Alexandria Engineering Journal (2023) 64, 923–935

H O S T E D BY
Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

ORIGINAL ARTICLE

A deep learning architecture for multi-class lung

diseases classification using chest X-ray (CXR)
images
Goram Mufarah M. Alshmrani a,*, Qiang Ni a, Richard Jiang a, Haris Pervaiz a,
Nada M. Elshennawy b

a
School of Computing and Commutations, Lancaster University, Lancaster LA1 4YW, UK
b
Department of Computers and Control Engineering, Faculty of Engineering, Tanta University, Tanta 31733, Egypt

Received 31 August 2022; revised 10 October 2022; accepted 21 October 2022

Available online 02 November 2022

KEYWORDS Abstract In 2019, the world experienced the rapid outbreak of the Covid-19 pandemic creating an
Pneumonia; alarming situation worldwide. The virus targets the respiratory system causing pneumonia with
Lung cancer; other symptoms such as fatigue, dry cough, and fever which can be mistakenly diagnosed as pneu-
COVID-19, TB, Lung opac- monia, lung cancer, or TB. Thus, the early diagnosis of COVID-19 is critical since the disease can
ity; provoke patients’ mortality. Chest X-ray (CXR) is commonly employed in healthcare sector where
X-ray images; both quick and precise diagnosis can be supplied. Deep learning algorithms have proved extraordi-
Deep learning, VGG19 nary capabilities in terms of lung diseases detection and classification. They facilitate and expedite
+CNN; the diagnosis process and save time for the medical practitioners. In this paper, a deep learning
Multiclass diseases (DL) architecture for multi-class classification of Pneumonia, Lung Cancer, tuberculosis (TB),
classification
Lung Opacity, and most recently COVID-19 is proposed. Tremendous CXR images of 3615
COVID-19, 6012 Lung opacity, 5870 Pneumonia, 20,000 lung cancer, 1400 tuberculosis, and
10,192 normal images were resized, normalized, and randomly split to fit the DL requirements.
In terms of classification, we utilized a pre-trained model, VGG19 followed by three blocks of con-
volutional neural network (CNN) as a feature extraction and fully connected network at the clas-
sification stage. The experimental results revealed that our proposed VGG19 + CNN
outperformed other existing work with 96.48 % accuracy, 93.75 % recall, 97.56 % precision,
95.62 % F1 score, and 99.82 % area under the curve (AUC). The proposed model delivered supe-
rior performance allowing healthcare practitioners to diagnose and treat patients more quickly and
efficiently.
Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria
University. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/
4.0/).

* Corresponding author.
E-mail address: [email protected] (G.M.M. Alshmrani).
Peer review under responsibility of Faculty of Engineering, Alexandria University.
https://doi.org/10.1016/j.aej.2022.10.053
1110-0168 Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
924 G.M.M. Alshmrani et al.

1. Introduction In late 2019, the coronavirus (COVID-19) pandemic

invaded the planet, leading to an alarming scenario. The virus
Numerous studies have discussed the efficacy of computer- was first formally discovered in Wuhan, China in December
aided diagnoses in the medical context, based on collaboration 2019, and the World Health Organization (WHO) designated
between medical researchers and computer scientists. Certain it as an emergency health problem at the beginning of 2020.
computer-aided diagnosis systems in medicine may be classi- By March 2020, WHO classified it as a pandemic [4]. The
fied as expert systems since they seek to replicate the decisions Coronavirus causes pneumonia, persistent cough, high fever,
of medical professionals. In addition, computer-aided detec- and fatigue, among other symptoms. Reverse transcription-
tion systems in medicine can process complicated and large polymerase chain reaction (RT-PCR) is employed to identify
clinical data [1]. Computer-aided detection systems can also positive cases of the virus. However, it can take several hours,
assist clinicians to gain new insights into data and apply the even days, to generate results using this form of diagnosis. RT-
knowledge to improve diagnostic accuracy. As a result, the PCRs are both time-consuming and expensive. Subsequently,
systems are considered intelligent systems since they employ experts are facing significant challenges in developing alterna-
a process of feedback to continuously enhance their perfor- tives via detection technologies. AI is being used to automate
mances. Large clinical data is complicated to analyze. Intelli- the diagnosis of many diseases today, and AI has been proven
gent Computer-aided diagnosis systems using data mining, to achieve superior performance during automatic image cate-
artificial intelligence (AI), and deep learning methodologies gorization using various machine learning algorithms. The
are beneficial in diagnosing an array of illnesses and medical detection based on the image processing is based on the classi-
disorders. fication of the features extracted from the CXR or CT, as
In the last century, researchers have accumulated substan- shown in Fig. 1. Furthermore, machine learning specifies mod-
tial knowledge regarding human anatomy and physiology. els that have the capability of learning and making decisions
In recent years, chest X-rays (CXR), ultrasounds, and mag- based on a massive input of data samples.
netic resonance imaging (MRI) have played vital roles in In the context of deep learning, the extraction and classifi-
enhancing the accurate diagnosis of human diseases. Signifi- cation of features from images is the primary goal. Deep learn-
cant improvements in healthcare and medical research have ing has been a huge success in a wide range of industries,
helped people to improve their quality of life as new technolo- including healthcare [3]. In addition, deep learning can develop
gies have facilitated the accurate diagnoses of patients’ ail- models that can accurately predict and diagnose illnesses using
ments and diseases. In the last few decades, medical experts images. It has been effective in diagnosing TB [5–10], pneumo-
faced challenges in conducting an accurate diagnosis of dis- nia [11–18], lung cancer [19–23], and COVID-19 diagnosis,
eases, which compounded unnecessary healthcare and mal- without need of human expertise. Unlike traditional machine
practice claims for both the doctor and patients. Machine learning, the fundamental reason behind using deep learning
learning, deep learning, and statistical analysis are effective techniques is its ability to build the model of input as the size
tools for computer-aided diagnosis. These tools are used in of network deeply grows. Because of this, the model automat-
solving difficult computer vision tasks in medical imaging, ically gathers data and generates findings that are more accu-
such as segmenting lungs, classifying lung diseases, and so rate. Deep learning models, in contrast to typical machine
on. With recent developments in deep learning, machines can learning algorithms, describe features using a sequence of
perform equally or better than humans on a wide range of non-linear functions that are incorporated to optimize the
activities. For example, deep learning can be used to calculate accuracy of the utilized model.
the treatment outcomes, such as cancer therapy. With huge- The remaining part of this paper is organized as follows:
labelled datasets and deep learning-based approaches, promis- Section 2 presents the research-related works and background
ing findings are developed in the categorization of thoracic dis- on chest diseases detection and classification based deep learn-
orders using a CXR modality. In addition, machine learning is ing. The details of the materials and methods, the dataset for
the model that can learn and make decisions based on a vast the study, data preprocessing, and the proposed deep learning
number of input data sets. Artificial intelligence performs framework are described in Section 3. Section 4 incorporates
activities that require human intellect, such as voice recogni- the results, including the experimental parameters and the per-
tion, translation, and the ability to analyze colors and shapes formance metrics for our proposed system, with comparisons
by evaluating the incoming data and making predictions. A to current state-of-the-art systems of chest diseases detection
combination of machine learning algorithms, known as deep and classification. Finally, conclusions and possible ongoing
learning, has demonstrated remarkable success in various sec- future work are given in Section 5.
tors, particularly in the healthcare sector [2,3]. Deep learning
models can accurately predict and categorize numerous dis- 2. Literature review
eases, such as tuberculosis (TB), lung cancer, pneumonia,
and currently COVID-19 using images, without human inter- In most countries, chest computed tomography (CT) and X-
vention. As the network becomes larger, data representation ray pictures are widely utilized as a feasible option for identi-
becomes deeper, making deep learning to be more effective, fication of COVID-19. However, COVID-19 identification is a
contrary to classical machine learning. Consequently, the complex process, which requires clinical imaging of patients
model automatically collects characteristics and generates [24–30]. Lung cancer represents a major source of mortality
more accurate outcomes. Since the models use a combination in humans. The immediate diagnosis could improve human
of non-linear functions rather than linear functions, deep survival [19–22]. Applying machine learning and image pro-
learning algorithms are more accurate than typical machine cessing have presented considerable promise for lung cancer
learning methods. diagnosis. In this section, an exhaustive evaluation of deep
A deep learning architecture for multi-class lung diseases classification 925

Fig. 1 Image processing-based classification model.

learning models for TB, COVID-19, lung cancer, and pneumo- patients were grouped into two (174 patients) training and
nia are discussed. (43 patients) testing. The authors created a 3-dimensional deep
Transferring learning tools, such as VGG-16, ResNet-50, learning network using the clipped segments and multivariable
and InceptionV3, to clinical pictures of lung illnesses and logistic regression to integrate relevant radiomics characteris-
COVID-19 has offered promising results [24]. It is discovered tics and deep learning scores. To test the robustness of their
that pneumonia is among major symptoms of COVID-19. methods, they used stratified analysis, cross-validation, deci-
Transfer learning helps discover that the same virus causes sion curve analysis, and survival analysis. An AUC of 0.909
pneumonia and COVID-19. A study demonstrates that the distinguishes between critical patients in the test and training
information obtained by a model trained to detect viral pneu- groups [31]. Another study applied InceptionV3, NASNet,
monia may be applied to identify COVID-19 [24]. As a result, Xception, DenseNet, MobileNet, VGGNet, Incep-
Haralick features can be used to facilitate feature extraction. tionResNetV2, and ResNet for classifying the COVID-19,
This approach involves statistical analyses that focus on a which was tested on the mixed dataset of CXR and CT images.
specific area of COVID-19 diagnosis. In comparison to the tra- DenseNet121 offered the best performance with an accuracy of
ditional classifications, transfer learning has consistently pro- 99 % [29].
ven to offer statistically significant outcomes [24]. Some Image segmentation is used to categorize chest CTX into
studies developed and analyzed a fully automated COVID-19 pneumonia, COVID-19, and normal illnesses using four
detection framework utilizing CTX. To diagnose COVID-19, CNN base learners, a modified stack ensemble model, and
the visual features were extracted from volumetric chest CT Naive Bayes as the meta-learner in one research. For
images using COVID-19 neural network approach. The out- COVID-19, pneumonia, and normal classes, the suggested
comes show that the approach has outperformed the existing technique beats current techniques by 0.9867 on standard
work. Pre-trained models-based convolutional neural network datasets and 0.98 Kappa on the same datasets [32] based on
(CNN) architecture such as Inception-ResNetV2, ResNet152, CT scans. By reducing manually labelled CT images, the sug-
ResNet50, InceptionV3, and ResNet101 was used in related gested technique may accurately detect COVID-19 infections
work to identify COVID-19 pneumonia based on the CXR and rule out the case of COVID-19. Based on the positive
images. Among the existing models, the ResNet50 exhibited qualitative and quantitative results, we envision our created
the most accurate classification outcomes [28]. The comparison approach is widely used in large-scale clinical trials [33]. The
and modelling were based on CT images of 101 pneumonia, 88 convolution neural networks are effective in converting 360
COVID-19 patients, and 86 healthy cases from two areas in X-ray and CT scan pictures into a categorization on a binary
China. A details relation extraction neural network (DRENet) class pneumonia-based translation of decision tree, Inception
learning-based CT diagnostic algorithm identified COVID-19 V2, and VGG-19 models. Compared to decision tree (60 %)
patients. The model correctly distinguished between COVID- models and Inception V2 (78 %), the fine-tuned version
19 patients with a recall of 0.93, AUC of 0.99, and accuracy VGG-19 (91 %) exhibits the greatest increase in training and
of 0.96. The research showed that deep learning based on validation precision [33].
CT scans may help to detect COVID-19 patients and automat- The GSA-DenseNet121-COVID-19 is a unique mixed
ically identify possible abnormal changes. Another study cate- CNN architecture that utilizes DenseNet121 and the optimiza-
gorized COVID-19 CXR images by applying modified tion technique as gravitational search (GSA). The
MobileNet and a ResNet architecture. With this approach, DenseNet121-COVID-19 could identify COVID-19 better
characteristics from multiple CNN layers are dynamically than other DenseNet121, which could only diagnose 94 % of
combined to overcome the gradient vanishing problem. The the test set. The suggested method was contrasted to an
proposed approaches outperform over the current methods Inception-v3 CNN architecture and manual analysis when
by 99.3 % on the CT image dataset and by 99.6 % on the computing hyperparameter estimates. The GSA-
CXR. [27]. DenseNet121-COVID-19 outclassed the comparison tech-
Some studies developed a model to distinguish between crit- nique, which could only categorize 95 % of the test set samples
ical and severe COVID-19 instances using deep learning char- [34].
acteristics and radiomics based on D-Resnet [29,31]. These EfficientNet-based pre-trained models were lowered using
authors studied 217 individuals in three Chinese hospitals, 82 kernel principal component analysis. Then, multiple retrieved
with extreme severity and 135 with serious disease. The features were merged using a feature fusion technique. Finally,
926 G.M.M. Alshmrani et al.

stacked ensemble meta-classifiers were used to classify the (ODNN) and Linear Discriminate Analysis (LDA) to evaluate
model into two stages. CT lung images that reduces the dimensionality of deep fea-
Predictions were made in the first step using a support vec- tures. The ODNN is used with CT scans and optimized using
tor machine (SVM) and a random forest, which were then the Gravitational Search Algorithm to classify lung cancer,
pooled and fed into the second stage. Next, a logistic regres- thereby offering a 96.2 % as sensitivity, 94.2 % as specificity,
sion classifier divides the X-ray and CT data into 2 classes and 94.56 % as accuracy [19].
(COVID, NON-COVID). The model’s performance was com- Since medical specialists’ face challenges in distinguishing
pared to other CNN-based pre-trained models. The new model between Covid-19 and pneumonia, one study utilized an artifi-
outperforms previous approaches and may be used by clini- cial neural network, ensemble classifier, SVM, and KNN for
cians for point-of-care diagnosis [35]. In a comparable work, categorization. However, a RNN with a LSTM has been pro-
researchers used ResNet32 and the deep transfer learning tech- posed as a deep learning architecture to identify lung condi-
nique to categorize COVID-19-infected patients, and the tions. Outcomes of the experiments demonstrated the
results were published. Comparing the COVID-19 classifier resilience and effectiveness of the suggested model [18].
to earlier supervised learning models, experimental data Another work uses ensemble of InceptionResNet_V2,
demonstrated that it delivered superior outcomes when com- ResNet50, and MobileNet_V2 for classifications. The out-
pared to previous learning models [36]. comes revealed that the ResNet50, MobileNet_V2, and Incep-
A cutting-edge attention-based deep learning model with tionResNet_V2 models provide an F1 score of 94.84 %, which
VGG-16 and a fine-tuned classification process was designed is higher than other models [17]. In addition, the CNN with
using a unique deep learning model that uses a convolution pre-trained weights is utilized to categorize Covid-19, pneumo-
layer of the VGG-16 models for COVID-19. The experimental nia, and healthy individuals using transfer learning techniques.
analysis shows steady and promising performance after com- Those who have active SARS-CoV-2 and pneumonia were
paring the suggested approach to the existing models [38]. accurately categorized in the dataset, which is one of the most
The integrated stacking deep convolutional network using important discoveries of that work [16]. Another study exam-
pre-trained models like ResNet101 and XceptionV3 was ined the potential of using machine learning to delineate and
applied for InstaCovNet-19. The accuracy of 0.99 for 3 classes pinpoint pneumonia in CXR using RetinaNet and Mask R-
(Normal, Pneumonia, COVID-19) and 0.9953 for 2 classes CNN as an ensemble for the identification and localization
(COVID, non-COVID) is achieved. In ternary classification, of pneumonia, thereby achieving a recall of 0.793 for a large
the suggested model obtained 98 % accuracy, whereas binary dataset [15]. For a variety of lung diseases, the transfer learn-
classification achieved 100 % precision and 98 % recall [38]. ing approach was used to capture images on CXR and CT. As
The CNN is used to implement binary and multiclass clas- COVID-19 resembles pneumonic viral lung illness, COVID-19
sification. The model was trained on 3877 CT and X-ray pic- detection is challenging and relies on a thorough examination
tures, of which 1917 were of Covid-19 affected people. The of a patient’s clinical pictures. The goal is attained using a
binary classifier achieved a 99.64 % accuracy and exhibited a novel architecture trained to identify virus-related pneumonia
99.58 % recall, a 99.56 % precision, a 99.59 % F1 score, for COVID-19 detection. When compared to traditional cate-
and a 100 % ROC. The model was trained with 6077 pho- gories, the findings of transfer learning are strikingly different
tographs. A total of 1917 of which were of Covid-19 infected [14].
patients, 1960 healthy people, and 2200 pneumonia patients. One study develops the CNN model from scratch to extract
The suggested technique obtained 98.28 % accuracy, characteristics from an image of pneumonia infected person’s
98.25 % recall (or sensitivity), 98.22 % precision, 98.23 % chest X-ray and categorize it. This concept might alleviate
F1-score, and 99.87 % ROC for multiclass classification [40]. some of the issues associated with dealing with medical images.
The early detection of lung cancer increases survival It is difficult to obtain a significant number of pneumonia
chances from 14 % to 49 %. Although CT approaches are datasets for this classification assignment due to the limited
found to deliver more accuracy than X-rays, a conclusive diag- availability of such data. The multiple data augmentation
nosis that relies on many imaging modalities. An artificial strategies were used to increase the training and validation
DNN can spot lung cancer in CT images. Therefore, studies classification accuracy of the proposed model. This has
have proposed an adaptive boosting technique and a Dense- achieved a significant precision of 0.94814 the validation phase
Net to classify the lung image as normal or malignant. A total [13]. The transfer learning system automatically differentiates
of 201 lung pictures have been included in the training dataset, between 3883 CXR pictures classified as exhibiting pneumonia
with 85 percent of them being utilized for training and 15 per- and 1349 that are designated normal. As an initialization, the
cent being used for testing and classification. The proposed suggested technique makes use of weights pre-trained on Ima-
approach was shown to achieve a 90 % accuracy in testing geNet using the Xception Network. When compared to cur-
[22]. The MLP classifier offered a higher accuracy of rent approaches, the model is competitive to obtain 0.84,
88.55 % than the alternative classifiers, according to the out- 0.91, 0.99, 0.97 as precision, recall, F1, and ROC respectively
come of the analysis of a study [21]. The CNN, DNN, and [12]. In a separate study, researchers studied 180 X-ray images
sparse auto-encoder deep neural networks were employed to of persons who had been infected with COVID-19. The
identify lung cancer calcification. CT scans of benign and research attempted to employ the most successful systems,
malignant lung nodules were classified using these networks. such as ResNet50V2 and Xception networks to detect the
The Lung Image Database Consortium image collection virus. Overall, the suggested model achieved a 91.4 % accuracy
(LIDC) database examined the networks where accuracy was for all classes and a 99.50 % accuracy for instances of COVID-
84.15 %, sensitivity 83.96 %, and a specificity 84.32 % [20]. 19 [11]. Using a CXR dataset from the National Library of
The CNN was the most accurate of the three networks. Medicine Shenzhen No.3 Hospital, researchers developed a
Another work called applied Optimal Deep Neural Network DCNN model to detect tuberculosis. This dataset was com-
A deep learning architecture for multi-class lung diseases classification 927

pared with a non-TB-specific chest X-ray dataset of a different The hierarchical feature extraction for abnormality detec-
population. The DCNN offered an AUC as 0.9845 and 0.8502. tion method uses two levels of hierarchy to classify character-
The AUC of the supervised DCNN model in the CXR dataset, istics into healthy and unhealthy categories. Two levels of
on the other hand, was much lower, at 0.7054, than in the feature extraction are identified: level one is handmade geo-
other datasets. A total of 36.51 % of aberrant radiographs metrical feature extraction, and level two is typical statistical
in the CXR dataset associated to tuberculosis were predicted feature extraction and textural feature extraction from seg-
by the final DCNN model [5]. mented lung fields. They were tested on 800 CXR images
Another study combined ResNet and depth-ResNet in pre- derived from two public datasets to verify their performance.
dicting severity scores and an analysis of TB’s likelihood. A AUC = 0.99 0.01 for Shenzhen and 0.95 0.06 for Mont-
depth-ResNet of 92.70 % and ResNet-50 of 67.15 % were pro- gomery, which illustrated that the two TB detection
duced for the TB detection. The study used the overall severity approaches offered a promising performance as compared to
probability, different likelihoods for high severity (1 to 3 the existing techniques, as demonstrated by the obtained find-
scores), and low severity (4 and 5 scores), where scores of 1 ings. Furthermore, Friedman’s posthoc multiple comparison
to 5 were converted into the probabilities of 0.9, 0.7, 0.50, methods are demonstrated to statistically validate the sug-
0.30, and 0.2. A 75.88 % and 85.29 %, respectively are the gested method [8]. In [43,44] COVID-19 classification-based
averaged accuracies for both approaches [6]. Other studies pro- detection based on machine learning was introduced. The com-
posed three standard designs in the ensemble technique, mon reviewed studies about the chest diseases detection and
namely AlexNet, GoogleNet, and ResNet. As a result, a new classification are summarized in Table 1.
classifier for TB categorization has been developed from
scratch. A combined dataset of publicly accessible standard 3. Proposed methdology
datasets is used to train and test the suggested approach. Accu-
racy of 88 % and the AUC of 0.93 %, which is better than The human respiratory system is attacked by a variety of lung
most existing approaches, are achieved [7]. illnesses. These diseases include pneumonia, tuberculosis, lung

Table 1 Comparative analysis of existing work.

Disease Study Method Medical Image Performance
Acc. Prec. Sens.
COVID-19 [24] VGG-16, ResNet-50, InceptionV3 CXR + CT 93 91 90
[25] VGG-19 + ResNet-50 CT 94 95 90
[26] DRE-Net CT 86 96 93
[27] Modified ResNet CXR + CT 99.3 99.7 99.1
[28] ResNet50 CXR 96.1 76.5 91.8
[29] DenseNet121 CXR + CT 98 96 96
[30] VGG-16 CXR 98.67 100 98
[31] D-Resnet-10 network CT 81.4 79.8 87.5
[32] VGG + CNN CT 96.2 97.3 94.5
[33] VGG-16, InceptionV2, DT CXR + CT 91 94 97
[34] GSA-DenseNet121 CXR 98.38 98.5 98.5
[35] Deep learning Meta classifier CXR + CT 99 99 99
[36] ResNet32 + DTL CT 93 95 91
[37] VGG-16 CXR 79.58 92 95
[38] InstaCovNet-19 CXR 99.08 99 99
[39] CNN CXR + CT 98.28 98.22 98.25
Lung Cancer [22] FPSO-CNN CT 95.62 96.32 97.93
[21] Multi-layer Perceptron (MLP) CT 88.55 86.59 89.84
[20] CNN CT 84.15 84.32 83.96
[19] MGSA CT 94.56 94.2 96.2
Pneumonia [18] RNN-LSTM CXR 95.04 88.89 95.41
[17] ResNet50 + MobileNetV2 + InceptionResNetV2 CXR 95.09 95.53 94.43
[16] CNN with pre-trained weights on ImageNet CXR 91 92 87
[15] RetinaNet and Mask R-CNN CXR 83.80 75.8 79.3
[14] Transfer learning CXR + CT 94.9 93 93
[13] CNN CXR 93.73 – –
[12] Xception Network pre-trained weights on ImageNet CXR 97.3 84.3 99
[11] Xception + ResNet50V2 CXR 99.50 92.69 80.53
Tuberculosis [5] DCNN CXR 98.45 82 72
[6] Depth-ResNet CT 85.29 – 84.16
[7] Ensemble (AlexNet, GoogleNet and ResNet) CXR 88.24 88.0 88.42
[8] (SVM + FOSF + GLCM) CXR 99.40 99.42 99.40
Lung Opacity [41] Xception CXR 95.71 – 97.19
[42] ResNet2 CXR 88.68 – –
928 G.M.M. Alshmrani et al.

cancer, and lung opacity, among others. These diseases can opacity, and most recently Covid-19 were accessed and col-
cause similar effects on human lungs therefore X-ray images lected from reliable sources. To begin with, for COVID-19,
are commonly employed for diagnosing these diseases. AI in 3615 CXR images from various sources of public datasets
the form of deep learning algorithms has increasingly played and published studies [45–51] were included in this study.
a key role in diseases identification and classification. Deep Secondly, 5856 CXR images of pneumonia were extracted
learning facilitates the diagnosis process and saves time for from the Radiological Society of North America (RSNA),
healthcare providers. the Italian Society of Medical and Interventional Radiology
The study presents a multiclass deep learning classification (SIRM), and Radiopaedia which are publicly available for
model to identify the most common chest diseases. The aim of research purpose [52,53]. To distinguish COVID-19 from
the research work is to design a deep learning framework and pneumonia as a part of our experiments, these datasets have
classify multi-class of Pneumonia, Lung Cancer, TB, Lung been utilized for training our proposed deep model. Further-
Opacity, and most recently COVID-19. A thorough search more, the Radiological Society of North America [52] repre-
of the literature shows that our study is the first attempt to sents 6012 CXR images of Lung opacity whilst [54,55]
use the single deep learning framework, incorporating and indicate the datasets resource for a total of 20,000 X-ray
classifying all these six classes at a time. Fig. 2 represented images of lung cancer. The fourth dataset describes 5870
the proposed framework in a block diagram. The framework CXR images of pneumonia cases obtained from a variety of
is divided into three phases as pre-processing, feature extrac- research articles [53,54]. Ultimately, the total of 1400 X-ray
tion, and classification. As observed in Fig. 3, X-ray pictures images for tuberculosis [56,57] were collected and employed
were used as inputs, and the categorization of the input X- in our research.
ray image on a disease level was the final output of the model. 80,000 specifies the total number of CXR images used in
During the first phase, the input images undergo pre- our experiments. Samples of chest X-ray images for COVID-
processing function such as normalization, resizing, and data 19, normal, pneumonia, TB, lung opacity and lung cancer
image splitting into 80 % training and 20 % validation at ran- are shown in Fig. 3. The number of patients for each disease
dom. Then, deep learning algorithms are used during the sec- dataset with respect to ages: ages were frequently between 38
ond and third stages. The second phase involves feature and 65 for the COVID-19 dataset, 26 and 62 for the pneumo-
extraction, which is performed using VGG19 and CNN tech- nia dataset, 28 and 58 for the lung cancer dataset, and for nor-
niques. The fully connected network technique is employed mal patients the ages were between 33 and 58 years.
during the image categorization step.
3.2. Dataset Pre-processing
3.1. Dataset
Some pre-processing processes were employed to adjust the
For the experimental purpose, in addition to healthy cases, input data to meet the requirements of the deep learning
tremendous X-ray images of pneumonia, TB, lung cancer, lung model: 1) The images were resized; 2) the images were normal-

Fig. 2 Diagram of the proposed framework.

A deep learning architecture for multi-class lung diseases classification 929

Fig. 3 Chest X-ray images: (a) Tuberculosis images, (b) Pneumonia images, (c) Normal images, (d) Lung Opacity images, (e) COVID-19
images, (f) Lung cancer images.

ized; 3) the images were converted to an array to be employed The use of X-ray pictures to identify specific forms of chest
an input in the model’s next phase. To ensure that the varia- ailments is demonstrated using a VGG19 followed by a CNN
tion of the images meets the requirements to train the pro- model. The model is depicted in detail in Fig. 4. X-ray chest
posed model, the data were randomly divided into validation pictures with a dimension of 224 * 224 * 3 were used as input
and training subsets at 20 % and 80 %, respectively. To meet data for our model. The VGG19 pre-trained model is followed
the criteria of the framework, all images were scaled to by three CNN blocks during the feature extraction stage.
224 * 224 * 3. VGG19 is designed to provide great accuracy for large-scale
After normalizing each pixel in the image to the interval image applications. [58] feature architecture was used that
[0,1], all images were transformed to array data representation. comprised 19 CNN with 3 convolution filters and 1 stride.
Multiple deep learning models were merged with the VGG19
3.3. Proposed deep learning VGG19 + CNN model to improve picture categorization accuracy. A convolution
layer with a ReLU as an activation function is included in each
Supervised deep learning for multiclass classification of the CNN block. Following these three CNN blocks, a batch nor-
most common chest diseases is presented in this research. malization and a max-pooling layer were applied, which were
For classification, we used a pre-trained model, VGG19, and then followed by a dropout layer, as indicated in Fig. 4.
the CNN as a feature extraction model which has fully In the feature extraction step, the output was turned into a
connected. one-dimensional data vector, which was then used as an input
930 G.M.M. Alshmrani et al.

Fig. 4 The model architecture.

has storage of 2 TB, RAM 25 GB, and CPU-P100. The

ImageDataGenerator class in Keras [60] was used during the
pre-processing stage, which included picture scaling, normal-
ization, and conversion to an array of data.
The suggested multi-chest illnesses classification deep learn-
ing model input was created using the outcome of the pre-
processing step. An optimizer and appropriate fit algorithms
were used with 5000 epochs to train and validate the model.
Eight iterations and 32 batch sizes were employed in each
epoch. With the greatest precision, the performance metrics
formulae were entered into the validation data outputs. The
Adam [58] optimizer was employed, with a learning rate of
0.000009 (LR). Our suggested deep learning model’s code
was published on the GitHub website [61].
Fig. 5 Pseudo-Code for the proposed framework. Precision, loss, F1-Score, accuracy, precision, the AUC,
and recall were applied to measure the performance of the
in the classification stage after being modified through the flat- model. The accuracy was determined in terms of the number
tening layer. The remaining components of the categorization of instances with correct predictions out of the total number
step are comprised of three thick layers, each having 512, 256, of instances. The positive predictive value represented the pre-
and 128 neurons. It is a thick layer with six neurons and the cision. The accuracy was determined in line with the number of
SoftMax activation function that generates the final classifica- samples that was positive versus those forecasted as positive.
tion output. This layer is responsible for classifying the output Th F1-Score was calculated based on the harmonic mean of
image into one of the six chest diseases classes: pneumonia, precision. According to this, the percentage of samples pre-
tuberculosis, lung cancer, and lung opacity. A total of dicted to be positive from the total number of samples which
24,622,470 model parameters are span into two categories. considered to be the true positive rate. These matrices are
First were the trainable parameters (24,622,342), which were defined as in equations (1) to (4) [62]:
revised throughout the training process. The best value for Tp þ Tn
Accuracy ¼ ð1Þ
these parameters was required to ensure the training accuracy. Tp þ Tn þ Fp þ Fn
The second category was the untrainable parameters (1 2 8),
which were those that did not change at the time of training. 2 Tp
F1  score ¼ ð2Þ
Fig. 5 illustrate the pseudo-code for the proposed framework. 2 Tp þ Fp þ Fn

4. Results Tp
Precision ¼ ð3Þ
T p þ Fp
With the help of Python 3 and the Keras framework, we were
able to create a classification model for chest disease. The Tp
RecallðSensitivityÞ ¼ ð4Þ
model was simulated on a Google Colab Pro edition [59] that Tp þ Fn
A deep learning architecture for multi-class lung diseases classification 931

Where the actual positive and negative parameters are denoted COVID-19 images. To show the effectiveness of the proposed
by Tp and Tn respectively. The False positive and false negative model, Table 3 introduces a comparative analysis of fifteen
values are denoted by Fp and Fn , respectively. In addition, a existing works. The comparison process was evaluated based
confusion matrix is computed for the proposed model. on the architecture performance, the number of classes, and
Fig. 6 shows how the performance of the varying their deep learning approaches.
epochs during the training and validation phases.
4.3. Architecture performance
4.1. Experimental results
Accuracy, precision, and recall (sensitivity) are the major
The suggested VGG19 + CNN model validation metric values parameters used to measure the performance of the model.
are shown in Table 2. These values represent those acquired The accuracy of the proposed framework produced the highest
during the highest accuracy validation iteration. The accuracy results with 96.48 overcoming the rest models in Fig. 7.
ðTp Þ, precision (Fp ), F1-score (Tn ), and recall (Sensitivity) (Fn ) As revealed in Fig. 8, the best precision value was 97.56
were calculated using equations 1–4. with our proposed model. However, Fig. 9 confirms that our
proposed model achieves the highest sensitivity of 93.75 com-
4.2. Comparative anaysis pared to others.
As presented in Fig. 10, various architectures of individual
To the best of our knowledge, there is no recent existing pre-trained models, transfer learning, and ensemble techniques
research has employed a single deep learning model for evalu- based on deep learning have been investigated and compared
ating and classifying the following chest diseases together: with our multi-class proposed framework. The results show
Tuberculosis, Pneumonia, Lung Opacity, Lung cancer, and that the proposed VGG19-CNN achieved the best perfor-

Fig. 6 Performance metrics changes with epochs in training and validation phases.

Table 2 The performance validation of the VGG19 + CNN model.

Methods Loss Acc Pre AUC F1 Recall
VGG19 + CNN model 0.1792 96.48 97.56 99.82 95.62 93.75
932 G.M.M. Alshmrani et al.

Table 3 The performance comparison between the proposed VGG19 + CNN and existing related work.
Ref Number of Classes Method Medical Image Performance
Acc. Prec. Sens.
[24] 3 VGG-16, ResNet-50, InceptionV3 CXR + CT 93 91 90
[25] 3 VGG-19 + ResNet-50 CT 94 95 90
[26] 3 DRE-Net CT 86 96 93
[28] 2 ResNet50 CXR 96.1 76.5 91.8
[36] 2 ResNet32 + DTL CT 93 95 91
[31] 2 D-Resnet-10 network CT 81.4 79.8 87.5
[21] 2 Multi-layer Perceptron (MLP) CT 88.55 86.59 89.84
[20] 2 CNN CT 84.15 84.32 83.96
[16] 3 CNN with pretrained weights on ImageNet CXR 91 92 87
[15] 2 RetinaNet and Mask R-CNN CXR 83.80 75.8 79.3
[14] 3 Transfer learning CXR + CT 94.9 93 93
[13] 2 CNN CXR 93.73 – –
[6] 1 Class with 5 Levels of Severity Depth-ResNet CT 85.29 – 84.16

[7] 2 Ensemble (AlexNet, GoogleNet and ResNet) CXR 88.24 88.0 88.42
[42] 3 ResNet2 CXR 88.68 – –
Proposed 6 VGG19 + CNN CXR 96.48 97.56 93.75

Fig. 7 Competitive analysis based on Accuracy. Fig. 9 Competitive analysis based on Sensitivity.

Fig. 8 Competitive analysis based on Precision. Fig. 10 Competitive analysis based on a variety of utilized deep
learning approaches.

mance. ResNet50 [28] was better than Transfer Learning [14].

However, the Ensemble model [7] records the lowest. chest diseases: tuberculosis, pneumonia, lung opacity, lung
Moreover, Table 3 illustrates the proposed multi-class cancer, and COVID-19, in addition to normal cases. The
framework used to classify-six classes of the most popular model significantly outperformed binary classes presented by
A deep learning architecture for multi-class lung diseases classification 933

Funding statement

This research was funded by the Lancaster University, U.K.

Declaration of Competing Interest

The authors declare that they have no known competing

financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgement

Fig. 11 The confusion matrix. This work was supported in part by The Engineering and
Physical Sciences Research Council (EPSRC) under Grant
[28,21,15], and [7]. Likewise, the model got over multi-class as EP/P009727/1, and in part by the Leverhulme Trust under
observed by [25,16,14], and [42]. Grant RF-2019-492.
The confusion matrix for the VGG19 + CNN proposed
model is shown in Fig. 11, revealing that the References
VGG19 + CNN model can successfully classify the six chest
diseases with the highest ratio to the COVID-19, starting from [1] J. Yanase, E. Triantaphyllou, A systematic survey of computer-
lung opacity, normal chest, lung cancer, pneumonia, and lastly aided diagnosis in medicine: Past and present developments,
the tuberculosis disease. Expert Syst. Appl. 38 (2019) 112821, https://doi.org/10.1016/j.
A multiclass deep learning classification model has been eswa.2019.112821.
used in this work to incorporate and classify-six classes of [2] M.A. Khan, An IoT Framework for Heart Disease Prediction
COVID-19, lung opacity, TB, lung cancer, and pneumonia Based on MDCNN Classifier, IEEE Access 8 (2020) 34717–
34727, https://doi.org/10.1109/ACCESS.2020.2974687.
using the VGG19 + CNN approach. The architecture of the
[3] M.A. Khan, M.T. Quasim, N.S. Alghamdi, M.Y. Khan, A
model was based on VGG19 + CNN for feature extraction Secure Framework for Authentication and Encryption Using
and a fully linked network for classification. Improved ECC for IoT-Based Medical Sensor Data, IEEE
The recall, accuracy, AUC, F1 score, and precision of the Access 8 (2020) 52018–52027, https://doi.org/10.1109/
suggested model were all tested. The findings showed that ACCESS.2020.2980739.
the VGG19 + CNN provided satisfactory classification per- [4] S. Albahli, Efficient gan-based chest radiographs (CXR)
formance with 96.48 % accuracy, as shown in Table 3. augmentation to diagnose coronavirus disease pneumonia, Int.
Based on X-ray images, the VGG19 + CNN can identify J. Med. Sci. 17 (10) (2020) 1439–1448, https://doi.org/10.7150/
various chest disorders with 96.48 % accuracy, 93.75 % recall, IJMS.46684.
97.56 % precision, 95.62 % F1 score, and 99.82 % AUC. We [5] S. Sathitratanacheewin, P. Sunanta, K. Pongpirul, Deep
learning for automated classification of tuberculosis-related
expect that the deep learning model will contribute to the
chest X-Ray: dataset distribution shift limits diagnostic
development of a model for diagnosing chest disorders from performance generalizability, Heliyon 6 (8) (2020) e04614.
CXR chest pictures, improving patient outcomes, and saving [6] X.W. Gao, C. James-Reynolds, E. Currie, Analysis of
lives. tuberculosis severity levels from CT pulmonary images based
on enhanced residual deep learning architecture,
5. Conclusion and future work Neurocomputing 392 (2020) 233–244, https://doi.org/10.1016/j.
neucom.2018.12.086.
[7] R. Hooda, A. Mittal, S. Sofat, Automated TB classification
In this study, a multi class chest diseases classification based a using ensemble of deep architectures”, Multimed. Tools Appl.
deep learning architecture was developed and evaluated for 78 (22) (2019) 31515–31532, https://doi.org/10.1007/s11042-019-
classifying TB, lung opacity, lung cancer, pneumonia, normal, 07984-5.
and COVID-19 using CXR images. In terms of classification, a [8] T.B. Chandra, K. Verma, B.K. Singh, D. Jain, S.S. Netam,
pre-trained model, VGG19 followed by three blocks of convo- Automatic detection of tuberculosis related abnormalities in
lutional neural network (CNN) as a feature extraction and Chest X-ray images using hierarchical feature extraction scheme,
fully connected network at the classification stage were intro- Expert Syst. Appl. 158 (2020), https://doi.org/10.1016/j.
duced. The experimental results revealed that our proposed eswa.2020.113514 113514.
VGG19 + CNN outperformed other existing work with [9] U.K. Lopes, J.F. Valiati, Pre-trained convolutional neural
networks as feature extractors for tuberculosis detection,
96.48 % accuracy, 93.75 % recall, 97.56 % precision,
Comput. Biol. Med. 89 (2017) 135–143, https://doi.org/
95.62 % F1 score, and 99.82 % area under the curve (AUC). 10.1016/j.compbiomed.2017.08.001.
The CT scan can accurately detect the aberrant pattern [10] J.S. Kumar, S.A. Alias Balamurugan, S. Sasikala, Analysis of
even before symptoms appear. Therefore, employing a combi- Deep Learning Techniques for Tuberculosis Disease, SN
nation of CXR and CT images is a potential enhancement Comput. Sci. 2 (4) (2021) 1–10, https://doi.org/10.1007/s42979-
parameter for the future work. Moreover, the identification 021-00680-y.
of the region of interest (ROI) in conjunction with severity [11] M. Rahimzadeh, A. Attar, A modified deep convolutional
levels classification based on a powerful segmentation model neural network for detecting COVID-19 and pneumonia from
is another direction for the exploration of the future work. chest X-ray images based on the concatenation of Xception and
934 G.M.M. Alshmrani et al.

ResNet50V2, Inform. Med. Unlocked 19 (2020), https://doi.org/ [27] G. Jia, H.K. Lam, Y. Xu, Classification of COVID-19 chest X-
10.1016/j.imu.2020.100360 100360. Ray and CT images using a type of dynamic CNN modification
[12] J.E. Luján-Garcı́a, C. Yáñez-Márquez, Y. Villuendas-Rey, O. method, Comput. Biol. Med. 134 (2020), https://doi.org/
Camacho-Nieto, A transfer learning method for pneumonia 10.1016/j.compbiomed.2021.104425 104425.
classification and visualization, Appl. Sci. 10 (8) (2020), https:// [28] A. Narin, C. Kaya, Z. Pamuk, Automatic detection of
doi.org/10.3390/APP10082908. coronavirus disease (COVID-19) using X-ray images and deep
[13] O. Stephen, M. Sain, U.J. Maduh, D.U. Jeong, An Efficient convolutional neural networks, Pattern Anal. Appl. 24 (3)
Deep Learning Approach to Pneumonia Classification in (2021) 1207–1220, https://doi.org/10.1007/s10044-021-00984-y.
Healthcare, J. Healthc. Eng. (2019), https://doi.org/10.1155/ [29] S.H. Kassania, P.H. Kassanib, M.J. Wesolowskic, K.A.
2019/4180949. Schneidera, R. Detersa, Automatic Detection of Coronavirus
[14] M.R. Lascu, Deep Learning in Classification of Covid-19 Disease (COVID-19) in X-ray and CT Images: A Machine
Coronavirus, Pneumonia and Healthy Lungs on CXR and CT Learning Based Approach, Biocybern. Biomed. Eng. 41 (3)
Images, J. Med. Biol. Eng. 41 (4) (2021) 514–522, https://doi. (2021) 867–879, https://doi.org/10.1016/j.bbe.2021.05.013.
org/10.1007/s40846-021-00630-2. [30] R.K. Singh, R. Pandey, R.N. Babu, COVIDScreen: explainable
[15] I. Sirazitdinov, M. Kholiavchenko, T. Mustafaev, Y. Yixuan, R. deep learning framework for differential diagnosis of COVID-19
Kuleev, B. Ibragimov, Deep neural network ensemble for using chest X-rays, Neural Comput. Appl. 33 (14) (2021) 8871–
pneumonia localization from a large-scale chest x-ray 8892, https://doi.org/10.1007/s00521-020-05636-6.
database, Comput. Electr. Eng. 78 (2019) 388–399, https://doi. [31] C. Li, D. Dong, L. Li, W. Gong, X. Li, Y. Bai, M. Wang, Z. Hu,
org/10.1016/j.compeleceng.2019.08.004. T. Zha, J. Tian, Classification of Severe and Critical Covid-19
[16] J.E. Luján-Garcı́a, M.A. Moreno-Ibarra, Y. Villuendas-Rey, C. Using Deep Learning and Radiomics, IEEE J. Biomed. Heal.
Yáñez-Márquez, Fast COVID-19 and pneumonia classification Informatics 24 (12) (2020) 3585–3594, https://doi.org/10.1109/
using chest X-ray images, Mathematics 8 (2020) 9, https://doi. JBHI.2020.3036722.
org/10.3390/MATH8091423. [32] J. Shi, X. Yuan, M. Elhoseny, X. Yuan, Weakly Supervised
[17] K. El Asnaoui, Design ensemble deep learning model for Deep Learning for Objects Detection from Images 8 (2020) 231–
pneumonia disease classification, Int. J. Multimed. Inf. 242, doi: 10.1007/978-3-030-45099-1_18.
Retr. 10 (1) (2021) 55–68, https://doi.org/10.1007/s13735-021- [33] D. Dansanan, R. Kumar, A. Bhattacharjee, D.J. Hemanth, D.
00204-7. Gupta, A. Khanna, O. Castillo, Early diagnosis of COVID-19-
[18] S. Goyal, R. Singh, Detection and classification of lung diseases affected patients based on X-ray and computed tomography
for pneumonia and Covid-19 using machine and deep learning images using deep learning algorithm, Soft Comput.
techniques, J. Ambient Intell. Humaniz. Comput. (0123456789) (0123456789) (2020), https://doi.org/10.1007/s00500-020-05275-
(2021), https://doi.org/10.1007/s12652-021-03464-7. y.
[19] S.K. Lakshmanaprabu, S.N. Mohanty, K. Shankar, N. [34] D. Ezzat, A.E. Hassanien, H.A. Ella, An optimized deep
Arunkumar, G. Ramirez, Optimal deep learning model for learning architecture for the diagnosis of COVID-19 disease
classification of lung cancer on CT images, Futur. Gener. based on gravitational search optimization, Appl. Soft Comput.
Comput. Syst. v92 (2019) 374–382, https://doi.org/10.1016/ 98 (2021), https://doi.org/10.1016/j.asoc.2020.106742 106742.
j.future.2018.10.009. [35] V. Ravi, H. Narasimhan, C. Chakraborty, T.D. Pham, Deep
[20] Q.Z. Song, L. Zhao, X.K. Luo, X.C. Dou, Using Deep Learning learning-based meta-classifier approach for COVID-19
for Classification of Lung Nodules on Computed Tomography classification using CT scan and chest X-ray images,
Images, J. Healthc. Eng. (2017), https://doi.org/10.1155/2017/ Multimed. Syst. (0123456789) (2021), https://doi.org/10.1007/
8314740. s00530-021-00826-1.
[21] G.A.P. Singh, P.K. Gupta, Performance analysis of various [36] Y. Pathak, P.K. Shukla, A. Tiwari, S. Stalin, S. Singh, Deep
machine learning-based approaches for detection and Transfer Learning Based Classification Model for COVID-19
classification of lung cancer in humans, Neural Comput. Appl. Disease, Irbm 1 (2020) 1–6, https://doi.org/10.1016/j.
31 (10) (2019) 6863–6877, https://doi.org/10.1007/s00521-018- irbm.2020.05.003.
3518-x. [37] C. Sitaula, M.B. Hossain, Attention-based VGG-16 model for
[22] N. Kalaivani, N. Manimaran, S. Sophia, D.D. Devi, Deep COVID-19 chest X-ray image classification, Appl. Intell. 51 (5)
Learning Based Lung Cancer Detection and Classification, IOP (2021) 2850–2863, https://doi.org/10.1007/s10489-020-02055-x.
Conf. Ser. Mater. Sci. Eng. 994 (1) (2020) 7731–7776, https:// [38] A. Gupta, Anjum, S. Gupta, R. Katarya, InstaCovNet-19: A
doi.org/10.1088/1757-899X/994/1/012026. deep learning classification model for the detection of COVID-
[23] J.A. ALzubi, B. Bharathikannan, S. Tanwar, R. Manikandan, 19 patients using Chest X-ray, Appl. Soft Comput. 99 (2021)
A. Khanna, C. Thaventhiran, Boosted neural network ensemble 106859, https://doi.org/10.1016/j.asoc.2020.106859.
classification for lung cancer disease diagnosis, Appl. Soft [39] S. Thakur, A. Kumar, X-ray and CT-scan-based automated
Comput. J. 80 (2019) 579–591, https://doi.org/10.1016/j. detection and classification of covid-19 using convolutional
asoc.2019.04.031. neural networks (CNN), Biomed. Signal Process. Control 69
[24] V. Perumal, V. Narayanan, S.J.S. Rajasekar, Detection of (2021), https://doi.org/10.1016/j.bspc.2021.102920 102920.
COVID-19 using CXR and CT images using Transfer Learning [40] K.F. Monowar, M.A.M. Hasan, J. Shin, Lung Opacity
and Haralick features, Appl. Intell. 51 (1) (2021) 341–358, Classification With Convolutional Neural Networks Using
https://doi.org/10.1007/S10489-020-01831-Z/FIGURES/11. Chest X-rays, 2020 11th International Conference on Electrical
[25] L. Li, Z. Xu, et al, Artificial Intelligence Distinguishes COVID- and Computer Engineering (ICECE), (2020), 169-172, doi:
19 from Community Acquired Pneumonia on Chest CT, 10.1109/ICECE51571.2020.9393135.
Radiology 296 (2) (2020) E65–E71, https://doi.org/10.1148/ [41] G. Latif, FTA. Lung Opacity Pneumonia Detection with
RADIOL.2020200905. Improved Residual Networks, J. Med. Biol. Eng. 41 (5) (2021)
[26] Y. Song, S. Zheng, L. Li, X. Zhang, X. Zhang, Z. Huang, J. 581–591.
Chen, R. Wang, H. Zhao, Y. Chong, J. Shen, Y. Zha, T. Yang, [42] BIMCV-COVID19, 2022, https://bimcv.cipf.es/bimcv-projects/
Deep learning Enables Accurate Diagnosis of Novel bimcv-covid19/#1590858128006-9e640421-6711 (accessed 5
Coronavirus (COVID-19) with CT images, IEEE/ACM Trans. February 2022).
Comput. Biol. Bioinforma. (2021), https://doi.org/10.1109/ [43] Shahin, Osama R., Hamoud H. Alshammari, Ahmed I. Taloba,
TCBB.2021.3065361. Rasha M. Abd El-Aziz, Machine Learning Approach for
A deep learning architecture for multi-class lung diseases classification 935

Autonomous Detection and Classification of COVID-19 Virus, diographs with and without a lung nodule: receiver operating
Comput. Electr. Eng. 101 (2022) 108055. char acteristic analysis of radiologists’ detection of pulmonar y
[44] Shahin, Osama R., Rasha M. Abd El-Aziz, Ahmed I. Taloba, nodules, Am. J. Roentgenol. 174 (2000) 71–74.
Detection and classification of Covid-19 in CT-lungs screening [55] S. Jaeger, A. Karargyris, S. Candemir, L. Folio, J. Siegelman, F.
using machine learning techniques, J. Interdisciplinary Math. 25 Callaghan, Z. Xue, K. Palaniappan, R.K. Singh, S. Antani, G.
(3) (2022) 791-813. Thoma, Y.X. Wang, P.X. Lu, C.J. McDonald, Automatic
[45] ml-workgroup covid-19-image-repository, 2022, Public https:// tuberculosis screening using chest radiographs, IEEE Trans.
github.com/ml-workgroup/covid-19-image-repository/ Med. Imaging 33 (2) (2014) 233–245, https://doi.org/10.1109/
tree/master/png, (accessed 5 February 2022). TMI.2013.2284099, PMID: 24108713.
[46] COVID-19 DATABASE, 2022, https://sirm.org/category/senza- [56] S. Candemir, S. Jaeger, K. Palaniappan, J.P. Musco, R.K.
categoria/covid-19/, (accessed 5 February 2022). Singh, Z. Xue, A. Karargyris, S. Antani, G. Thoma, C.J.
[47] https://eurorad.org, (accessed 5 February 2022). McDonald, Lung segmentation in chest radiographs using
[48] ieee8023 covid-chestxray-dataset, 2002 https://github.com/ anatomical atlases with nonrigid registration, IEEE Trans.
ieee8023/covid-chestxray-dataset, (accessed 5 February 2022). Med. Imaging 33 (2) (2014 Feb) 577–590, https://doi.org/
[49] COVID-19 Chest X-Ray Image Repository, 2002, 10.1109/TMI.2013.2290491, PMID: 24239990.
https://figshare.com/articles/COVID-19_Chest_X-Ray_Image_ [57] K. Simonyan, A. Zisserman, Very deep convolutional networks
Repository/12580328, (accessed 5 February 2022). for large-scale image recognition, arXiv (2014),
[50] Armiro COVID-CXNet, 2022, https://github.com/armiro/ arXiv:1409.1556.
COVID-CXNet, (accessed 5 February 2022). [58] E. Bisong, Building Ma chine Learning and Deep Learning
[51] RSNA Pneumonia Detection Challenge, Can you build an Models on GoogleCloud Platform, Springer, 2019.
algorithm that automatically detects potential pneumonia [59] A. Gulli, S. Pal, Deep Lear ning with Kera s, Packt Publishing
cases?, 2022, https://www.kaggle.com/c/rsna-pneumonia- Ltd, 2017.
detection-challenge/data, (accessed 5 February 2022). [60] D.P. Kingma, J. Ba, Adam: A Method for Stochastic
[52] Chest X-Ray Images (Pneumonia) 5,863 images, 2 categories, Optimization, 2014arXiv preprint ar Xiv:1412.6980.
2022, https://www.kaggle.comC/paultimothymooney/chest- [61] GitHub: A Deep Learning Architecture for Multi-Class Lung
xray-pneumonia, (accessed 5 February 2022). Diseases Classification using Chest X-ray Images: https://
[53] N.M. Elshennawy, D.M. Ibrahim, Deep-Pneumonia github.com/abunaif544/vgg.git, (accessed August 20, 2022).
Framework Using Deep Learning Models Based on Chest X- [62] M.A. Khan, F. Algarni, A Healthcare Monitoring System for
Ray Images, Diagnostics 10 (2020) 649. the Diagnosis of Heart Disease in the IoMT Cloud Environment
[54] J. Shiraishi, S. Ka tsuragawa, J. Ikezoe, T. Ma tsumoto, T. Using MSSO-ANFIS, IEEE Access 8 (2020) 122259–122269,
Kobayashi, K.-I. Komatsu, M. Ma tsui, H. Fujita, Y. Kodera, https://doi.org/10.1109/ACCESS.2020.3006424.
K. Doi, Development of a digital image database for chest ra