1. Introduction
In recent years, the incidence rate of Infectious Disease (ID) has been gradually increasing in humans, and timely detection and treatment are essential to cure and control the spread of the disease. The ID usually affects people with a weak immune system, and this disease is commonly transmitted from one individual to another. Thus, it is regarded as a communicable disease as well. Furthermore, the ID which infects the inner body organ is harsher in comparison to the ID which infects the eternal body organ [
1,
2,
3,
4].
In hospitals, a unique diagnostic and treatment procedure is employed to cure/control the disease in internal organs infected with ID. As part of human physiology, the lung exchanges air between the outer atmosphere and other body parts, and any infection in the lung will affect this exchange. There is evidence that the lungs are commonly affected by infections such as pneumonia, tuberculosis, and COVID-19 [
5,
6,
7]. Several diagnostic procedures exist to detect pneumonia with the help of medical images or chosen methods. Pneumonia is an infectious disease commonly affecting people with lower immunity (children and older adults). COVID-19 is also a pathogen that causes severe pneumonia in humans, and many medication procedures have been recommended to reduce the severity of COVID-19 infection in recent years [
8,
9,
10].
As a result of TB infection, people will experience mild to severe breathing problems, and if TB is not detected, it can result in death. A common cause of tuberculosis is Mycobacterium tuberculosis (M. tuberculosis), a bacterium responsible for infecting the lungs and other soft tissues. As TB infection in the lung is a common illness, the clinical level diagnosis is performed using biomedical images of the chosen modality. A computed tomography (CT) or chest radiograph can be used to examine the lung section. X-ray images have been used in medical clinics to examine tuberculosis more frequently than CT images. Since tuberculosis is a communicable disease, and can spread quickly from one person to another, timely detection and treatment are essential [
5].
As stated by the World Health Organization (WHO), tuberculosis is a serious disease and a significant cause of death worldwide. TB is considered among the top 10 deadliest diseases because the mortality rate is higher than HIV/AIDS. According to the WHO statement, 1.4 million people worldwide died of tuberculosis in 2019, while ten million cases of tuberculosis have been estimated for 2020. It is estimated that approximately 90% of the community’s adults are infected with tuberculosis (TB), with men contracting the disease at a significantly higher rate than women. Further, TB infection rates are lower in high-income countries than in low-income countries. TB will result in financial suffering, weakness, marginalization, and bias among TB-affected citizens. TB that is detected in a timely fashion is entirely curable, and >85% of those who develop TB can recover with a six-month prescription regimen [
11,
12,
13].
Various clinical tests, including a bio-image-based examination, are used to diagnose TB at the clinical level. Using the bio-images (CT/X-ray) analysis, an experienced pulmonologist can determine the severity of the lung infection caused by tuberculosis. In addition, CT/X-ray-supported tuberculosis detection helps the pulmonologist confirm the disease’s severity, decision-making, and treatment execution processes.
There are several automated disease detection schemes proposed and employed to evaluate CT/X-rays to reduce the diagnostic burden on pulmonologists. Furthermore, modern hospitals utilize standard procedures to assist physicians during the diagnosis and treatment of diseases, including segmentation, Machine-Learning-Scheme (MLS), and Deep-Learning-Scheme (DLS) [
14,
15,
16,
17].
This research aims to develop an automated TB detection system using optimally selected Machine-Learning-Features (MLF) and Deep-Learning-Features (DLF) to examine chest X-ray images with better diagnostic accuracy. The proposed scheme consists of the following phases: (i) image pre-processing, (ii) implementing a pre-trained DLS and mining of the DLF, (iii) extracting the necessary MLF using Local-Binary-Pattern (LBP) and Discrete-Wavelet-Transform (DWT), (iv) optimization of DLF and MLF using Seagull-Algorithm (SA), (v) serial concatenation of DLF and MLF, and (vi) binary classifier implementation and validation.
In this work, 7000 test images are collected for assessment from the database supplied by Rahman et al. [
18,
19]. Initially, all these images are resized to
pixels, and the converted images are then evaluated using the pre-trained DLS existing in the literature. For the AlexNet, the chest X-ray images are resized to
pixels. Initially, the conventional DLF obtained from every DLS is considered to classify the chest X-ray images using the SoftMax classifier, and the performance is verified. This investigation confirms that the VGG19 helped improve classification accuracy (92.9048%). Then, this DLF is optimized using the SA and combined with the SA-optimized MLF (DLF+MLF). These concatenated features are once again considered for the classification task, and the binary classification obtained with SVM-Medium Gaussian (SVM-MG) classifier offered an accuracy of ≈98.62%. This result is then compared and validated with the earlier result by Rahman et al. [
18].
The chief involvement of this research comprises.
- (i)
Implementation of pre-trained DLS-based TB detection from chest X-ray;
- (ii)
Generation of MLF using LBP and DWT;
- (iii)
SA-based feature optimization and serial feature concatenation to obtain DLF+MLF.
3. Methodology
This section presents essential information about the developed procedure to examine the chest X-ray and detect the normal/TB class images.
In this work, the necessary images are collected from the dataset of Rahman et al. [
18] and each picture are resized to a dimension of
pixels.
Figure 1 depicts the various stages employed in the proposed scheme. The test images are initially processed with VGG19 to obtain the necessary deep features of dimension
. These features are then reduced using the SA and the optimized DLF are then considered for the examination. The test images are then considered to extract the necessary MLF based on the LBP and DWT and are then combined (LBP+DWT) to obtain a single 1D feature vector and these features are then reduced using the SA to obtain the optimized MLF. The optimal values of DLF and MLF are then serially concatenated to obtain the hybrid feature vector, which is then considered to train and validate the binary classifiers considered to categorize the X-ray pictures. In this work, a 10-fold cross validation is employed to obtain better classification accuracy.
3.1. Image Dataset
The worth of the automatic infection analysis is then confirmed using the clinical ranking or benchmark medicinal information. In this work, the chest X-ray pictures provided by Rahman et al. [
19] are considered. This database consists of 7000 test images in which 3500 pictures fit in normal class and remaining 3500 is with TB traces. From these images, 70% images (2450 images) are considered as training class (in which 490 images are chosen for validation) and the remaining 30% (1050 images) are considered as testing class, as depicted in
Table 2.
Figure 2 presents the sample test images of the collected X-ray database.
3.2. Pre-Trained VGG19
DLS based clinical data evaluation is a frequently considered method and most of these approaches are used to implement automatic image categorization tasks [
5,
27,
28,
29]. In this work VGG19 is considered to examine the database and to support better TB diagnosis. Initially, the VGG19 is trained using X-ray pictures with the following tasks: (i) conventional augmentation (rotation and zoom) to boost the training images, (ii) fixing the learning rate as 1 × 10
−5 to improve the training and validation accuracy, (iii) training with Linear-Dropout-Rate (LDR) and Adam optimization. During this task, other vital parameters are assigned as follows; total iteration = 2000, total epochs = 50, dropout rates in fully connected layer = 50%, and classification with SoftMax unit using a 10-fold cross validation.
3.3. Feature Extraction
This part of research outlines the extraction of DLF and MLF with chosen technique.
3.3.1. Deep-Learning-Features
The proposed scheme initially aims to extract the DLF from VGG19. The final MaxPool layer section provides a feature vector of dimension and it is then passed through 3 numbers of fully connected layers with a dropout rate of 50% to obtain a reduced the DLF value to . This DLF is further reduced using the SA based optimization and the reduced feature is then combined with the MLF to obtain the hybrid feature vector.
3.3.2. Machine-Learning-Features
In the literature, it can be noted that the combined DLF and MLF helps to obtain a better diagnostic accuracy on various medical images and in this work, the necessary MLF is obtained using the LBP and DWT methods. Initially, the resized images are separately treated with the LBP and DWT and from these images, the necessary features are extracted.
During the LBP feature extraction task, the necessary LBP patterns are produced by assigning its weights as
and from these images, essential features with a dimension of
are extracted from every image. The LBP technique adopted in this work is found in [
30,
31,
32]. During the DWT feature extraction process, the test images are processed with the DWT technique discussed in [
33,
34,
35,
36,
37]. This technique separates the considered test image into four equal sections, such as approximation, vertical, horizontal, and diagonal images and from every image, the necessary feature are then extracted. The extracted LBP and DWT features are then combined to obtain the original MLF, which is then reduced with the SA to obtain the optimized MLF. The necessary images of the LBP and DWT can be found in
Figure 3 and
Figure 4, respectively.
Figure 3a–d presents LBP with various weights and
Figure 4 present the results of DWT.
Figure 4a,b depicts the outcome for normal and TB class images.
3.4. Seagull-Algorithm Based Feature Optimization
Recently, Heuristic-Algorithm (HA) based techniques are widely adopted in the literature to find the finest solution for a variety of real-world problems. The performance of any automatic data examination system depends mainly on the features considered to train and validate the classifier and to avoid the over-fitting problem, it is necessary to reduce the raw feature set. Feature selection is a recommended procedure, in which a chosen mathematical procedure (ex. Student’s
t-test) or heuristic algorithm approach is widely adopted. The implementation of mathematical expression is widely discussed in the literature [
38,
39] and to automate the feature selection process, the HA based techniques are also widely considered by the researchers [
40,
41,
42]. In the proposed work, the SA is considered to find the finest features from the raw DLF and MLF by finding the maximal value of the Cartesian Distance (CD) between the healthy and TB features.
The SA was invented in 2019 by Dhiman and Kumar [
43] to find the finest solution for a chosen industrial optimization problem. This algorithm is developed by mimicking the hunting behaviors found in Seagull birds. Due to its eminence, the SA is widely adopted to solve a variety of optimization tasks [
43,
44,
45].
The basic operation in SA includes (i) migration as a group to find fish swarm (exploration) and (ii) attacking the fish (exploitation). Usually, the seagull lives as a group and always search for the food source (fish swarm) in sea surface. When it finds the food source, every seagull in the group attacks the fish swarm with its own approach. During the search, every group is led by an experienced leader and other birds will follow the leader without collision. When the group finds the fish swarm, the leader will decide the best food source and other birds follow the best source for attacking. The pictorial representation of the SA is depicted in
Figure 5.
In the proposed work, traditional SA is considered for the feature selection problem and the various stages of this scheme are discussed below.
3.4.1. Migration as a Group to Find Fish Swarm (Exploration)
This part of the SA helps to recognize the movement of Seagull (agent) from one location to other in search of the fish group by satisfying the conditions, such as collision prevention, association towards best neighbor and being close to the best neighbor till the food source is detected.
The collision prevention task is mathematically depicted in Equation (1):
where
is considered to compute the new position of agent,
is current position, and
denotes the updated position.
The value of
A is gradually varied to adjust the positions of agents and this work is mathematically denoted in Equation (2):
where
= 2,
= 3000 and
.
During the exploration, the movement of an agent towards the best agent can be defined in Equation (3):
where
is a new position of search agents is,
denotes fittest agent, and
is a randomization parameter and is computed in Equation (4).
where
is a random value of range [0, 1].
During the exploration, every agent will closely fly with the best agent with a close distance as depicted in Equation (5):
3.4.2. Attacking the Fish (Exploitation)
During this phase, the agent (Seagull) takes a spiral path to dive into the sea to catch the fish and also takes the inverted path to reach the sky from the sea as depicted in
Figure 5. The operation in
x,
y, and
z-plane is defined in Equations (6)–(9):
where
r is radius of the path,
k is a random parameter [
], other parameters are constants, and these values are assigned in the earlier works [
43,
44,
45].
The necessary parameters of the SA are assigned as follows: number of Seagull (agent) = N = 30, search dimension (D) = 2,
= 3000 and stopping criteria=maximization of Cartesian Distance (CD) between features or
. In this work, identification of maximized CD is chosen as the objective value and the SA helps to find the individual features of healthy/TB class whose value of CD is maximal. The SA based optimization and serial concatenation work is depicted in
Figure 6. After finding the optimal features, the necessary hybrid features are obtained using the serial feature concatenation procedure.
The number of DLF and MLF considered in this work is presented in Equations (10)–(12) and in this work, the SA is then used to reduce these features to a lower value.
The number of DF and HF available for the optimization is depicted in Equations (10)–(12).
Initially, the DLF is optimized with the SA and this process helps to reduce the feature set to a value of
. Similar procedure is then implemented on the MLF (
+
) and this procedure helped to obtain a feature vector with dimension
. These two features are then considered to obtain a hybrid feature as depicted in Equation (13):
The feature vector shown in Eqn.13 is then used to train and validate the classifier and, in this research, well known classifiers, such as SoftMax, Naïve-Bayes, Logistic Regression, Decision-Tree (DT) variants, K Nearest Neighbors (KNN) variants, and SVM variants are considered to validate the performance of the proposed system [
40,
41,
42].
3.5. Performance Validation
The merit of automatic disease discovery arrangement is to be confirmed by computing the essential performance values. In this work, the measures obtained from the confusion matrix are considered to substantiate the importance of the proposed system. These measures, includes True-Positive (TP), False-Negative (FN), True-Negative (TN), False-Positive (FP), Accuracy (ACC), Precision (PRE), Sensitivity (SEN), Specificity (SPE), and Negative-Predictive-Value (NPV). The mathematical expressions of these values are presented in Equations (14)–(18) [
27,
28].
4. Results and Discussion
This section of research demonstrates the experimental result, and this work is performed by a workstation, Intel i7 2.9GHz processor with 20 GB RAM and 4GB VRAM equipped with Matlab®.
The resized X-ray images are initially tested using the VGG19 scheme. Initially, the pre-trained DLS is trained using the test images depicted in
Table 2 and after getting the necessary accuracy, its performance is then validated with a 10-fold cross validation procedure.
Figure 7 depicts the outcome during the deep-feature extraction procedure with VGG19.
Figure 7a presents the heat-map obtained for both the normal and TB class trial images,
Figure 7b–f shows the sample outcome extracted from each convolutional layer. This procedure helps to obtain a DLF of dimension
. This feature vector is then considered to verify the TB detection performance of VGG19 with the SoftMax based binary classification and the attained results are shown in
Table 3. This table confirms that Trial 8 helped to obtain the better classification accuracy compared to other trials. A similar procedure is then repeated with other pre-trained schemes and the attained results are presented in
Table 4. The results existing in
Table 4 confirm that the classification accuracy achieved with VGG19 is better compared to other methods considered for the assessment.
Figure 8 graphically shows the accuracy achieved during the 10-fold cross validation and
Figure 9 presents the Spider-Plot of
Table 4 to verify the overall performance of the chosen schemes. The Spider-Plot is one of the graphical procedures normally considered to find the overall performance during the classification task. The plot which covers maximal area is considered superior compared to other plots. In
Figure 9, it is noted that the plot area of VGG19 is more compared to other approaches, which confirms its superiority over other DLS. The obtained Confusion-Matrix (CM) for VGG19 with SoftMax classifier is depicted in
Figure 10a,b presents the result by SVM- Medium Gaussian (SVM-MG) classifier.
The performance of the proposed scheme is initially verified using the optimized value of DLF and MLF, and the results are depicted in
Table 5. This result confirms that the DLF-supported classification presented better accuracy (93.6667%) when considering KNN-Coarse. Similarly, the MLF-based classification achieved 82.1905% accuracy with SVM- Fine Gaussian. The confusion matrix achieved during this work is presented in
Figure 10.
Figure 10b,c depict the confusion matrix achieved with optimal DLF and MLF.
After confirming the performance of VGG19 with traditional features (
), the classification of X-ray is once again repeated with the SA optimized hybrid features presented in Equation (13) (DLF+MLF). The classification achieved with the SoftMax helps to provide an accuracy of 95.09%. Further, the classification performance of VGG19 with other binary classifiers is verified using 10-fold cross validation and the attained results are presented in
Table 6. This table confirms that the result obtained with SVM-Medium Gaussian (SVM-MG) classifier is better compared to other approaches and the CM attained with this classifier is depicted in
Figure 10d.
To verify the overall performance of the results shown in
Table 6, a Glyph-Plot is then constructed, and the result is demonstrated in
Figure 11. This plot also confirms that the SVM-MG provides better outcome compared to other classifies adopted in this study. This result confirms that the VGG19 with concatenated feature helps to obtain a better overall result when SVM-MG classifier is employed.
The merit of the proposed scheme is then confirmed with a comparative analysis with the results provided by Rahman et al. [
18]. The plot depicted in
Figure 12 verifies that the accuracy by the proposed system with SVM-MG classifier is very close to DenseNet201 result of [
18].
This research work automatically classified X-ray images using SA-optimized hybrid image features (DLF+MLF) with a VGG19 scheme. The classification task is executed using SoftMax and other binary classifiers existing in the literature to achieve better classification accuracy. The experimental result of this work confirms that the proposed scheme works well on chest X-ray images. In the future, this scheme can be considered to evaluate the clinically collected chest X-ray images. Further, the merit of this system can be tested and confirmed using other benchmark chest X-ray pictures of different lung irregularities. The main limitation of the proposed scheme is the integration of deep and handcrafted features, and this procedure can be improved by considering the ensemble of features technique in future.