Hindawi

Computational Intelligence and Neuroscience

Volume 2022, Article ID 9263379, 12 pages
https://doi.org/10.1155/2022/9263379

Research Article
Tuberculosis Detection in Chest Radiographs Using Spotted Hyena
Algorithm Optimized Deep and Handcrafted Features

Seifedine Kadry ,1,2,3 Gautam Srivastava ,4,5 Venkatesan Rajinikanth ,6

Seungmin Rho ,7 and Yongsung Kim 8
1
Department of Applied Data Science, Noroff University College, Kristinasand 4612, Norway
2
Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon
3
Artificial Intelligence Research Center (AIRC), College of Engineering and Information Technology, Ajman University,
Ajman, UAE
4
Department of Mathematics and Computer Science, Brandon University, Brandon, R7A 6A9, Canada
5
Research Center for Interneural Computing, China Medical University, Taichung 40402, Taiwan
6
Department of Computer Science and Engineering, Saveetha School of Engineering,
Saveetha Institute of Medical and Technical Sciences, Chennai 602105, Tamil Nadu, India
7
Department of Industrial Security, Chung-Ang University, Seoul, Republic of Korea
8
Department of Technology Education, Chungnam National University, Daejeon, Republic of Korea

Correspondence should be addressed to Yongsung Kim; [email protected]

Received 17 February 2022; Revised 3 September 2022; Accepted 13 September 2022; Published 6 October 2022

Academic Editor: Andrea Loddo

Copyright © 2022 Seifedine Kadry et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Lung abnormality in humans is steadily increasing due to various causes, and early recognition and treatment are extensively suggested.
Tuberculosis (TB) is one of the lung diseases, and due to its occurrence rate and harshness, the World Health Organization (WHO) lists
TB among the top ten diseases which lead to death. The clinical level detection of TB is usually performed using bio-medical imaging
methods, and a chest X-ray is a commonly adopted imaging modality. This work aims to develop an automated procedure to detect TB
from X-ray images using VGG-UNet-supported joint segmentation and classification. The various phases of the proposed scheme
involved; (i) image collection and resizing, (ii) deep-features mining, (iii) segmentation of lung section, (iv) local-binary-pattern (LBP)
generation and feature extraction, (v) optimal feature selection using spotted hyena algorithm (SHA), (vi) serial feature concatenation,
and (vii) classification and validation. This research considered 3000 test images (1500 healthy and 1500 TB class) for the assessment, and
®
the proposed experiment is implemented using Matlab . This work implements the pretrained models to detect TB in X-rays with
improved accuracy, and this research helped achieve a classification accuracy of >99% with a fine-tree classifier.

1. Introduction diseases caused by Mycobacterium tuberculosis

(M. tuberculosis), and it can cause severe breathing problems
In the healthcare industry, there is a heavy diagnostic burden in human patients. Therefore, it is imperative to detect and
because of the steady increase in disease incidence in treat tuberculosis in a timely manner. It is also a commu-
humans due to various reasons. The burden of disease de- nicable illness that will affect a human quickly and easily if
tection can be reduced in hospitals by developing and one has a weak immune system.
implementing automated disease detection systems using A recent report by World Health Organization (WHO)
artificial intelligence (AI) [1–5]. lists TB as one of the top 10 causes of death globally and the
The lungs are one of the vital internal organs, and an foremost reason for death from a solitary infectious agent
infection in the lungs can cause severe illness, including [9]. This report also confirms that, in 2019, TB caused 1.4
death [6–8]. Tuberculosis (TB) is one of the severe lung million deaths worldwide, and this report estimated that ten
2 Computational Intelligence and Neuroscience

million people would be diagnosed with TB. Most people then identified using SHA, and then, a serial concatenation is
infected with TB (>t 90%) are adults, and the infection rate considered to combine these optimal features (DF + HF). This
in males is higher than in women. Increased TB rate in a feature vector is then considered to validate the performance of
country is due to poverty, which causes financial distress, the binary classifier with a 5-fold cross-validation, and the
susceptibility, marginalization, and bias in TB-infected employed scheme helped to achieve a classification accuracy of
people. Furthermore, this report also verifies that about a 98.73% with the fine-tree classifier.
quarter of the world’s population is infected with TB. The main contribution of this research includes the
Usually, TB is curable and preventable when diagnosed in its following:
early phase, and >85% who develop TB can be completely
(i) Execution of CNN-based joint segmentation and
recovered with a 6-month drug regimen [10, 11].
classification is implemented using VGG16
The clinical level diagnosis of TB is usually performed
with various clinical tests, including the bio-images. The (ii) LBP pattern generation with various weights is
infected lung section is typically recorded using computed presented
tomography (CT) and radiographs (X-ray). The recorded (iii) SHA-based feature selection and serial feature
image is then examined using a computer algorithm or by an concatenation is discussed
experienced doctor to identify the harshness of TB. The
Other sections are arranged as follows: Section 2 shows
former research on TB detection confirms that early diag-
earlier related work, Section 3 demonstrates methodology,
nosis is essential to reduce the disease burden; hence, the
and Sections 4 and 5 present the experimental outcome and
researchers suggest several automated diagnostic procedures
conclusion of this research.
[12, 13]. In literature, the detection of TB with chest X-ray is
widely discussed due to its clinical significance. Several
machine learning (ML) and deep learning (DL) procedures
2. Related Research
are developed and employed to assess chest X-ray pictures. Automated disease detection schemes are developed to reduce
The DL-supported scheme helps to achieve a better detection the diagnostic burden in hospitals, and most of these schemes
accuracy compared to the ML, and hence, the DL-supported also support the decision-making and treatment planning
TB detection is considered in this research. The proposed processes. In the literature, several ML and DL schemes are
research proposes a TB detection framework using the discussed to identify the TB from chest X-rays with the help of
pretrained DL scheme, which implements combined seg- benchmark and clinically collected images. Every procedure
mentation and classification to achieve better detection, as aims to get better detection accuracy. This section summarizes
discussed in [14]. The earlier work by Rahman et al. [14] chosen procedures employed to examine the X-ray, and the
implemented UNet for the segmentation and pretrained DL necessary information is presented in Table 1.
schemes for the classification. In the earlier work, the The research by Rahman et al. [14] employed a combined
performance of VGG16 is not discussed, and hence, this segmentation and thresholding concept to improve disease
research attempted the detection of TB using the VGG- detection performance. This work employed the proposed
UNet-based technique. The different stages of this frame- technique on 7000 images (3500 healthy and 3500 TB class)
work consist of (i) image collection and resizing, (ii) and presented a detailed examination using various pre-
implementation of pretrained VGG-UNet to segment the trained CNN methods in the literature. With an experi-
lung section from X-ray, (iii) collection of deep features mental investigation, the proposed work confirmed that
(DF), (iv) local-binary-pattern (LBP) generation using dif- joint segmentation and classification help to get a better
ferent weights and LBP feature extraction, (v) spotted hyena disease diagnosis. With this motivation, the proposed work
algorithm (SHA) based DF and HF reduction, (vi) gener- of this research also adopted the joint segmentation and
ating a new feature section with the serial concatenation of classification concept to examine the TB from the database
features, and (vii) binary classification and validation. provided by Rahman et al. [15]. In the earlier work, the
In this work, 3000 test images (1500 healthy and 1500 TB) VGG16 was not employed for the segmentation and clas-
are collected from the dataset provided by Rahman et al. sification task. Hence, the proposed research work adopted
[14, 15]. Initially, every test image is resized to 224 × 224 × 3 the VGG-UNet scheme for the investigation, in which the
pixels, and the converted images are then evaluated using the VGG16 acts as the encoder unit. The experimental outcome
pretrained VGG-UNet. UNet is a well-known convolutional of this study confirms that the proposed scheme worked well
neural network (CNN)-based encoder-decoder assembly, and on the chest X-ray database and helped to achieve a clas-
the enhancement of this scheme is already reported in the sification accuracy of >99% with the fine-tree classifier.
literature. The enhancement methods, such as VGG-UNet [16]
and ResNet [17], are already employed in which the encoder 3. Methodology
section is modified using the DL scheme. In the considered
VGG-UNet, the well-known VGG16 architecture is considered This research division shows the scheme developed to
to implement the encoder-decoder assembly, and the earlier identify the TB by joint segmentation and classification task.
work on this scheme can be accessed from [4]. In this work, the First, the necessary test pictures are collected from a
encoder section provides the necessary DF, and the decoder benchmark image database represented by Rahman et al.
section supplies the segmented lung section, which is then [15], and after the collection, every image is resized to a
considered to extract HF. The optimal value of DF and HF is dimension of 224 × 224 × 3 pixels. After the resizing task,
Computational Intelligence and Neuroscience 3

Table 1: Summary of automated TB detection schemes employed to examine X-ray images.

Reference Developed procedure
A customized DL system is proposed to examine the Shenzhen CXR pictures, and the proposed system provided
Rajaraman and Antani
an accuracy of 83.7%. However, this work confirms that implementing a customized DL approach is complex
[18]
and time-consuming
Examination of TB from X-ray using ensemble DL system and canny-edge detection is implemented and
achieved better values of accuracy (89.77%), sensitivity (90.91%), and specificity (88.64%). However, the
Hwa et al. [19]
implementation of canny-edge detection along with the ensemble DL scheme needs a larger image preprocessing
task, and it will increase the detection time
The development of a customized DL technique called TB-Net is proposed, and this work helped to achieve
Wong et al. [20] better performance measures, such as accuracy (99.86%), sensitivity (100%), and specificity (99.71%). This
research also proposes a customary model, which is relatively more complex than the pretrained models
Seven convolutional layers and three fully connected layer-based customized DL method are proposed for TB
Hooda et al. [21]
detection and achieved a classification accuracy of 94.73%
This work employed the conventional AlexNet and VGG16 methods to examine the X-ray images and attained
Rohilla et al. [22]
an accuracy of >81%
X-ray diagnosis performance of pretrained DL schemes is presented, and the employed technique helped to
Nguyen et al. [23]
provide better TB recognition
The contour-based silhouette descriptor technique is employed to detect TB, and the selected features provided
Afzali et al. [24]
an accuracy of 92.86%
The CNN-based disease diagnosis with lossless and lossy data expansion is employed, and the proposed method
Stirenko et al. [25]
offers a better TB diagnosis with X-ray pictures
Implementation of combined CNN segmentation and categorization is presented to identify TB from X-ray
images. This work implemented the classification task with and without segmentation and achieved a TB
Rahman et al. [14]
detection accuracy of 96.47% and 98.6%, respectively. This work also presented a detailed evaluation
methodology for TB detection using various pretrained DL methods

Encoder (VGG16) Decoder (UNet) Classifier

Conv (64)

Sigmoid
224×224×3

Conv (128)

Conv (256)

Conv (512)

Conv (512) 224×224×1

Deep-
Features Conv Batch
(1×4096) (512) Normalization

Healthy
LBP Feature Classifier
TB
Max pool

Up-Conv
Copy and Crop
Figure 1: Joint segmentation and classification implemented for TB detection using X-ray.
4 Computational Intelligence and Neuroscience

Table 2: Chest X-ray image dataset information.

Images
Class Dimension
Total Training Validation
Healthy 224 × 224 × 3 1500 1050 450
TB 224 × 224 × 3 1500 1050 450

every picture is evaluated by the VGG-UNet. Then, the are considered to form the encoder and decoder section,
encoder section presents the necessary DF, and the final which supports the feature extraction and segmentation for
layer (SoftMax) of the decoder section provides the binary medical images of a chosen dimension. In this work, the
form of the segmented lung section. The outcome of the pretrained VGG16 is then considered to implement the
encoder unit provides a DF of value, which is then reduced VGG-UNet scheme, and the necessary information about
by 1 × 1 × 1024 using a chosen dropout rate (2 dropout this architecture can be found in [4, 31].
layers with 50% dropout value to reduce 1 × 1 × 4096 to Initially, the considered VGG-UNet is trained using
1 × 1 × 1024), and these features are further reduced using X-ray images with the following tasks.
the SHA to get the DF of a chosen dimension. The binary
(i) Predictable augmentation (rotation and zoom) to
image obtained at the decoder section is then combined with
increase the number of training images
its original test image to extract the lung section. The
necessary LBP features are extracted from the extracted lung (ii) Assignment of learning rate as 1e-4 for better
section, and these features are further reduced with SHA. accuracy
Finally, a serial concatenation is then implemented to get (iii) Training with linear dropout rate (LDR) and Adam
DF + HF, and these features are then chosen to test and optimization
validate the performance of the developed system on the
During this task, other vital parameters are assigned as
considered database.
follows: total iteration � 2000, total epochs � 50, dropout
The performance of the proposed scheme is tested using
rates in the fully connected layer � 50%, and the final layer is
(i) DF alone and (ii) SHA-optimized DF + HF. During this
the SoftMax unit with 5-fold cross-validation.
assessment, the SoftMax-based binary classification is
employed, and later, other binary classifiers existing in the
literature are considered for testing the performance of the 3.3. Feature Extraction. This section presents the outline of
proposed scheme. The various stages presented in this the DF and HF extraction procedure.
scheme are shown in Figure 1. The concatenated feature is
employed in this work to classify the X-ray images into
healthy/TB classes. 3.3.1. Deep Feature. The necessary deep features from the
proposed scheme are extracted from the encoder section
(VGG16) of the VGG-UNet. This section offered a feature
3.1. Image Dataset. The merit of the automated disease vector of dimension 1 × 1 × 4096, and it is then passed
diagnosis is then tested and verified using the clinically grade through three fully connected (FC) layers with a dropout
or benchmark medical data. In this research, the chest X-ray rate of 50% to get a reduced feature vector of dimension
images considered by Rahman et al. [15] are adopted. From 1 × 1 × 1024. This feature is the DF, which is then considered
this dataset, 3000 images are collected to assess which 1500 to classify the X-ray images using a chosen binary classifiers.
images belong to the healthy group and the remaining 1500 In this work, the classification task is executed using the
with TB traces. Every collected image is resized to 224 × conventional DF and the DF optimized using the SHA. The
224 × 3 pixels (approved size for VGG16). Of the total experimental outcome of this study confirms that the
images, 70% (1050 images) are considered to train the de- proposed work helped to get better classification accuracy
veloped scheme, and the remaining 30% (450 images) are with optimized DF compared to the conventional DF.
considered for validation. The information about the test
images is shown in Table 2, and the sample test images for
the healthy/TB class are presented in Figure 2. 3.3.2. Handcrafted Feature. The HF is considered in ML-
based automatic disease detection systems, and in this work,
the HF is obtained using LBP of various weights as discussed
3.2. Pretrained VGG-UNet. Deep-learning-supported med- in [32]. The various procedures to extract the HF from the
ical data assessment is a commonly employed technique, and chosen X-ray are as follows: the implemented VGG-UNet
most of these approaches are adopted to implement auto- helps to extract the lung section in binary form. This binary
matic segmentation and classification operations [26–28]. image is then combined with its original test image to get the
The CNN-based segmentation using the traditional UNet necessary lung section without the artifacts. After getting the
[29] and SegNet [30] is employed in the literature to extract lung image, the necessary LBP pattern is generated by
and evaluate the disease-infected section from various assigning its weights as W � 1, 2, 3 an d 4, and from these
modality medical images. The limitation of traditional CNN images, the necessary LBP features with dimension 1 × 1 ×
segmentation schemes is rectified by enhancing its perfor- 59 are extracted, and the extracted features are then opti-
mance using the pretrained DL schemes. The DL schemes mized using the SHA.
Computational Intelligence and Neuroscience 5

Healthy
TB

Figure 2: Sample X-ray images of healthy/TB class.

(a) (b) (c)

Figure 3: The result achieved with VGG-Unet and the extracted lung section. (a) Test image, (b) binary image, (c) section to be examined.

Figure 3 depicts the procedure to remove the artifact. The earlier works related to medical image assessment
Figures 3(a) and 3(b) show the sample test image and the confirm that the MA is widely adopted in various image
extracted binary region by the decoder section of VGG- examination works, such as thresholding, segmentation, and
UNet. When Figures 3(a) and 3(b) are combined (pixel-wise feature selection [33, 34]. The MA-based feature selection
multiplication), then we get Figure 3(c), which is the lung procedure is already discussed in various ML and DL
section without the artifact. This section is then considered techniques, and this procedure helps to get the finest feature
to get the LBP pattern with various weights. The generated vector, which avoided the overfitting problem during the
LBP pattern of a sample image is presented in Figure 4, in automated classification. The MA-based feature selection
which Figures 4(a)–4(d) depict the outcome for the chosen can be used as an alternative technique for the traditional
weights. Every image provides 59 number of one-dimen- feature reduction procedures discussed in [35].
sional (1D) features, and the total features obtained with LBP In this work, the feature reduction task is imple-
are 236 features (59 × 4) which are then reduced using the mented for both the DF and HF using the SHA. It is a
SHA to avoid the overfitting difficulty in X-ray classification. nature-motivated procedure invented in 2017 by mim-
Other information related to this task can be found in the icking the hunting events found in spotted hyena (SH)
earlier research works [4, 16]. packs. The SH are the skillful animal that hunts as a pack,
and this operation consists of the following stages: (i)
choice making and following the prey, (ii) chasing the
3.4. Feature Reduction with Spotted Hyena Algorithm. prey, (iii) surrounding the prey, and (iv) killing. The
Metaheuristic algorithms (MA) are adopted in the literature arithmetical replica developed by Dhiman and Kumar
to find the finest solution for various real-world problems. [36, 37] considered all constraints to improve the
6 Computational Intelligence and Neuroscience

(a) (b) (c) (d)

Figure 4: Generated LBP pattern of X-ray for various weights. (a) W � 1, (b) W � 2, (c) W � 3, (d) W � 4.

(A*-A, B) (A*, B)
(A, B)

(A*-A)
(B*-B)

(A*, B*)
(A*-A, B*)
(A, B*)
Target Target

(A*-A, B*-B) (A*, B*-B) (A, B*-B)

(a) (b)

Target

(c)

Figure 5: Various stages of SHA. (a) Encircling, (b) hunting, (c) attacking.

converge capability of the SHA. A similar kind of algo- using notation A and B. This adjustment is carried out in the
rithm, known as the Dingo optimizer, is also developed algorithm using mathematical operations such as multi-
and implemented by Bairwa et al. [38]. plication and subtraction.
The various stages of the SHA are depicted in Figure 5, in The encircling process is mathematically represented as
which Figures 5(a)–5(c) present the operations, such as follows:
identifying and tracking the prey as in Figure 5(a), tracking → 􏼌􏼌􏼌→ → → 􏼌􏼌
and encircling the prey depicted in Figure 5(b), and hunting D h � 􏼌􏼌 B . P p (x) − P (x)􏼌􏼌􏼌, (1)
as presented in Figure 5(c). This operation is as follows: the
leader in pack identifies the prey, and the leader and its pack → → →→
P (x + 1) � P p (x) − E .D h , (2)
will chase it till it is tired. When the prey is tired, the leader →
and its group will encircle the prey as depicted in Figure 5. In where D h →� distance among the hyena and prey,→x � current
this context, every group member will adjust their location iteration, P p � position vector of prey, and P � position
concerning the prey. This process is depicted in the figure vector of hyena.
Computational Intelligence and Neuroscience 7

→ →
The coefficient vectors B and E are computed as Healthy class image features
follows:
→ �→
B � 2.R d1 , (3)
Optimized
→ → �→ → Spotted Hyena Algorithm Deep/LBP features
E � 2 h .R d2 − h , (4)

→ 5
h � 5 − 􏼠Iteration ∗ 􏼠 􏼡􏼡, (5) TB class image features
Itermax
→ Figure 6: The SHA-based feature selection process.
where Itermax � maximum iterations assigned, h � a�→ linearly
decreasing
�→ value from 5 to 0 insteps of 0.1, R d 1 and LBP2w2(1×1×59) � W2(1,1) , W2(1,2) , · · · , W2(1,59), (12)
R d2 � random number [0, 1] number
In this figure, (A, B) are the hyena, and it will adjust its
location towards the prey (A ∗ , B ∗ ) based on the values of LBP3w3(1×1×59) � W3(1,1) , W3(1,2) , · · · , W3(1,59) , (13)
Eqns. (3) to (5).
In the hunting stage, the hyena pack will move close to LBP4w4(1×1×59) � W4(1,1) , W4(1,2) , · · · , W4(1,59) , (14)
the prey and proceed for the attack. This phase is represented
as follows: HF(1×1×236) � LBP1 + LBP2 + LBP3 + LBP4. (15)
→ 􏼌􏼌􏼌→ → �→􏼌􏼌􏼌
D h � 􏼌􏼌 B . P h − Pk 􏼌􏼌, (6)
During the feature reduction process, the DF in Eqn. (10)
→ → →→ and HF in (8) are examined by the SHA, and the reduced
P k � P h − E .D h , (7) features are then serially concatenated to get a hybrid feature
vector (DF + HF). In this work, the SHA helps to reduce the
→ �→ ���→ ����→ DF to 1 × 1 × 427and HF to 1 × 1 × 166, and these features
C h � Pk + Pk+1 + · · · + Pk+N , (8)
→ are then combined as in Eqn. (16) to get the new feature
where
�→ P h � leader which moves closer to prey and vector.
Pk � positions of other hyenas in the pack, and N � total
hyenas in the pack. DF + HF(1×1×593) � DF(1×1×427) + HF(1×1×166) . (16)
In the attacking phase, the hyena moves and attacks the
The feature presented in (9) is then considered to train
prey, other hyenas in the group also follow the same
and test the classifiers considered in this study. The various
technique, and the group attach will kill the prey. When the
binary classifiers considered in this research include Soft-
prey is dead, every hyena in the pack is on or nearer to the
Max, Naı̈ve-Bayes (NB), random forest (RF), decision tree
prey. This process is the convergence of the chosen agents
(DT) variants, K-nearest neighbors (KNN) variants, and
towards the optimal location as in (7).
�→ SVM with linear kernel [40–43].
→ C (9)
P (x + 1) � h ,
N
→
where P (x + 1) is the best position, in which every hyena of 3.5. Performance Validation. The merit of an automated
the pack converges. In this work, the SHA is initiated with disease detection system is to be verified by computing the
the following parameters: number of hyena (agent) � N � 20, necessary performance values. In this work, the measures
search dimension (D) � 2, Itermax � 2000 and stopping cri- obtained from the confusion matrix are considered to
teria � maximization of Cartesian distance (CD) between confirm the eminence of the proposed scheme. These
features or Itermax . measures include true positive (TP), false negative (FN), true
The feature reduction with SHA is graphically depicted negative (TN), false positive (FP), accuracy (ACC), precision
in Figure 6, and from this procedure, it is clear that this task (PRE), sensitivity (SEN), specificity (SPE), and negative
selects the image features of the healthy/TB class based on predictive value (NPV). The mathematical expressions of
the maximal CD, and this procedure is already discussed in these values are presented in the following equations
earlier works [5, 34, 39]. As depicted in Figure 6, the SHA- [42–46]:
based feature selection is separately employed to identify the TP + TN
optimal features, and after getting these features, it is then ACC � , (17)
TP + TN + FP + FN,
serially concatenated, and the concatenated features are then
considered for classifier training and validation. TP
The number of DF and HF available for the optimization PRE � , (18)
TP + FP
is depicted in the following equations:
DFVGG(1×1×1024) � VGG(1,1) , VGG(1,2) , · · · , VGG(1,1024) , TP
SEN � , (19)
TP + FN
(10)
TN
LBP1w1(1×1×59) � W1(1,1) , W1(1,2) , · · · , W1(1,59) . (11) SPE � , (20)
TN + FP
8 Computational Intelligence and Neuroscience

(a) (b)

Figure 7: Sample test images obtained the convolution operation. (a) Normal, (b) TB.

Table 3: Classification result achieved with DF alone for a 5-fold cross-validation.

Classifier TP FN TN FP ACC (%) PRE (%) SEN (%) SPE (%) NPV (%)
Fold 1 418 32 414 36 92.44 92.07 92.89 92.00 92.82
Fold 2 421 29 419 31 93.33 93.14 93.56 93.11 93.53
Fold 3 420 30 424 26 93.78 94.17 93.33 94.22 93.39
Fold 4 423 27 425 25 94.22 94.42 94.00 94.44 94.03
Fold 5 421 29 423 27 93.78 93.97 93.56 94.00 93.58

SoftMax SoftMax

423 25 94.4% 441 8 98.2%

Healthy Healthy
47.0% 2.8% 5.6% 49.0% 0.9% 1.8%
Output Class

Output Class

27 425 94.0% 9 442 98.0%

TB TB
3.0% 47.2% 6.0% 1.0% 49.1% 2.0%

94.0% 94.4% 94.2% 98.0% 98.2% 98.1%

6.0% 5.6% 5.8% 2.0% 1.8% 1.9%

Healthy TB Healthy TB
Target Class Target Class
(a) (b)

Figure 8: Confusion matrix attained with traditional and optimized features. (a) DF with SoftMax, (b) DF + HF with SoftMax.

TN workstation; Intel i7 2.9 GHz processor with 20 GB RAMS

NPV � . (21)
TN + FN and 4 GB VRAM equipped with Matlab .
Initially, the pretrained VGG-UNet scheme is trained
®
using resized chest X-ray images till it extracts the lung
4. Results and Discussion section with better accuracy. After the training, its seg-
mentation performance is tested using test images, and the
This part of the research presents the present research’s outcome is recorded. Then, the extracted section is com-
investigational outcome. This work is executed using a bined with the original image to extract the lung section
Computational Intelligence and Neuroscience 9

100
90
80
70
Accuracy (%)

60
50
40
30
20
10
0
0 50 100 150 200 250 300 350
Iteration

3
2.5
2
Loss

1.5
1
0.5
0
0 50 100 150 200 250 300 350
Iteration
(a)
Fine-Tree 1

0.95
Trye Positive Rate (Sensitivity)

446 3 99.3%
Healthy 49.6% 0.3% 0.7%
0.9

0.85
Output class

4 447 99.1%
TB 0.4% 49.7% 0.9% 0.8
AUC=0.982
0.75 P<0.001
99.1% 99.3% 99.2% 0.7
0.9% 0.7% 0.8%
0.65
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Healthy TB False Positive Rate (1 - specificity)
Target class
(b) (c)

Figure 9: Investigational outcome of the fine-tree classifier with DF + HF. (a) Convergence, (b) confusion matrix, (c) ROC.

without the artifact, and the necessary LBP is generated and 7(b) depict the hot color map image obtained for healthy
when the HF is extracted. Similarly, the necessary DF is and TB class sample images, respectively. After extracting
extracted from the encoder section (VGG16), and these the necessary deep features from the test images (with a
features are then passed through the fully connected layers to VGG16-like scheme), the necessary classification task is
get a feature vector of size 1 × 1 × 1024. This procedure is implemented using the SoftMax classifier with a 5-fold
similar to employing a traditional VGG16 scheme, and this cross-validation. The achieved results are presented in Ta-
feature is initially considered to classify the images with a ble 3. This table confirms that when the DF vector of di-
SoftMax classifier, and the necessary performance is then mension 1 × 1 × 1024 is considered, SoftMax provided a
recorded. classification accuracy of 94.22%. This procedure is then
During the convolutional operation, the layers of the repeated using the SHA-selected DF + HF presented in (9),
VGG-UNet help to recognize the necessary image features to and the achieved confusion matrix (CM) is presented in
support the necessary feature extraction and segmentation. Figure 8. Figure 8(a) shows the CM of the DF case, and
The sample test image textures identified during a con- Figure 8(b) shows the CM of optimized DF + HF, and this
volutional operation are presented in Figure 7. Figures 7(a) confirms that the accuracy achieved with the proposed
10 Computational Intelligence and Neuroscience

Table 4: Sample results attained with the implemented UNet scheme.

Classifier TP FN TN FP ACC (%) PRE (%) SEN (%) SPE (%) NPV (%)
SoftMax 441 9 442 8 98.11 98.22 98.00 98.22 98.00
NB 442 8 443 7 98.33 98.44 98.22 98.44 98.23
RF 441 9 443 7 98.22 98.44 98.00 98.44 98.01
Coarse tree 442 8 442 8 98.22 98.22 98.22 98.22 98.22
Medium tree 442 8 444 6 98.44 98.66 98.22 98.67 98.23
Fine tree 446 4 447 3 99.22 99.33 99.11 99.33 99.11
Coarse KNN 445 5 446 4 99.00 99.11 98.89 99.11 98.89
Medium KNN 444 6 445 5 98.78 98.88 98.67 98.89 98.66
Fine KNN 447 3 442 8 98.78 98.24 99.33 98.22 99.32
SVM linear 441 9 445 5 98.44 98.88 98.00 98.89 98.02

SoftMax NB RF Coarse-Tree
VGG19 ResNet18 ResNet50 ResNet101

Medium-Tree Fine-Tree Coarse KNN Medium KNN ChexNet Inception-V3 DenseNet201 SquuzeNet

Fine KNN SVM-Linear MobileNet Proposed

(a) (b)

Figure 10: Overall performance of TB detection system demonstrated with glyph plot. (a) Results with DF + HF, (b) comparison with earlier
work.

method is superior to the traditional technique. Hence, the comparison of the fine-tree classifier with earlier works is
performance of the proposed scheme is then confirmed with presented in Figure 10(b), which confirms its superiority
various binary classifiers using the DF + HF. over other classifiers. Figure 10(b) compares the ACC. PRE,
Figure 9 presents the experimental outcome achieved SEN, and SPE of the earlier research by Rahman et al. [14]
with the RF variant and the fine-tree classifier. Figure 9(a) with the fine-tree result, and this comparison confirms that
depicts the convergence of the search, and Figures 9(b) and the proposed system’s outcome is better.
9(c) show the CM and the receiver operating characteristic This research implemented a joint segmentation and
(ROC) curve, respectively. The results achieved with other classification scheme to detect TB from chest X-rays with
chosen classifies are presented in Table 4. This confirms that better accuracy. The main limitation of the proposed scheme
the classification accuracy of the fine tree is >99%, which is that it considered the artifact-removed image for getting
confirms its merit over other techniques. In order to verify the necessary HF from the LBP images. In the future, the
the performance of the proposed scheme, its best result is LBP can be combined with other HF existing in the liter-
compared with the results of Rahman et al. [14] and con- ature. Furthermore, the performance of the proposed
firmed that the proposed joint segmentation and classifi- scheme can be tested and verified with other benchmark
cation scheme with SHA-selected DF + HF help to achieve a chest X-ray images with various lung abnormalities.
better outcome compared to the earlier works.
Table 4 confirms that the result of the fine tree is better 5. Conclusion
than other binary classifiers, and the coarse KNN also helped
to achieve a classification accuracy of 90% compared to other In humans, TB is a severe disease that widely affects the
techniques. This confirmed that the optimized DF + HF lungs, and early diagnosis and treatment will help to reduce
supported classification helps to get a better overall result, as the severity. Furthermore, the timely detection and rec-
presented in Figure 10. Figure 10(a) shows the glyph plot for ommended medication will help to cure the TB completely.
the overall performance of binary classifiers, and the pattern Due to its significance, a considerable number of research
covering a maximum area is considered superior. The works are performed by researchers to support the
Computational Intelligence and Neuroscience 11

automated diagnosis. This research aims to develop and classification of brain tumor in clinical MRI slices,” Infor-
implement a joint segmentation and classification scheme mation Technology and Control, vol. 50, no. 2, pp. 342–356,
with the help of a pretrained VGG-UNet scheme. The VGG- 2021.
UNet system consists of an encoder-decoder assembly, in [5] Z. F. khan and S. R. Alotaibi, “Applications of artificial in-
which the encoder helps to get the necessary DL features as telligence and big data analytics in m-health: a healthcare
system perspective,” Journal of Healthcare Engineering,
in the traditional VGG16 system, and the decoder associated
pp. 1–15, 2020.
with the SoftMax classifier helps to extract the binary form of [6] S. Vj and J. F. D, “Deep learning algorithm for COVID-19
the lung image. This work considered the LBP pattern of the classification using chest X-ray images,” Computational and
lung image to extract the necessary HF. This work con- Mathematical Methods in Medicine, vol. 2021, no. 10, Article
sidered the LBP with varied weights and helped to get a1D ID 9269173, 2021.
feature vector of size 236. The extracted DF and the HF are [7] T. Akram, M. Attique, S. Gul et al., “A novel framework for
then optimized using the SHA, and these features are then rapid diagnosis of COVID-19 on computed tomography
serially united to get the concatenated features vector scans,” Pattern Analysis & Applications, vol. 24, no. 3,
(DF + HF). This feature vector is then considered for testing pp. 951–964, 2021.
and validating the performance of the binary classifiers using [8] D. Singh, V. Kumar, and M. Kaur, “Densely connected
5-fold cross-validation. The experimental outcome of this convolutional networks-based COVID-19 screening model,”
study confirmed that the binary classification with the fine- Applied Intelligence, vol. 51, no. 5, pp. 3044–3051, 2021.
[9] https://www.who.int/health-topics/tuberculosis#tab=tab_1.
tree classifier helped to achieve an accuracy of >99% for the
[10] G. Who, Global tuberculosis report, Glob. Tuberc. Rep, 2020.
considered chest X-ray images. This result is then compared [11] E. Harding, “WHO global progress report on tuberculosis
and validated with the result of other DL methods available elimination,” The Lancet Respiratory Medicine, vol. 8, no. 1,
in the literature. This research confirmed the merit of the pp. 19–30, 2020.
proposed DF + HF-based TB detection from the chest X-ray [12] A. H. Al-Timemy, R. N. Khushaba, Z. M. Mosaand, and
images. In the future, this scheme can be enhanced with J. Escudero, “An efficient mixture of deep and machine
other HF available in the literature. Furthermore, the per- learning models for covid-19 and tuberculosis detection using
formance of the proposed scheme can be tested and vali- x-ray images in resource limited settings,” Artificial Intelli-
dated with other chest X-ray image datasets available in the gence for COVID-19, pp. 77–100, 2021.
literature. [13] U. Z. M. Azmi, N. A. Yusof, J. Abdullah, S. A. A. Ahmad, and
F. N. M. Faudziet al, “Portable electrochemical immunosensor
for detection of Mycobacterium tuberculosis secreted protein
Data Availability CFP10-ESAT6 in clinical sputum samples,” Microchimica
The X-ray images considered in this research work can be Acta, vol. 188, no. 1, pp. 1–11, 2021.
[14] T. Rahman, A. Khandakar, M. A. Kadir et al., “Reliable tu-
accessed from https://ieee-dataport.org/documents/
berculosis detection using chest X-ray with deep learning,
tuberculosis-tb-chest-x-ray-database. segmentation and visualization,” IEEE Access, vol. 8, no. 2,
pp. 191586–191601, 2020.
Conflicts of Interest [15] https://ieee-dataport.org/documents/tuberculosis-tb-chest-x-
ray-database.
The authors declare that they have no conflicts of interest to [16] V. Rajinikanth and S. Kadry, “Development of a framework
report regarding the present study. for preserving the disease-evidence-information to support
efficient disease diagnosis,” International Journal of Data
Acknowledgments Warehousing and Mining, vol. 17, no. 2, pp. 63–84, 2021.
[17] F. I. Diakogiannis, F. Waldner, P. Caccetta, and C. Wu,
This work was supported in part by the National Research “ResUNet-a: a deep learning framework for semantic seg-
Foundation of Korea (NRF) grant funded by the Korean mentation of remotely sensed data,” ISPRS Journal of Pho-
government (MSIT) (No. 2020R1G1A1099559). togrammetry and Remote Sensing, vol. 162, no. 2, pp. 94–114,
2020.
References [18] S. Rajaraman and S. K. Antani, “Modality-specific deep
learning model ensembles toward improving TB detection in
[1] V. Kumar, D. Singh, M. Damaševičius, and R. Damaševičius, chest radiographs,” IEEE Access, vol. 8, no. 2,
“Overview of current state of research on the application of pp. 27318–27326, 2020.
artificial intelligence techniques for COVID-19,” PeerJ [19] M. H. A. Hijazi, S. Kieu Tao Hwa, A. Bade, R. Yaakob, and
Computer Science, vol. 7, no. 1, pp. 5644–e634, 2021. M. Saffree Jeffree, “Ensemble deep learning for tuberculosis
[2] V. Chouhan, S. K. Singh, A. Khamparia et al., “A novel detection using chest X-Ray and canny edge detected images,”
transfer learning based approach for pneumonia detection in IAES International Journal of Artificial Intelligence, vol. 8,
chest X-ray images,” Applied Sciences, vol. 10, no. 2, no. 4, pp. 429–442, 2019.
pp. 559–617, 2020. [20] A. Wong, J. R. H. Lee, H. Rahmat-Khah, and A. S. A. Alaref,
[3] S. J. Priya, A. J. Rani, M. S. P. Subathra, M. A. Mohammed, “TB-Net: a tailored, self-attention deep convolutional neural
and N. Damaševičius, “pattern transformation based feature network design for detection of tuberculosis cases from chest
extraction for recognition of Parkinson’s disease based on gait X-ray images,” 2021, https://arxiv.org/abs/2104.03165.
signals,” Diagnostics, vol. 11, no. 8, pp. 1395–1417, 2021. [21] R. Hooda, S. Sofat, S. Kaur, A. Mittal, and F. Meriaudeau,
[4] V. Rajinikanth, S. Kadry, and Y. Nam, “Convolutional- “Deep-learning: a potential method for tuberculosis detection
Neural-Network assisted segmentation and SVM using chest radiography,” in Proceedings of the 2017 IEEE
12 Computational Intelligence and Neuroscience

International Conference on Signal and Image Processing [35] K. H. Cheong, K. J. W. Tang, X. Zhao et al., “An automated
Applications (ICSIPA), pp. 497–502, IEEE, Kuching, Malaysia, skin melanoma detection system with melanoma-index based
September 2017. on entropy features,” Biocybernetics and Biomedical Engi-
[22] A. Rohilla, R. Hooda, and A. Mittal, “TB detection in chest neering, vol. 41, no. 3, pp. 997–1012, 2021.
radiograph using deep learning architecture,” ICETETSM, [36] G. Dhiman and V. Kumar, “Spotted hyena optimizer: a novel
vol. 17, pp. 136–147, 2017. bio-inspired based metaheuristic technique for engineering
[23] Q. H. Nguyen, B. P. Nguyen, S. D. Dao, B. Unnikrishnan, and applications,” Advances in Engineering Software, vol. 114,
R. Dhingraet al, “Deep learning models for tuberculosis de- no. 2, pp. 48–70, 2017.
tection from chest X-ray images,” in Proceedings of the 2019 [37] G. Dhiman and V. Kumar, “Spotted hyena optimizer for
26th International Conference on Telecommunications (ICT), solving complex and non-linear constrained engineering
pp. 381–385, IEEE, Hanoi, Vietnam, April 2019. problems,” in Harmony Search and Nature Inspired Opti-
[24] A. Afzali, F. B. Mofrad, and M. Pouladian, “Feature selection mization Algorithms, pp. 857–867, Springer, Singapore, 2019.
for contour-based tuberculosis detection from chest x-ray [38] A. K. Bairwa, S. Joshi, and D. Singh, “Dingo optimizer: a
images,” in Proceedings of the 2019 26th National and 4th nature-inspired metaheuristic approach for engineering
International Iranian Conference on Biomedical Engineering problems,” Mathematical Problems in Engineering, vol. 2021,
(ICBME), pp. 194–198, IEEE, Tehran, Iran, November 2019. no. 2, pp. 1–12, 2021.
[25] S. Stirenko, Y. Kochura, O. Alienin, O. Rokovyi, and [39] V. Rajinikanth and N. Dey, Magnetic Resonance Imaging:
Y. Gordienkoet al, “Chest X-ray analysis of tuberculosis by Recording, Reconstruction and Assessment, Elsevier, Amster-
deep learning with segmentation and augmentation,” in dam, Netherlands, 2022.
Proceedings of the 2018 IEEE 38th International Conference on [40] J. Ramya, M. P. Rajakumar, and B. U. Maheswari, “Deep CNN
Electronics and Nanotechnology (ELNANO), pp. 422–428, with hybrid binary local search and particle swarm optimizer
IEEE, Kyiv, UKraine, April 2018. for exudates classification from fundus images,” Journal of
[26] V. Rajinikanth, A. N. Joseph Raj, K. P. Thanaraj, and Digital Imaging, vol. 35, pp. 56–67, 2022.
G. R. Naik, “A customized VGG19 network with concate- [41] M. Ahmed, M. Ramzan, H. Ullah Khan et al., “Real-Time
nation of deep and handcrafted features for brain tumor violent action recognition using key frames extraction and
detection,” Applied Sciences, vol. 10, no. 10, pp. 3429–3439, deep learning,” Computers, Materials & Continua, vol. 69,
2020. no. 2, pp. 2217–2230, 2021.
[27] L. K. Ramasamy, S. G. Padinjappurathu, S. Kadry, and [42] M. P. Rajakumar, R. Sonia, B. Uma Maheswari, and
R. Damaševičius, “Detection of diabetic retinopathy using a S. P. Karuppiah, “Tuberculosis detection in chest X-ray using
fusion of textural and ridgelet features of retinal images and Mayfly-algorithm optimized dual-deep-learning features,”
sequential minimal optimization classifier,” PeerJComputer Journal of X-Ray Science and Technology, vol. 29, no. 6,
Science, vol. 7, no. 1, pp. 4566–e517, 2021. pp. 961–974, 2021.
[28] M. Wieczorek, J. Siłka, D. Połap, M. Woźniak, and [43] S. Rajesh Kannan, J. Sivakumar, and P. Ezhilarasi, “Automatic
R. Damasevicius, “Real-time neural network-based predictor detection of COVID-19 in chest radiographs using serially
for cov19 virus spread,” PLoS One, vol. 15, no. 12, p. e243217, concatenated deep and handcrafted features,” Journal of
Article ID 02431899, 2020. X-Ray Science and Technology, vol. 30, no. 2, pp. 231–244,
[29] O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolu- 2022.
tional networks for biomedical image segmentation,” in [44] M. Hmoud Al-Adhaileh, E. Mohammed Senan, W. Alsaade
Proceedings of the International Conference on Medical Image et al., “Deep Learning Algorithms for Detection and Classi-
Computing and Computer-Assisted Intervention, Springer, fication of Gastrointestinal Diseases,” Complexity, vol. 2021,
Cham, pp. 234–241, November 2015. Article ID 6170416, 2021.
[30] V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: a deep [45] T. Wen, Y. Du, T. Pan, C. Huang, and Z. Zhang, “A deep
convolutional encoder-decoder architecture for image seg- learning-based classification method for different frequency
mentation,” IEEE Transactions on Pattern Analysis and EEG data,” Computational and Mathematical Methods in
Machine Intelligence, vol. 39, no. 12, pp. 2481–2495, 2017. Medicine, vol. 2021, pp. 1–13, 2021.
[31] S. Kadry, R. Damaševičius, D. Taniar, V. Rajinikanth, and [46] S. Sharma, S. Gupta, D. Gupta et al., “Deep Learning Model
I. A. Lawal, “U-net supported segmentation of ischemic- for the Automatic Classification of White Blood Cells,”
Computational Intelligence and Neuroscience, vol. 2022, Ar-
stroke-lesion from brain MRI slices,” in Proceedings of the
ticle ID 7384131, 2022.
2021 Seventh International Conference on Bio Signals, Images,
and Instrumentation (ICBSII), pp. 1–5, IEEE, Chennai, India,
March 2021.
[32] A. Gudigar, U. Raghavendra, T. Devasia et al., “Global
weighted LBP based entropy features for the assessment of
pulmonary hypertension,” Pattern Recognition Letters,
vol. 125, no. 2, pp. 35–41, 2019.
[33] L. Wei, S. X. Pan, Y. A. Nanehkaran, and V. Rajinikanth, “An
optimized method for skin cancer diagnosis using modified
thermal exchange optimization algorithm,” Computational
and Mathematical Methods in Medicine, vol. 2021, no. 2,
pp. 1–11, 2021.
[34] Y. Y. Wang, H. Zhang, C. H. Qiu, and S. R. Xia, “A Novel
Feature Selection Method Based on Extreme Learning Ma-
chine and Fractional-Order Darwinian PSO,” Computational
intelligence and neuroscience, vol. 2018, pp. 1–8, 2018.
Copyright of Computational Intelligence & Neuroscience is the property of Hindawi Limited
and its content may not be copied or emailed to multiple sites or posted to a listserv without
the copyright holder's express written permission. However, users may print, download, or
email articles for individual use.