents a structuring element with value 21 and o denotes the opening operator.

Figure 2. Image enhancement. (a) Unprocessed

Figure 2. Image enhancement. images
(a) Unprocessed andand
images (b)(b)
processed images.
processed images.

Equation (1) improves the contrast of the image. Next, an improved Weiner filter
2.2.1. EnhancedisCNN
used to remove the noise in the medical images. The mathematical representation is
as follows:
We developed an enhanced CNN model to detect
w( p.q)lung illness accurately. The pro-
W m ( p, q) = (2)
posed model consists of convolution blocks and includes the
|w( p, q)|2 + λ improved kernel convolution
function, drop-out layer, and ( dense layers. Figures 3 and 4 show the overall enhanced
r
3
p3 + q
)
where W ( p, q) = e 2s2 , λ is constant with value 1, and W m is the minimized error for
stacked CNN and detailed convolution kernel functioning.
the Weiner filter.
Finally, to obtain the enhanced image for further processing, we integrate the results
of Equations (1) and (2), as follows:
I e ( p, q) = ( I t ( p, q) + W m ( p, q)) − I k ( p, q) (3)
 Technologies 2023, 11, 128 5 of 17

Figure 2. Image enhancement. (a) Unprocessed images and (b) processed images.

2.2.1. Enhanced
2.2.1. Enhanced CNN
CNN
We developed
We developed an an enhanced
enhanced CNN
CNN model
model toto detect
detect lung
lung illness
illness accurately.
accurately. The
The pro-
pro-
chnologies 2023, 11, x FOR PEER REVIEW posed model
posed model consists
consists of
of convolution
convolution blocks
blocks and
and includes
includes the
the improved
improved kernel
kernel convolution
6 of convolution
19
function, drop-out
function, drop-out layer,
layer, and
and dense
dense layers.
layers. Figures
Figures 33 and
and 44 show
show the
the overall
overall enhanced
enhanced
stacked CNN and detailed convolution kernel functioning.
stacked CNN and detailed convolution kernel functioning.
network’s generalization ability. Additionally, the last few layers of the proposed model
were adjusted to better match the learnt high-level properties.
The deep feature map was extracted in the proposed model by concatenating the fea-
ture map outputs from the final convolutional layer of the sacked CNN models along the
channels. To provide flexibility to the stacking model and reduce generalization error, we
used a fully connected neural network as a meta-learner. The final scores were then cal-
culated by concatenating the final feature maps and running them through the softmax
activation.
To obtain the channels of the maps to the same depth, 1 × 1 pointwise convolution
was first applied to each of these feature maps. These maps were then provided to five
stacks as input feature embedding. The final output was then obtained by flattening the
final feature map and running it through a dense layer with softmax activation. Stacking
works by determining the biases in relation to a given training set. A final classification
layer with Sigmoid activation and a few fully connected layers with rectified linear unit
(ReLU) activation were added to the model to improve its performance on the dataset. To
achieve a better outcome than the individual forecasts, we have combined the predictions
from the sub-models in this case. We combined the CNN models into a single one by con-
catenatingFigure
them, 3.and we then
Enhanced addedCNN
stacked twomodel.
dense layers with the softmax and ReLU activa-
Figure 3. Enhanced stacked CNN model.
tion functions.
All layers in the proposed model receive input from the prior ones and pass along
their feature mappings to all subsequent layers. Instead of adding the layer’s incoming
and outgoing feature maps, the proposed method concatenates them. As a result, the pro-
posed model is split up into dense blocks, where the filter numbers may change inside a
block, but the feature map dimensions never change. The transition layers, which are the
layers between the blocks, go through batch normalization in order to reduce the number
of channels by 50%. However, the proposed model requires fewer parameters and allows
feature reuse, which makes the networks smaller.
The proposed model encourages the reuse of features by creating dense relationships
between layers inside a dense block. A robust feature transmission between layers is made
possible by the dense connections, which also increase the gradient flow. The computa-
tional efficiency of the proposed model is also improved by the fact that it has many fewer
trainable parameters. The bias and skewness caused by the asymmetric class distribution
were reduced by proportionally adjusting the class weights. Additionally, early stopping
was used to address the overfitting problem while preserving the convolution neural

Figure 4. Convolution kernel functioning.

Figure 4. Convolution kernel functioning.

Enhanced Kernel
AllConvolution
layers in the proposed model receive input from the prior ones and pass along
In order to feature
their guarantee featureto
mappings independence
all subsequentduring
layers.stacking
Instead modelling, thelayer’s
of adding the deep fea-
incoming and
tures wereoutgoing
extractedfeature maps, the
from trained proposed
models. methodcan
The model concatenates them.
fit the training As more
data a result, the proposed
accu-
modelto
rately according is the
splitincreased
up into dense
modelblocks, wherewhich
complexity, the filter numbers
lowers may
the bias. change
The inside a block,
proposed
models learnbutextensive
the feature map dimensions
independent aspectsnever
aboutchange.
the dataThe transition layers,
by combining the deepwhich are the layers
feature
maps, which enables the models to generalize more effectively on the test database. The
proposed feature extractor’s depth-wise separable convolution requires fewer parame-
ters, which reduces overfitting. The proposed model’s dense block enables improved gra-
dient propagation and feature sharing. The proposed models have fewer parameters,
Technologies 2023, 11, 128 6 of 17

between the blocks, go through batch normalization in order to reduce the number of
channels by 50%. However, the proposed model requires fewer parameters and allows
feature reuse, which makes the networks smaller.
The proposed model encourages the reuse of features by creating dense relationships
between layers inside a dense block. A robust feature transmission between layers is
made possible by the dense connections, which also increase the gradient flow. The
computational efficiency of the proposed model is also improved by the fact that it has
many fewer trainable parameters. The bias and skewness caused by the asymmetric class
distribution were reduced by proportionally adjusting the class weights. Additionally, early
stopping was used to address the overfitting problem while preserving the convolution
neural network’s generalization ability. Additionally, the last few layers of the proposed
model were adjusted to better match the learnt high-level properties.
The deep feature map was extracted in the proposed model by concatenating the
feature map outputs from the final convolutional layer of the sacked CNN models along
the channels. To provide flexibility to the stacking model and reduce generalization
error, we used a fully connected neural network as a meta-learner. The final scores were
then calculated by concatenating the final feature maps and running them through the
softmax activation.
To obtain the channels of the maps to the same depth, 1 × 1 pointwise convolution
was first applied to each of these feature maps. These maps were then provided to five
stacks as input feature embedding. The final output was then obtained by flattening the
final feature map and running it through a dense layer with softmax activation. Stacking
works by determining the biases in relation to a given training set. A final classification
layer with Sigmoid activation and a few fully connected layers with rectified linear unit
(ReLU) activation were added to the model to improve its performance on the dataset. To
achieve a better outcome than the individual forecasts, we have combined the predictions
from the sub-models in this case. We combined the CNN models into a single one by
concatenating them, and we then added two dense layers with the softmax and ReLU
activation functions.

Enhanced Kernel Convolution

In order to guarantee feature independence during stacking modelling, the deep
features were extracted from trained models. The model can fit the training data more
accurately according to the increased model complexity, which lowers the bias. The
proposed models learn extensive independent aspects about the data by combining the
deep feature maps, which enables the models to generalize more effectively on the test
database. The proposed feature extractor’s depth-wise separable convolution requires
fewer parameters, which reduces overfitting. The proposed model’s dense block enables
improved gradient propagation and feature sharing. The proposed models have fewer
parameters, since they used meta-learners after the convolutional layers. Algorithm 1
presents the enhanced kernel function algorithm.
The latent dimensions of the kernel weights are defined using the kernel weights as
functions of the latent values. Each channel weight is represented by the proposed curves
weighted sum, as follows:
2 eu + e−u
w(b) = ŵ[F ∗ (b + A) + C], b ∈ [−π, π] (4)
π eu − e−u
where b is the latent variable and is continuous, so the latent space becomes continuous,
w(b) represents the kernel weights based on constant b, and F, A, and C are the matrices.
The convolution operation is an integral operation from −π to π. Equation (5) is as follows:

x(b) = ∑ d n ∈ D w ( Dn , b ) y ( Dn ) T (5)
The curve parameters can be obtained with the convolution operations. The ReLU
activation function is used to process the features, and the operation becomes max (0,x(b)).
Technologies 2023, 11, 128 7 of 17

Algorithm 1: Enhanced Kernel Function

1. X = {}
2. for all F, A, C ∈ {Fj , Aj , Cj }
3. obtain the kernel convolution weights
u
e−u
4. w(b) < −w[F π 2
∗ eeu +
−e−u (b + A) + C]
5. evaluate the convolution with the enhanced kernel weights
6. X(b) < − ∑ Dn ∈D w(Dn fb) ∗ (Dn )T
7. Discertize the dimensions of latent
8. For all b ∈ {b1, b2,. . .bn} do
9. X < −XU{}X(b)}
10. end for
11. end for
12. return list (X)

2.2.2. Classification
After extracting the features, the enhanced SVM classifies the lesions. Initially, the SVM
calculates the score for all the extracted features by using linear mapping on feature vectors
and uses this to evaluate the loss. The Algorithm 2 presents the improved SVM algorithm.

Algorithm 2: Improved SVM

1. Initialize the values in the training set.
2. Repeat for j = 1 to M
• Calculate the loss using the enhanced regularization for all values of j
• Compared the extracted regions in the liver images.
3. end for.
4. Repeat for every score vector j − 1 to M.
• Compute the weight and bias.
• argmax((w ∗ pj) + b)
5. end for.
6. Compute for all weights and finally evaluate the output.

The improved uses the linear mapping on extracted features to calculate the feature
score to the parts of the region of interest used to differentiate the lesion types, which helps
in the evaluation of loss function, which helps in obtaining the classification results.

3. Experimental Results
All experiments have been implemented on Keras. The data split was performed
based on an 80:20 ratio, where 80% of the data were used for training and 20% for testing
and implemented the proposed kernel convolution function in the CNN for the chest X-ray
images and CT scan images. Three classes of chest X-ray or CT scan images were used,
including healthy, COVID-19 affected, and patients with pneumonia.

3.1. Evaluation of Chest X-ray Dataset

Figure 5 shows the confusion matrix for the chest X-ray dataset. The proposed model
is compared with the existing models DeepCNN [27], Unet [9], nCONVnet [28], Dark-
net [29], infet [8], and Unet++ [30]. The proposed model accurately classifies 316 out of
317 and misclassifies 1 for the Normal class, accurately classifies 853 cases out of 855 and
misclassifies 2 for Pneumonia class, and accurately classifies 15 out of 16 and misclassifies
1 for the COVID-19 class. The DeepCNN model accurately classifies 305 for the Normal
class, accurately classifies 845 cases for Pneumonia class, and accurately classifies 10 for the
COVID-19 class. The Unet model accurately classifies 310 for the Normal class, accurately
classifies 848 cases for Pneumonia class, and accurately classifies 12 for the COVID-19
class. The nCONVnet model accurately classifies 315 for the Normal class, accurately
 [29], infet [8], and Unet++ [30]. The proposed model accurately classifies 316 out of 317
and misclassifies 1 for the Normal class, accurately classifies 853 cases out of 855 and mis-
classifies 2 for Pneumonia class, and accurately classifies 15 out of 16 and misclassifies 1
for the COVID-19 class. The DeepCNN model accurately classifies 305 for the Normal
class, accurately classifies 845 cases for Pneumonia class, and accurately classifies 10 for
Technologies 2023, 11, 128 the COVID-19 class. The Unet model accurately classifies 310 for the Normal class, accu- 8 of 17
rately classifies 848 cases for Pneumonia class, and accurately classifies 12 for the COVID-
19 class. The nCONVnet model accurately classifies 315 for the Normal class, accurately
classifies 850 cases for the Pneumonia class, and accurately classifies 15 for the COVID-19
Darknetmodel
class. The Darknet modelaccurately
accuratelyclassifies
classifies315
315for
forthe
the Normal
Normal class,
class, accurately
accurately classi-
classifies
fies cases
853 853 cases for Pneumonia
for the the Pneumonia class,class,
and and accurately
accurately classifies
classifies 14the
14 for forCOVID-19
the COVID-19class.class.
The
The infet
infet model model accurately
accurately classifies
classifies 310 for310the
forNormal
the Normalclass, class, accurately
accurately classifies
classifies 848
848 cases
cases
for thefor the Pneumonia
Pneumonia class,
class, and and accurately
accurately classifies
classifies 12 for the12 COVID-19
for the COVID-19 class.
class. The The
Unet++
Unet++accurately
model model accurately classifies
classifies 315 for315thefor the Normal
Normal class,class, accurately
accurately classifies
classifies 852 852
casescases
for
the Pneumonia
for the Pneumonia class, andand
class, accurately classifies
accurately 14 for
classifies COVID-19
14 for COVID-19 class. From
class. Figure
From 5, we
Figure 5,
observed
we observed thatthat
the proposed
the proposedmodel mostmost
model accurately classifies
accurately the chest
classifies X-rayX-ray
the chest class followed
class fol-
by the Unet++
lowed and Darknet
by the Unet++ models.
and Darknet models.

Figure 5.
Figure 5. Confusion
Confusion matrix
matrix comparison
comparison for
for the proposed model
the proposed with the
model with the existing
existing models
models for
for the
the
chest X-ray dataset.
chest X-ray dataset.

The overall performance of the chest X-ray dataset is shown in Figure 6. Each violin
Technologies 2023, 11, x FOR PEER REVIEW 9 of 19
plot’s width corresponds to the frequency of the mean intensity. In general, the probability
plot’s width corresponds to the frequency of the mean intensity. In general, the probability
that the images have the corresponding mean intensity increases with the width of the
violin plot.
violin plot.The
Theintensity
intensitydistribution
distribution still
still exhibits
exhibits variance.
variance. TheThe accuracy
accuracy of multi-class
of the the multi-
lung
class infection classification
lung infection from the
classification fromX-ray images
the X-ray is99.8%.
images TheseThese
is99.8%. findings demonstrate
findings demon-
how
strateeffectively the models
how effectively identify
the models lesionslesions
identify in chestinX-ray
chestimages.
X-ray images.

Figure 6. Performance
Figure6. Performance of
of the
the complete
complete chest
chest ray
ray dataset.
dataset.

The individual classifier performance for the chest X-ray dataset is shown in Figure
7. The ISVM classifier shows the outstanding performance compared to the base classifiers
of k-nearest neighbor (KNN), naïve Bayes (NB), random forest (RF), and the standard
support vector machine (SVM).
Technologies 2023, 11, 128 9 of 17

Figure 6. Performance of the complete chest ray dataset.

The
The individual
individualclassifier
classifierperformance
performancefor
forthe
thechest
chestX-ray
X-raydataset
datasetis is
shown
showninin
Figure
Figure7.
The ISVM
7. The ISVMclassifier shows
classifier thethe
shows outstanding performance
outstanding performancecompared
compared to the base
to the classifiers
base of
classifiers
k-nearest neighbor
of k-nearest (KNN),
neighbor naïvenaïve
(KNN), BayesBayes
(NB), (NB),
random forest (RF),
random forestand the and
(RF), standard support
the standard
vector machine (SVM).
support vector machine (SVM).

(a) (b)

(c) (d)
Figure 7. Comparison of classifier performance on the chest X-ray dataset. (a) Accuracy; (b) preci-
Figure 7. Comparison of classifier performance on the chest X-ray dataset. (a) Accuracy; (b) precision;
sion; (c) recall; (d) F1-score.
(c) recall; (d) F1-score.

The chest X-ray dataset consists of three classes, i.e., Normal (healthy), COVID-19
affected, and patients with pneumonia.
The individual class categorization performance for the chest X-ray dataset is shown
in Figure 8.

3.2. Evaluation of CT Scan Dataset

Figure 9 shows the confusion matrix for the CT scan dataset. The proposed model is
compared with the existing models DeepCNN [27], Unet [9], nCONVnet [28], Darknet [29],
infet [8], and Unet++ [30]. The proposed model accurately classifies 67 out of 67 for the
Normal class, accurately classifies 71 cases out of 72 and misclassifies 1 for the Pneumonia
class, and accurately classifies 69 out of 70 and misclassifies 1 for the COVID-19 class. The
DeepCNN model accurately classifies 60 for the Normal class, accurately classifies 63 cases
for the Pneumonia class, and accurately classifies 61 for the COVID-19 class. The Unet
model accurately classifies 64 for the Normal class, accurately classifies 68 cases for the
Pneumonia class, and accurately classifies 67 for the COVID-19 class. The nCONVnet
model accurately classifies 64 for the Normal class, accurately classifies 70 cases for the
Pneumonia class, and accurately classifies 65 for the COVID-19 class. The Darknet model
Technologies 2023, 11, 128 10 of 17

accurately classifies 65 for the Normal class, accurately classifies 68 cases for the Pneumonia
Technologies 2023, 11, x FOR PEER REVIEW 10 of 19
class, and accurately classifies 67 for the COVID-19 class. The infet model accurately
classifies 64 for the Normal class, accurately classifies 68 cases for the Pneumonia class, and
accurately classifies 67 for the COVID-19 class. The Unet++ model accurately classifies 65
for the The chest X-ray dataset consists of three classes, i.e., Normal (healthy), COVID-19
Normal class, accurately classifies 70 cases for the Pneumonia class, and accurately
affected, and patients with pneumonia.
classifies 68 for the COVID-19 class. From Figure 9, we observed that the proposed model
The individual class categorization performance for the chest X-ray dataset is shown
most accurately classifies the chest X-ray class, followed by the Unet++ model.
in Figure 8.

(a) (b)

(c) (d)
Technologies 2023, 11, x FOR PEER REVIEW 11 of 19
Figure 8. Class-wise performance on the chest X-ray dataset. (a) Accuracy; (b) precision; (c) recall;
Figure 8. Class-wise performance on the chest X-ray dataset. (a) Accuracy; (b) precision; (c) recall;
(d) F1-score.
(d) F1-score.
3.2. Evaluation of CT Scan Dataset
Figure 9 shows the confusion matrix for the CT scan dataset. The proposed model is
compared with the existing models DeepCNN [27], Unet [9], nCONVnet [28], Darknet
[29], infet [8], and Unet++ [30]. The proposed model accurately classifies 67 out of 67 for
the Normal class, accurately classifies 71 cases out of 72 and misclassifies 1 for the Pneu-
monia class, and accurately classifies 69 out of 70 and misclassifies 1 for the COVID-19
class. The DeepCNN model accurately classifies 60 for the Normal class, accurately clas-
sifies 63 cases for the Pneumonia class, and accurately classifies 61 for the COVID-19 class.
The Unet model accurately classifies 64 for the Normal class, accurately classifies 68 cases
for the Pneumonia class, and accurately classifies 67 for the COVID-19 class. The nCON-
Vnet model accurately classifies 64 for the Normal class, accurately classifies 70 cases for
the Pneumonia class, and accurately classifies 65 for the COVID-19 class. The Darknet
model accurately classifies 65 for the Normal class, accurately classifies 68 cases for the
Pneumonia class, and accurately classifies 67 for the COVID-19 class. The infet model ac-
curately classifies 64 for the Normal class, accurately classifies 68 cases for the Pneumonia
class, and accurately classifies 67 for the COVID-19 class. The Unet++ model accurately
classifies 65 for the Normal class, accurately classifies 70 cases for the Pneumonia class,
and accurately classifies 68 for the COVID-19 class. From Figure 9, we observed that the
proposed model most accurately classifies the chest X-ray class, followed by the Unet++
Figure
Figure 9.
9.Confusion
Confusionmatrix
matrixcomparison
comparisonfor
forthe
theproposed
proposedmodel
modelwith
withthe
theexisting
existingmodels
modelsfor
forthe
theCT
CT
model.
scan dataset.
scan dataset.

The overall performance of the CT scan dataset is shown in Figure 10. Each violin
plot’s width corresponds to the frequency of the mean intensity. In general, the probability
that the images have the corresponding mean intensity increases with the width of the
violin plot. The intensity distribution still exhibits variance. The accuracy of the multi-
class lung infection classification from the CT scan images is99.9%. These findings demon-
Technologies 2023, 11, 128 11 of 17

Figure 9. Confusion matrix comparison for the proposed model with the existing models for the CT
scan dataset.
The overall performance of the CT scan dataset is shown in Figure 10. Each violin
plot’s width correspondsThe to overall
the frequency
performanceof the mean
of the intensity.
CT scan Inshown
dataset is general, the probability
in Figure 10. Each violin
that the images have thewidth
plot’s corresponding
corresponds tomean intensity
the frequency of theincreases withIn the
mean intensity. width
general, of the
the probability
violin plot. The intensity distribution
that the images havestill exhibits variance.
the corresponding The accuracy
mean intensity increasesofwith
the the
multi-class
width of the
violin plot.
lung infection classification The the
from intensity distribution
CT scan imagesstill exhibits These
is99.9%. variance. The accuracy
findings of the multi-
demonstrate
class lung infection classification from the CT scan images is99.9%. These findings demon-
how effectively the models identify lesions in CT scan images.
strate how effectively the models identify lesions in CT scan images.

Figure
Figure 10. Performance 10.complete
of the Performance
CTof the
scancomplete CT scan dataset.
dataset.
The individual classifier performance for the chest X-ray dataset is shown in Figure
The individual classifier performance for the chest X-ray dataset is shown in Figure 11.
11. The ISVM classifier shows an outstanding performance when compared to the the base
The ISVM classifier classifiers,
shows an outstanding
namely performance
k-nearest neighbor (KNN), when compared
naïve Bayes to the
(NB), random the(RF),
forest baseand
classifiers,
Technologies 2023, 11, x FOR PEER REVIEW namely k-nearest
the standardneighbor (KNN),
support vector naïve
machine Bayes (NB), random forest
(SVM). 12 of 19(RF), and
the standard support vector machine (SVM).

(a) (b)

(c) (d)
Figure 11. Comparison of classifier performance on the CT scan dataset. (a) Accuracy; (b) precision;
Figure(c)11. Comparison
recall; (d) F1-score.of classifier performance on the CT scan dataset. (a) Accuracy; (b) precision;
(c) recall; (d) F1-score.
The CT scan consists of three classes, i.e., Normal (healthy), COVID-19 affected, and
patients with pneumonia. The individual class categorization performance for the CT scan
dataset is shown in Figure 12.
 (c) (d)
Technologies 2023, 11, 128 12 of 17
Figure 11. Comparison of classifier performance on the CT scan dataset. (a) Accuracy; (b) precision;
(c) recall; (d) F1-score.

TheCT
The CTscan
scanconsists
consistsof
ofthree
threeclasses,
classes,i.e.,
i.e.,Normal
Normal(healthy),
(healthy),COVID-19
COVID-19affected,
affected,and
and
patients
patientswith
withpneumonia.
pneumonia.The
Theindividual
individualclass
classcategorization
categorizationperformance
performancefor
forthe
theCT
CTscan
scan
dataset
datasetisisshown
shownin
inFigure
Figure12.
12.

Technologies 2023, 11, x FOR PEER REVIEW 13 of 19

(a) (b)

(c) (d)
Figure
Figure 12.
12. Class-wise
Class-wiseperformance
performanceon
onthe
theCT
CTscan dataset.
scan (a) (a)
dataset. Accuracy; (b) (b)
Accuracy; precision; (c) recall;
precision; (d)
(c) recall;
F1-score.
(d) F1-score.

3.3. Performance
Performance Comparison
Comparison
The model
model has beenbeen trained
trained for
for 900
900 epochs
epochs on on training
training data,
data, and
and the
the proposed
proposed models
models
are trained for 900 epochs with batch sizes of 64. To To determine how well the proposed
model
model would
would work
work andand how
how often
often itit would
would be be used,
used, every
every combination
combination has has been
been tested.
tested.
This has been accomplished
accomplished by combining
combining variousvarious models
models results
results in
in stacking
stacking systems,
systems,
which produces reliable forecasts while reducing reducing the the possibility
possibility ofof incorrect
incorrect predictions.
predictions.
The results
The resultsofofthe
thecomparison,
comparison, which
which was was performed
performed in terms
in terms of accuracy,
of accuracy, recall,recall,
pre-
precision,
cision, AUC,
AUC, andand
F1 F1 score,
score, areare displayed
displayed inin Figure13.
Figure 13.The
Thetable
tableshows
showsthatthatthe
thesuggested
suggested
model exceeds
exceeds all
allother
othermodels
modelstakentaken into
intoconsideration
consideration in terms of accuracy.
in terms The research
of accuracy. The re-
of [27] and [28] likewise used CT scan and X-ray images, and achieved
search of [27] and [28] likewise used CT scan and X-ray images, and achieved accuracy accuracy rates of
98.23%
rates of and 99.87%,
98.23% and respectively; however,however,
99.87%, respectively; they used binary
they usedcategorization. The Darknet
binary categorization. The
model was
Darknet utilized
model in the study
was utilized in theofstudy
[29], which
of [29],solely
whichusedsolelyX-ray
usedimages, and 96.42%
X-ray images, and
accuracy
96.42% was attained.
accuracy In contrast
was attained. to ourtoproposed
In contrast work,work,
our proposed which combined
which CT scan
combined and
CT scan
X-ray
and images,
X-ray the author
images, of [8]ofreported
the author [8] reportedan accuracy
an accuracyof 94.47% for the
of 94.47% forinfet model.
the infet Our
model.
proposed
Our workwork
proposed employed 5935 5935
employed CT scanCT images and 1043
scan images and X-ray images
1043 X-ray and reached
images 99.9%
and reached
99.9% and 99.8% accuracy, respectively, whereas the author of [30] used both CT scanX-ray
and 99.8% accuracy, respectively, whereas the author of [30] used both CT scan and and
imagesimages
X-ray and achieved
and achieved99.57%99.57%
and 99.47% accuracy,
and 99.47% respectively.
accuracy, Our proposed
respectively. model,
Our proposed
model, however, has the highest accuracy and precision values for the three-class catego-
rization of chest CT images and X-ray images.
Figure 14 shows the computational complexity in terms of the number of parameters
and the time. From Figure 14, it is observed that the proposed model exhibits a lower
number of parameters and takes less time to process. Depth-wise separable convolution
Technologies 2023, 11, 128 13 of 17

Technologies 2023, 11, x FOR PEER REVIEW 14 of 19

however, has the highest accuracy and precision values for the three-class categorization of
chest CT images and X-ray images.

Technologies 2023, 11, x FOR PEER REVIEW 14 of 19

Figure
Figure 13.
13. Performance
Performance comparison.
comparison.

Figure 14 shows the computational complexity in terms of the number of parameters

and the time. From Figure 14, it is observed that the proposed model exhibits a lower
number of parameters and takes less time to process. Depth-wise separable convolution
helps in
Figure 13.reducing the comparison.
Performance parameters and improves the performance of the model.

Figure 14. Complexity comparison.

Figure
Figure 14. Complexity
Complexity comparison.
comparison.

Figure 15 shows the comparison results for models with and without image processing.
We compared the performance metrics, i.e., the accuracy, precision, recall, and F1-score for
the existing models, to the proposed model with and without image processing. For the
proposed model, the accuracy without image processing is 98.46%, and with image process-
ing it is 99.90%. The proposed model achieved the highest accuracy, followed by Unet++
with an accuracy of 99.57% with image processing and 98.26% without image processing.

(a) (b)
Technologies 2023, 11, 128 14 of 17

Figure 14. Complexity comparison.

Technologies 2023, 11, x FOR PEER REVIEW 15 of 19

(a) (b)

(c) (d)
Figure 15. Pre-processing comparison. (a) Accuracy; (b) precision; (c) recall; (d) F1-score.
Figure 15. Pre-processing comparison. (a) Accuracy; (b) precision; (c) recall; (d) F1-score.
Statistical test
Statistical test To prove that the proposed model performs better than the existing approaches, we
To prove that have
the proposed
performed the model performs
statistical better
hypothesis usingthan the existing
the paired approaches,
sample t-test we
[31]. The t-test
have performed the finds the significant
statistical difference using
hypothesis betweenthe
the paired
sets of observations.
sample t-testIt only allows
[31]. Thefor t-test
testing
finds the significant thedifference
pairs of observations
betweenand a conclusion
the can be drawn. Therefore,
sets of observations. we havefor
It only allows compared
testing all
the independent DL models and the proposed model, selecting two at a time with the final
the pairs of observations and a conclusion can be drawn. Therefore, we have compared
stacking ensemble framework. The null hypothesis is assumed, as there exists no signifi-
all the independentcant DLdifference
modelsbetweenand the theproposed
performance model, selecting
of the individual two at
models andathe
time withframe-
stacked the
final stacking ensemble
work. On framework.
the other hand, The null hypothesis
an alternative hypothesisisstates
assumed,
there is aas there exists
statistically no
significant
significant differencedifference
between between
the the performancesof
performance of the
thetwo compared models.
individual modelsAll the the
and statistical
stackedtrials
are performed at a significance level of 5%, or the alpha is 0.05. The obtained p-values are
framework. On the other hand, an alternative hypothesis states there is a statistically
shown in Table 2. From the resulting p-values, which are all less than 0.05, we reject our
significant difference between
assumed the performances
null hypothesis. of the
This proves that thebetter
twoperformance
comparedofmodels.
the proposed Allmodel
the
statistical trials areis statistically
performed at a significance level of 5%,
significant and not a result obtained by chance.or the alpha is 0.05. The
obtained p-values are shown in Table 2. From the resulting p-values, which are all less than
Table 2. Statistical test results.
0.05, we reject our assumed null hypothesis. This proves that the better performance of the
proposed model isModel statistically significant and not a result p-Value
obtained by chance.
DeepCNN 0.01201
Unet 0.01818
Table 2. Statistical test results.
NCONVnet 0.01217
Model Darknet p-Value 0.00402
Infet 0.00395
DeepCNN Unet++ 0.01201 0.00245
Unet Proposed (chest X-ray) 0.01818 0.00135
Proposed (CT scan) 0.00129
NCONVnet 0.01217
Darknet 4. Discussion 0.00402
Infet Suppose that all of the features0.00395
are connected to the final fully connected layer, which
results in an increase in the complexity of the training set. To overcome this issue, we
Unet++ 0.00245
proposed kernel convolution functions used while training the model. The proposed ker-
nel convolution function reduces the computational overhead and enhances the network
Proposed (chest X-ray) 0.00135
compactness by regulating the number of feature maps. This study aimed to identify and
Proposed (CT scan)classify lung infections based on chest
0.00129
X-ray and CT scan images from different datasets.
The proposed kernel convolution functions outperform the existing predefined functions
with an accuracy of 99.8%, a sensitivity of 99.9%, and a specificity of 99% on chest X-ray
Technologies 2023, 11, 128 15 of 17

4. Discussion
Suppose that all of the features are connected to the final fully connected layer, which
results in an increase in the complexity of the training set. To overcome this issue, we
proposed kernel convolution functions used while training the model. The proposed kernel
convolution function reduces the computational overhead and enhances the network
compactness by regulating the number of feature maps. This study aimed to identify and
classify lung infections based on chest X-ray and CT scan images from different datasets.
The proposed kernel convolution functions outperform the existing predefined functions
with an accuracy of 99.8%, a sensitivity of 99.9%, and a specificity of 99% on chest X-ray
images and an accuracy of 99.9%, a sensitivity of 99.9%, and a specificity of 98% on CT
scan images.
Initially, all the images in the dataset were of different sizes; the images were resized to
225 × 225 using the RGB color. The hyperparameters were tuned to optimize the proposed
model. Model training can be accelerated, and the possibility of performance improved
using the kernel convolution function. There is no ideal batch size; we implemented the
experiments with various batch sizes. If we find the suitable batch size in addition to the
suitable activation function and hidden layers, the model will yield a high performance.
A batch size of 64 gives better results than batch sizes of 16 or 32. The batch size was set
to 64 for the CT scan images because this study’s CT scan image dataset was large. From
the previous studies, we observed that the batch sizes, in conjunction with the suitable
activation function and hidden layer, will yield a high performance. The parameters
adjusted to obtain the high-performance results are batch size (64), epochs (900), and
learning rate (0.001).
Hammoudi et al. [32] extracted pneumonia features using deep learning models,
namely Denset121, VGG16, VGG19, and Resnet50 for chest X-ray images, and classified
them using the SVMRBF to achieve an accuracy of 97.88%, 97.88%, 97.88%, and 97.88%,
respectively. Ibrahim et al. [33] implemented the Alexnet architecture via transfer learning
and achieved a classification accuracy of 98.19%, a precision of 97.29%, a recall of 97.29%,
and an F1-score of 97.29%.
One of the key limitations of the proposed approach is the small dataset, which is one
of the study’s shortcomings. However, in the future, the proposed performance should be
validated over a larger and more diversified dataset with more samples. We did not find
an increased risk of pneumonia infections, such as bronchopneumonia, in patients with
additional comorbidities. Future advancements of the suggested approach could include
employing 3D CT scans to train 3D CNN ensemble models.

5. Conclusions
This study proposed an enhanced feature extraction technique to minimize the ex-
traction error and an improved SVM classifier to classify the lesions using linear mapping.
The proposed model results were proven to outperform all the state-of-the-art models.
This study provided efficient accuracy results compared to the state-of-the-art models.
The manual identification and differentiation of COVID-19, pneumonia, and normal chest
X-rays and CT scans are time-consuming. The diagnostic results proved that the proposed
model achieved satisfactory diagnostic performance, which significantly assists the medical
professional in the decision-making process in the early stage of detection of the infection
itself, and timely treatment can alleviate the transmission of the infection and decrease
the risk of death. Automatic screening and differentiation of COVID-19, pneumonia, and
normal chest X-ray and CT scan images will significantly reduce the effort of the medical
professional and accelerate the diagnosis process.

Author Contributions: Conceptualization, U.B., N.C.; data curation, U.B.; funding acquisition, U.B.,
and G.B.; investigation, U.B., and N.C.; methodology, U.B., G.B. and N.C.; resources, U.B. and G.B.
All authors have read and agreed to the published version of the manuscript.
Funding: This study didn’t receive any funding.
Technologies 2023, 11, 128 16 of 17

Institutional Review Board Statement: Not Applicable.

Informed Consent Statement: Not Applicable.
Data Availability Statement: Not Applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Sanchez, C.A.; Rivera-Lozada, O.; Lozada-Urbano, M.; Best-Bandenay, P. Herd immunity in older adults from a mid-dle-income
country: A time-series trend analysis of community-acquired pneumonia mortality 2003–2017. Health Sci. Rep. 2023, 6, e1224.
[CrossRef] [PubMed]
2. Shen, Y.; Tian, Z.; Lu, D.; Huang, J.; Zhang, Z.; Li, X.; Li, J. Impact of pneumonia and lung cancer on mortality of women with
hypertension. Sci. Rep. 2016, 6, 20. [CrossRef] [PubMed]
3. Moses, D.A. Deep learning applied to automatic disease detection using chest X-rays. J. Med. Imaging Radiat. Oncol. 2021, 65,
498–517. [CrossRef]
4. Ahuja, S.; Panigrahi, B.K.; Dey, N.; Rajinikanth, V.; Gandhi, T.K. Deep transfer learning-based automated detection of COVID-19
from lung CT scan slices. Appl. Intell. 2021, 51, 571–585. [CrossRef]
5. Bharati, S.; Podder, P.; Mondal, M.R.H.; Podder, P.; Kose, U. A review on epidemiology, genomic characteristics, spread, and
treatments of COVID-19. Data Sci. COVID-19 2022, 2, 487–505.
6. 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. Performance
of radiologists in differentiating COVID-19 from non-COVID-19 viral pneumonia at chest CT. Radiology 2020, 296, E46–E54.
[CrossRef]
7. Rajaraman, S.; Candemir, S.; Thoma, G.; Antani, S. Visualizing and explaining deep learning predictions for pneumonia de-tection
in pediatric chest radiographs. In Proceedings of the Medical Imaging 2019: Computer-Aided Diagnosis, San Diego, CA, USA,
17–20 February 2019; Volume 10950, pp. 200–211. [CrossRef]
8. Fan, D.-P.; Zhou, T.; Ji, G.-P.; Zhou, Y.; Chen, G.; Fu, H.; Shen, J.; Shao, L. Inf-net: Automatic COVID-19 lung infection
seg-mentation from ct images. IEEE Trans. Med. Imaging 2020, 39, 2626–2637. [CrossRef]
9. Müller, D.; Soto-Rey, I.; Kramer, F. Robust chest CT image segmentation of COVID-19 lung infection based on limited data. Inform.
Med. Unlocked 2021, 25, 100681. [CrossRef]
10. Ranjbarzadeh, R.; Jafarzadeh Ghoushchi, S.; Bendechache, M.; Amirabadi, A.; Ab Rahman, M.N.; Baseri Saadi, S.; Aghamo-
hammadi, A.; Kooshki Forooshani, M. Lung infection segmentation for COVID-19 pneumonia based on a cascade convolu-tional
network from CT images. BioMed Res. Int. 2021, 2021, 5544742. [CrossRef]
11. Li, L.; Qin, L.; Xu, Z.; Yin, Y.; Wang, X.; Kong, B.; Bai, J.; Lu, Y.; Fang, Z.; Song, Q.; et al. Using artificial intelligence to detect
COVID-19 and community-acquired pneumonia based on pulmonary CT: Evaluation of the diagnostic accuracy. Radiology 2020,
296, E65–E71. [CrossRef]
12. Gu, X.; Pan, L.; Liang, H.; Yang, R. Classification of bacterial and viral childhood pneumonia using deep learning in chest
radiography. In Proceedings of the 3rd International Conference on Multimedia and Image Processing, Guiyang, China, 16–18
March 2018; pp. 88–93. [CrossRef]
13. Mazurowski, M.A.; Buda, M.; Saha, A.; Bashir, M.R. Deep learning in radiology: An overview of the concepts and a survey of the
state of the art with focus on MRI. J. Magn. Reson. Imaging 2019, 49, 939–954. [CrossRef]
14. Milletari, F.; Ahmadi, S.A.; Kroll, C.; Plate, A.; Rozanski, V.; Maiostre, J.; Levin, J.; Dietrich, O.; Ertl-Wagner, B.; Bötzel, K.; et al.
Hough-CNN: Deep learning for segmentation of deep brain regions in MRI and ultrasound. Comput. Vis. Image Underst. 2017,
164, 92–102. [CrossRef]
15. Lee, S.; Lee, Y.H. Improving emergency department efficiency by patient scheduling using deep reinforcement learning. Healthcare
2020, 8, 77. [CrossRef] [PubMed]
16. Zhang, J.; Xie, Y.; Li, Y.; Shen, C.; Xia, Y. COVID-19 screening on chest X-ray images using deep learning based anomaly detection.
arXiv 2020, arXiv:2003.12338.
17. Apostolopoulos, I.D.; Mpesiana, T.A. COVID-19: Automatic detection from x-ray images utilizing transfer learning with
convolutional neural networks. Phys. Eng. Sci. Med. 2020, 43, 635–640. [CrossRef]
18. Xia, X.; Xu, C.; Nan, B. Inception-v3 for flower classification. In Proceedings of the 2017 2nd International Conference on Image,
Vision and Computing (ICIVC), Chengdu, China, 2–4 June 2017; IEEE: New York, NY, USA, 2017; pp. 783–787.
19. Wang, C.; Chen, D.; Hao, L.; Liu, X.; Zeng, Y.; Chen, J.; Zhang, G. Pulmonary image classification based on inception-v3 transfer
learning model. IEEE Access 2019, 7, 146533–146541. [CrossRef]
20. Tsiknakis, N.; Trivizakis, E.; Vassalou, E.E.; Papadakis, G.Z.; Spandidos, D.A.; Tsatsakis, A.; Sánchez García, J.; López González,
R.; Papanikolaou, N.; Karantanas, A.H.; et al. Interpretable artificial intelligence framework for COVID 19 screening on chest X
rays. Exp. Ther. Med. 2020, 20, 727–735. [CrossRef]
21. Sethy, P.K.; Behera, S.K.; Rath a, P.K.; Biswas, P. Detection of coronavirus disease (COVID-19) based on deep features and support
vector machine. Int. J. Math. Eng. Manag. Sci. 2020, 5, 643–651. [CrossRef]
Technologies 2023, 11, 128 17 of 17

22. Saha, P.; Sadi, M.S.; Islam, M.M. EMCNet: Automated COVID-19 diagnosis from X-ray images using convolutional neural
network and ensemble of machine learning classifiers. Inform. Med. Unlocked 2021, 22, 100505. [CrossRef]
23. Horry, M.J.; Chakraborty, S.; Paul, M.; Ulhaq, A.; Pradhan, B.; Saha, M.; Shukla, N. X-ray image based COVID-19 detection using
pre-trained deep learning models. eng rXiv 2020. [CrossRef]
24. SARS- CoV-2 Ct-Scan Dataset. Available online: https://www.kaggle.com/datasets/plameneduardo/sarscov2-ctscan-dataset
(accessed on 20 May 2023).
25. COVID-CT. Available online: https://github.com/UCSD-AI4H/COVID-CT (accessed on 20 May 2023).
26. Rahimzadeh, M.; Attar, A.; Sakhaei, S.M. A fully automated deep learning-based network for detecting COVID-19 from a new
and large lung CT scan dataset. Biomed. Signal Process. Control 2021, 68, 102588. [CrossRef] [PubMed]
27. Abbas, A.; Abdelsamea, M.M.; Gaber, M.M. Classification of COVID-19 in chest X-ray images using DeTraC deep convolu-tional
neural network. Appl. Intell. 2021, 51, 854–864. [CrossRef] [PubMed]
28. Panwar, H.; Gupta, P.K.; Siddiqui, M.K.; Morales-Menendez, R.; Singh, V. Application of deep learning for fast detection of
COVID-19 in X-rays using nCOVnet. Chaos Solitons Fractals 2020, 138, 109944. [CrossRef]
29. Bhuiyan, M.R.; Khushbu, S.A.; Islam, M.S. A deep learning based assistive system to classify COVID-19 face mask for human
safety with YOLOv3. In Proceedings of the 2020 11th International Conference on Computing, Communication and Networking
Technologies (ICCCNT), Kānpur, India, 1–3 July 2020; pp. 1–5. [CrossRef]
30. Shan, F.; Gao, Y.; Wang, J.; Shi, W.; Shi, N.; Han, M.; Xue, Z.; Shen, D.; Shi, Y. Lung infection quantification of COVID-19 in CT
images with deep learning. arXiv 2020, arXiv:2003.04655.
31. Ross, A.; Willson, V.L. Paired samples T-test. In Basic and Advanced Statistical Tests; Brill: Leiden, The Netherlands, 2017; pp. 17–19.
32. Hammoudi, K.; Benhabiles, H.; Melkemi, M.; Dornaika, F.; Arganda-Carreras, I.; Collard, D.; Scherpereel, A. Deep learning
on chest X-ray images to detect and evaluate pneumonia cases at the era of COVID-19. J. Med. Syst. 2021, 45, 75. [CrossRef]
[PubMed]
33. Ibrahim, A.U.; Ozsoz, M.; Serte, S.; Al-Turjman, F.; Yakoi, P.S. Pneumonia classification using deep learning from chest X-ray
images during COVID-19. Cogn. Comput. 2021, 1, 1–13. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.