Xception Taylor Cascade Neuro Network based infection level identification of tuberculosis using sputum images

Abstract

Currently, one amongst most primary health problems and an enormously transmittable disease is Tuberculosis (TB). This disease spreads all over the world and is commonly developed by Mycobacterium TB (MTB). TB causes fatality if it is not identified at earlier stages. Thus, accurate and effectual model is necessary for detecting infection level of TB. Here, Xception Taylor Cascade Neuro Network (Xception T-Cascade NNet) is presented for infection level identification of TB utilizing sputum images. Firstly, input sputum image acquired from certain database is pre-processed by denoising and histogram equalization utilizing contrast limited adaptive histogram equalization (CLAHE). SegNet is utilized for bacilli segmentation and it is tuned by White Shark Optimizer (WSO). Thereafter, suitable features such as designed discrete cosine transform (DCT) with angled local directional pattern (ALDP), statistical features, shape features and gray-level co-occurrence model (GLCM) texture features are extracted for further processing. Lastly, infection level identification of TB is conducted by Xception T-Cascade NNet. However, Xception T-Cascade NNet is an integration of Xception with Cascade Neuro-Fuzzy Network (NFN) by Taylor concept. In addition, Xception T-Cascade NNet achieved 88.5% of accuracy, 90.8% of true negative rate (TNR) and 89.4% of true positive rate (TPR) and as well as minimal false negative rate (FNR) of 0.092 and false positive rate (FPR) of 0.106.

Keywords

Tuberculosis (TB)Xception cascade neuro-fuzzy network (NFN)SegNet white shark optimizer (WSO)

1. Introduction

TB is one amongst ordinary frightful and contagious diseases influencing several people across globe that is spread by the bacterium termed MTB. The MTB affects any organs of a human body, specifically lungs through blood stream as well as lymphatic system. With regard to count of mycobacteria in sputum accumulated for certain person, culture positivity and smear sensitivity are evaluated. Moreover, count of bacilli in sputum forecasts severity level of TB [22]. TB disease spreads by means of air while an affected person sneezes, coughs or speaks [2]. Since TB is the primary transmittable disease that causes high infection and huge count of fatalities [1]. In accordance to report of World Health Organization (WHO), numerous incident cases take place worldwide and several people have been died [3]. TB is managed by demands organizing diverse factors that involve policies corresponding to a public economic, humanitarian and health. Thus, TB is yet concerned as a vital health problem all over the world. TB diagnosing needs much concentration and wide-ranging treatment. The classical approaches adapted to detect TB have insufficiencies in its detection rate and time for respond. Furthermore, these methods are costly and needs complex structures for setups [19]. In several countries, detection of TB is accomplished manually by obtaining bacteria counts utilizing microscope. TB diagnosis utilizing sputum smear is a resultant obtained from the Ziehl–Neelsen (ZN) staining in microscopic images [2].

Automatic detection of TB needs image segmentation for identifying bacilli counts, feature extraction (FE) from the segmented sputum images and image classification into diverse categories based upon features. These sputum images have excellent visibility. An overall number of bacilli in sputum image specify a severity level of disease in the persons [25]. Therefore, it is significant to determine count of bacillus in sputum images before classifying particular image. The current techniques have chosen for manual screening of bacilli counts in an image. This screening process of an image needs much human interference and therefore, could direct to various errors [23]. Thus, automated identification of bacillus in sputum image is necessary [4]. An insufficiency of the Bacilli classification has been highlighted by molecular researches. The broad variety of compositions in chromosomal DNA signifies genetic diverseness and recommends that species of bacilli must be re-categorized into various genera [26]. The diagnostic examination of TB bacillus involve Polymerase Chain Reaction (PCR) analysis, chest X-ray, Gene Xpert, culture investigation, tuberculin skin assessment and sputum microscopic test [20]. As an incidence rate is higher in TB, WHO suggests earlier tests and effectual method to identify TB bacillus. There are two diverse microscopes namely fluorescence microscope and optical microscope. These microscopes are employed for detecting TB bacillus. The TB bacillus can be stained exploiting dyes and therefore, it is easier to differentiate at a time of microscopic test [3].

Various approaches are utilized for enhancing sensitivity, even though quality of TB identification is based upon diverse criterions such as worse devices, large workloads, unskilled personnel and so on, wherein it is normally conducted. Hence, researches of TB identification from the sputum images concentrate on automated schemes [2, 23]. Automated TB identification based upon classifiers has been gained admiration as they need minimum workloads [24]. The fully automated model involves auto focus process together with automatic detection. An auto focus of the microscopes has a crucial part in managing microscopes for capturing targeted images. It is a fundamental portion of an automatic TB assessment model. At present some fully automatic approaches are developed. Microscopy test of tainted sputum images has been remained a keystone of the pulmonary TB diagnosing worldwide. A diagnosing quality relies upon criterions that are usually conducted. Therefore, automated schemes for detecting TB from sputum images are developed [4, 23]. Deep learning (DL) methods have critical part in severity level identification of TB. Hence, artificial neural network (ANN) is accepted as information bases appear to be a favorable technique for detecting diagnosing and its treatment. In addition, ANN is employed for diagnosing progression of pulmonary TB [21]. The extracted features from ANN provide training ability, which makes a model accessible for newer diseases that are influenced by Mycobacterium strain utilizing mutation of causative organisms. It provides high flexibility in diagnosing system and therefore, makes a model opened-ended [1].

An essential aim is to present Xception T-Cascade NNet for infection level identification of TB utilizing sputum images. TB refers to an infectious disease influenced by the Mycobacterium. The human lungs are mostly affected by this disease. Firstly, input sputum image is obtained from specific database that is then pre-processed by denoising and histogram equalization utilizing CLAHE. Thereafter, bacilli segmentation is conducted on pre-processed sputum image employing SegNet and it is tuned by WSO. Next, features like DCT with ALDP, shape features namely area, coverage, length and density, statistical features namely kurtosis, variance, mean and skewness as well as GLCM texture features namely autocorrelation, energy, dissimilarity, homogeneity and entropy are extracted. Finally, infection level identification of TB is performed by Xception T-Cascade NNet that is introduced by merging Xception with cascade NFN, wherein modification of layers is done by Taylor concept.

The prime contribution is interpreted as follows.

•
Proposed Xception T-Cascade NNet for infection level identification of TB utilizing sputum images: Earlier infection level identification of TB is a vital chore that assists to avoid this disease from the severity. Here, Xception T-Cascade NNet is designed for infection level identification of TB. However, Xception T-Cascade NNet is formed by integrating Xception and cascade NFN, wherein layers of network are modified based upon Taylor concept.

The arrangement of residual sections are: Section 2 exhibits literature survey of classical methods gathered for infection level identification of TB, Section 3 specifies Xception T-Cascade NNet methodology, Section 4 indicates Xception T-Cascade NNet outcomes and Section 5 presents conclusion of Xception T-Cascade NNet.
2. Motivation

TB specifies to highly infective disease and it is one amongst significant health issues across globe. The complicated state of TB needs experienced professionals for identifying severity level of this disease utilizing sputum images. This motivated to introduce a technique for infection level identification of TB by collecting several methods regarding infection level identification of TB. The reviewed models and their experienced shortcomings are elucidated in this part.

2.1 Literature survey

Chithra and Jagatheeswari [1], developed adaptive fractional crow (AFC)-deep Convolutional Neural Network (CNN) for infection level identification of TB. It attained superior performance for severity detection of TB, but it did not specify the time utilized by the method for identifying TB infection level. Mithra and Emmanuel [2], designed Gaussian Decision Tree based Deep Belief Network (GDT-DBN) for diagnosing TB infection level utilizing sputum images. An overfitting issue was resolved by this approach, even though some error occurred while managing non-linearity data. Samuel and Baskaran [3], devised computer assisted system to identify TB bacilli for assisting pathologists with high sensitivity as well as specificity. This method was highly robustness, but still the pixels utilized in this approach did not characterize finite boundary. Mithra and Emmanuel [4], presented Gaussian-Fuzzy-Neural network (GFNN) in order to detect TB, which perfectly identified count of bacilli, even though, more training was required for better outcomes.

Kurmi et al. [5]introduced superellipse and supervised variational contour systems for TB assessment. It offered best segmentation accuracy, but it failed to reduce segmentation complexity. Panicker et al. [6]devised CNN for detecting TB bacilli. The count of parameters was significantly decreased in this model, though it was not capable to assist training procedure. Aung et al. [7]utilized Droplet digital PCR (ddPCR) to detect TB utilizing sputum samples. It was capable to identify MTB complexities in higher percentage of sputum samples, but it had high turnaround time, higher expenditure per response and required well-experienced personnel. Sirohi et al. [8]developed K-Means clustering for detecting pulmonary TB. This technique performed precisely in testing images of unacceptable sputum samples. However, it failed to examine its strength by conducting investigation on enormous count of images before utilizing it in regular microbiology practices.

2.2 Challenges

Few demerits experienced by the approaches considered for review regarding infection level identification of TB are elucidated beneath.

•
AFC-Deep CNN presented in [1] for infection level identification of TB obtained better outcomes, but still this model was not enhanced for the mycobacterial infection verification as well as drug-efficient researches.
•
An approach developed in [2] to identify TB infection level utilizing sputum image faced difficulty to distinguish uncommon MTB objects as both of them look like similar morphology.
•
In [4], GFNN devised for TB detection utilizing sputum images attained better outcomes and perfectly classified count of bacilli. However, this model was not appropriate for real-time scenarios.
•
CNN [6] structure was designed for identifying TB bacilli from the sputum images. The performance of this model was enhanced, even though it utilized similar count of samples to evade data imbalance problems.
•
Presently, TB is one amid most contagious disease that spreads through infective agents like MTB. The sputum images play as a significant tool for diagnosing pulmonary TB, though it faces some disadvantages like larger observation time, less sensitivity and so forth.

3. Proposed Xception T-Cascade NNet for infection level identification of TB utilizing sputum images

Figure 1.

Pictorial presentation of Xception T-Cascade NNet for infection level identification of TB utilizing sputum images.

Earlier and precise diagnosing of TB is crucial for exact treatment. Recently, infection level identification utilizing hybrid DL is much challenging. Here, Xception T-Cascade NNet is designed for infection level identification of TB utilizing sputum images. Firstly, considered input sputum image is pre-processed by denoising and histogram equalization utilizing CLAHE. Then, bacilli segmentation is accomplished utilizing SegNet that is tuned by WSO. Afterwards, features namely DCT with ALDP, shape features, GLCM features and statistical features are extracted. Lastly, infection level identification of TB is executed utilizing Xception T-Cascade NNet that is modeled by merging Xception and cascade NFN with Taylor concept. Figure 1 represents pictorial presentation of Xception T-Cascade NNet for infection level identification of TB utilizing sputum images.

3.1 Acquisition of input sputum image

An input sputum image is taken from particular database to conduct infection level identification of TB that can be described by,

$\displaystyle U=\left\{{U_{1},U_{2},\ldots,U_{t},\ldots,U_{\ell}}\right\}$ (1)

Here, $U_{\ell}$ implies overall sputum images in database $U$ whereas $U_{t}$ indicates $t^{\text{th}}$ sputum image.

3.2 Image pre-processing by denoising and histogram equalization utilizing CLAHE

The pre-processing stage should be conducted for an image to improve its quality before performing further processes. Here, image pre-processing is done by denoising and histogram equalization utilizing CLAHE. The sputum image $U_{t}$ is taken as input to obtain filtered image.

3.2.1 Denoising and histogram equalization utilizing CLAHE

CLAHE [10] is employed as a contrast emphasizing after denoising utilizing non-local means filter. A non-local means filter is developed for an image denoising. This technique is exploited to reduce noises that mention pixel value is substituted by a weighted average of the referenced pixels relying upon block similarity centered on pixel of interest as well as referenced pixel. When $d\,\left(y\right)$ relates to pixel of the interest in an image $Y$ , $\textit{vz}\,\left({y,\textit{si}}\right)$ relates to weight whereas $\textit{ac}\,\left({y,\textit{si}}\right)$ relates to similarity of block of interest as well as reference block. A pixel value $\textit{rw}\,\left({Y,\textit{si}}\right)$ afterwards denoising can be modeled by,

$\displaystyle\textit{rw}\,\left(y\right)=\sum\limits_{\textit{si}\in Y}{vz\,% \left({y,\textit{si}}\right)\,d\,}\left({y,\textit{si}}\right)$ (2)

$\displaystyle\textit{vz}\,\left({y,\textit{si}}\right)=\frac{\textit{ac}\,% \left({y,\textit{si}}\right)}{\sum\nolimits_{\textit{kn}\in Y}{\textit{ac}\,% \left({y,\textit{kn}}\right)}}$ (3)

$\displaystyle\textit{ac}\,\left({y,\textit{si}}\right)=\exp\,\left({-\frac{% \left\|{\,K\!\!\!\!/\,\left(y\right)-K\!\!\!\!/\,\left({\textit{si}}\right)\,}% \right\|^{2}}{E\!\!\!\!/^{2}}}\right)$ (4)

where, $K\!\!\!\!/\,\left(\textit{si}\right)$ specifies two-dimensional (2D) vector achieved by organizing pixel values in a block of $Z\!\!\!\!/\times Z\!\!\!\!/$ across pixel $y$ and $E\!\!\!\!/$ mentions the parameter to determine smoothing strength.

CLAHE refers to histogram equalization technique. An adaptive histogram equalization is utilized owing to brightness alters considerably in the considered image. Initially, image is separated into equally-sized regions, which do not overlap. Then, histogram is evaluated for individual area. Lastly, histogram is clipped and a clipped value is dispersed thereafter histogram equalization. A filtered image obtained by applying denoising and histogram equalization utilizing CLAHE can be specified as $C_{t}$ .

3.3 Bacilli segmentation using SegNet

Image segmentation is a most vital step in clinical image assessment. Its prime aim is to differentiate the objects influences the specific disease within a tissue. Here, bacilli segmentation is carried out utilizing SegNet and its tuning procedure is accomplished by WSO. An input fed to execute bacilli segmentation is $C_{t}$ .

3.3.1 Structure of SegNet

Figure 2.

SegNet model for bacilli segmentation.

SegNet [11, 29] is the framework that contain encoder as well as relevant decoder, which is pursued by pixel-wise classification layer at an end. The SegNet model for bacilli segmentation is revealed in Fig. 2.

At encoder network, each of an encoder executes convolutional (conv) processing and filter bank to develop sort of the feature maps. Then, feature maps are batch normalized and next, component-wise rectified-linear unit (ReLU) $\max\,\left({0,j}\right)$ is exploited. Assume an input $\zeta^{b-1}\times\zeta^{b-1}$ patch $D_{A\!\!\!\!/}^{\textit{ki}-1}$ is convolved with $w^{b}\times w^{b}$ locally relative range in specific image $D_{A\!\!\!\!/}^{\textit{ki}-1}$ surrounded by a filter $mn_{\textit{sx}}^{b}$ . At conv processing, each of a filter $\textit{mn}_{\textit{sx}}$ convolves across $\left({\zeta^{b-1}-w^{b}+1}\right)\times\left({\zeta^{b-1}-w^{b}+1}\right)$ pixels in $D_{A\!\!\!\!/}^{\textit{ki}-1}$ . An active operation outcome is represented as,

$\displaystyle\delta^{b}=F\left({\kappa_{\textit{sx}}^{b}}\right)=\frac{1}{1+% \exp\,\left({-\kappa_{\textit{sx}}^{b}}\right)}$ (5)

Here, $\kappa_{\textit{sx}}^{b}$ illustrates additional products of filter coefficient as well as image pixel of the locally corresponded region encircled by filters. Thereafter, max-pooling with stride and window is accomplished and last outcome is sub-sampled. For achieving translational invariance across tiny spatial differences in an input image, a max-pooling is utilized. The resultants of sub-sampling in largest input image models for individual pixel in a feature map. There occurs feature map’s spatial resolution loss while various sub-sampling layers as well as max-pooling is able to achieve extremely translational invariance for a purpose of classification. On account of max-pooling operation, resultant image size become $\frac{\left({\zeta^{b-1}-w^{b}+1}\right)}{\beta}\times\frac{\left({\zeta^{b-1}% -w^{b}+1}\right)}{\beta}$ , wherein $\beta$ denotes scale at which operation is exploited.

A decoder conducts feature map upsampling by trained max-pooling indices obtained from associated encoder feature map. Next, feature maps are convoluted by learnable decoder filter bank for developing densest feature maps. Afterwards, Batch normalization (BN) is exerted to individual maps. Lastly, high-dimension demonstration at an outcome of final decoder is fed to softmax classifier that classifies every pixel individually. An outcome from classifier is specified by VW channel image of probabilities, whereon VW mentions overall categories. The forecasted segmentation refers to a class having high feasibility of all pixels. This classifier signifies to supervised model, which generalizes logistic regression (LR) that can be modeled by,

$\displaystyle\Upsilon_{{\rm O}^{\left(b\right)}}\,\left({\tau\tau^{b}=ch\left|% {\,\varepsilon\varepsilon^{b}}\right.}\right)=\frac{1}{1+\exp\,\left({-{\rm O}% ^{\left(b\right)\,{\rm I}}\textit{ch}^{b}}\right)}$ (6)

where, $\Upsilon_{{\rm O}^{\left(b\right)}}\,\left(\bullet\right)\,\left({\textit{ch}% \in\left\{{0,1}\right\}}\right)$ implies sigmoid function with the parameters ${\rm O}^{\left(b\right)}$ . A segmented outcome from the SegNet is elucidated by $S_{t}$ .

3.3.2 Training of SegNet utilizing WSO

WSO [12] is presented for resolving optimization issues over continual search space. The fundamental concept of WSO is motivated by characteristics of huge white sharks that include their remarkable senses like smelling and hearing while foraging as well as navigation. WSO design integrates promising exploitative and explorative searches in its updating process to vary and update solutions in a random manner. WSO reliably achieves global optimum outcomes with generally high performance. Here, WSO is employed to train SegNet to obtain optimally best outcome.

White shark position encoding

In QP search space, best outcome is attained by continually tuning SegNet learning parameter denoted as LM, such a manner that $\textit{QP}=\left[{1\times\textit{LM}}\right]$ .

Fitness measure

The computation of fitness function is executed by determining variation amid targeted and predicted resultants, which is specified by,

$\displaystyle\Psi=\frac{1}{\ell}\sum\limits_{t=1}^{\ell}{\left[{{\rm T}_{t}-S_% {t}}\right]^{2}}$ (7)

Here, overall samples are given as $\ell$ whereas ${\rm T}_{t}$ and $S_{t}$ signifies targeted and segmented outcome from SegNet.

The beneath steps are followed by WSO to achieve optimum outcome.

Step 1:Initializing the solution

At first, population of the white sharks is initialized in a search space that can be specified as follows.

$\displaystyle W=\left\{{W_{1},W_{2},\ldots,W_{s},\ldots,W_{c}}\right\}$ (8)

Here, $c$ implies total variables and $W_{s}$ illustrates $s^{\text{th}}$ solution whereas $W$ represents population.

Step 2:Computing objective function

An objective function is computed based on difference amid targeted and predicted outcome from SegNet utilizing Eq. (7).

Step 3:Movement speed against prey

When white shark recognizes prey position based upon uncertainty of waves it listens as a prey is traveling, it move against a prey in rolling movement that is given by,

$\displaystyle t_{\upsilon+1}^{s}=\psi\,\left[{t_{\upsilon}^{s}+m_{1}\,\left({W% _{q\,\textit{best}_{\upsilon}}-W_{\upsilon}^{s}}\right)\times l_{1}+m_{2}\,% \left({W_{\textit{best}}^{t_{\upsilon}^{s}}-W_{\upsilon}^{s}}\right)\times l_{% 2}}\right]$ (9)

Here, $s=1,2,\ldots,h$ refers to index of white shark for population having size $h$ , $t_{\upsilon+1}^{s}$ mentions newer velocity vector of $s^{\text{th}}$ white shark in $\left({\upsilon+1}\right)^{\text{th}}$ step, $t_{\upsilon}^{s}$ indicates present speed vector of $s^{\text{th}}$ white shark in $\upsilon^{\text{th}}$ step, $W_{q\,\textit{best}_{\upsilon}}$ implies global better location vector achieved thus far by white shark in $\upsilon^{\text{th}}$ iteration, $W_{\upsilon}^{s}$ denotes present location vector of $s^{th}$ white shark in $\upsilon^{\text{th}}$ step, $W_{\textit{best}}^{t_{\upsilon}^{s}}$ describes $s^{\text{th}}$ better known location vector recognized to swarm whereas $t^{s}$ represents $s^{\text{th}}$ index vector of white shark reaching superior location. $l_{1}$ and $l_{2}$ specifies two random numbers generated uniformly in a range 0 and 1. $m_{1}$ and $m_{2}$ signifies forces of white sharks whereas $\psi$ depicts constriction factor recommended in WSO for assessing converging characteristic of the white sharks. The expressions are given as follows.

$\displaystyle t=\left[{h\times R\,\,\left({1,h}\right)\,}\right]+1$ (10)

Here, $R\,\,\left({1,h}\right)$ represents the random number vector developed with uniform distribution at a range 0 and 1.

$\displaystyle m_{1}=m_{\max}+\,\left({m_{\max}-m_{\min}}\right)\times e^{-% \left({4\upsilon/V}\right)^{2}}$ (11)

$\displaystyle m_{2}=m_{\min}+\,\left({m_{\max}-m_{\min}}\right)\times e^{-% \left({4\upsilon/V}\right)^{2}}$ (12)

where, $\upsilon$ and $V$ indicates present and maximal count of the iterations whereas $m_{\min}$ and $m_{\max}$ reveals initial as well as secondary velocities to acquire better movement for the white sharks.

$\displaystyle\psi=\frac{2}{\left|{\,2-\hbar-\sqrt{\hbar^{2}-4\hbar\,}}\right|}$ (13)

Here, $\hbar$ illustrates acceleration coefficient.

Step 4:Movement against optimum prey

During this condition, white sharks navigate in arbitrary position in prey search as a state of attribute of fish school searching for food source. Here, location update strategy presented in beneath expression is employed for describing characteristic of white shark as it move against prey.

$\displaystyle W_{\upsilon+1}^{s}=\left\{{\begin{array}[]{l}W_{\upsilon}^{s}\,.% \,\neg\oplus W_{0}+\textit{uu}\,.\,\textit{pq}+\textit{ll}\,.\,kr\,;\,R<% \textit{gt}\\ W_{\upsilon}^{s}+t_{\upsilon}^{s}/\textit{fe}\,;\,R\geqslant\textit{gt}\\ \end{array}}\right.$ (14)

Here, $W_{\upsilon+1}^{s}$ mentions newer location vector of $s^{\text{th}}$ white shark in $\left({\upsilon+1}\right)^{\text{th}}$ iteration, $\neg$ presents negation operator whereas pq and kr implies one-dimensional (1D) binary vector formulated employing Eqs (15) and (16). ll and uu specifies lower limit and upper limit of search space, $W_{0}$ indicates logical vector expressed in Eq. (17). fe symbolizes frequency of white shark’s wave movement evaluated by Eq. (18), $R$ denotes randomly chosen number generated at a range amid 0 and 1 whereas gt signifies motion force, which improved with count of iterations, as white shark reaches prey and it is modeled in Eq. (19).

$\displaystyle\textit{pq}=\textit{sgn}\,\left({W_{\upsilon}^{s}-\textit{uu}}% \right)>0$ (15)

$\displaystyle\textit{kr}=\textit{sgn}\,\left({W_{\upsilon}^{s}-\textit{ll}}% \right)<0$ (16)

$\displaystyle W_{0}=\oplus\,\left({\textit{pq},\textit{kr}}\right)$ (17)

Here, $\oplus$ denotes bit-wise XOR operation. Equations (15) and (16) are significant for supporting outcomes to randomly act in search space. These equations are also vital for assisting white sharks to find every probable regions of search space.

$\displaystyle\textit{fe}=\textit{fe}_{\min}+\frac{\textit{fe}_{\max}-\textit{% fe}_{\min}}{\textit{fe}_{\max}+\textit{fe}_{\min}}$ (18)

where, $\textit{fe}_{\min}$ and $\textit{fe}_{\max}$ signifies minimal and maximal frequency of undulating movement.

$\displaystyle\textit{gt}=\frac{1}{\left({\textit{pq}_{0}+e^{\left({V/2-% \upsilon}\right)/\textit{pq}_{1}}}\right)}$ (19)

Here, $\textit{pq}_{0}$ and $\textit{pq}_{1}$ reveals two positive constant utilized for managing exploration as well as exploitation characteristics.

Step 5:Movement against optimum white shark

Huge white sharks are capable to manage their locations against optimum one, which is closer to prey and it is estimated as,

$\displaystyle{W^{\prime}}^{s}_{\upsilon+1}=W_{q\,\textit{best}_{\upsilon}}+\Re% _{1}\vec{B}_{W}\,\textit{sgn}\,\left({\Re_{2}-0.5}\right)\,\,\Re_{3}<O_{O}$ (20)

where, ${W^{\prime}}_{\upsilon+1}^{s}$ signifies updated location of $s^{\text{th}}$ white shark regarding to prey location, $\textit{sgn}\,\left({\Re_{2}-0.5}\right)$ offers either value 1 or else $-$ 1 for changing search direction, $\Re_{1},\Re_{2}$ and $\Re_{3}$ denotes randomly selected numbers, which range amongst 0 and $-$ 1, $\vec{B}_{W}$ reveals distance amid white shark and prey computed by Eq. (21) whereas $O_{O}$ depicts parameter recommended for expressing strength of smelling as well as vision for the white sharks while they pursue other sharks, which are closer to optimum prey and it is calculated by Eq. (22).

$\displaystyle\vec{B}_{W}=\left|{\,R\times\,\left({W_{q\,\textit{best}_{% \upsilon}}-W_{\upsilon}^{s}}\right)\,}\right|$ (21)

Here, $R$ implies randomly chosen number between 0 and 1 whereas $W_{\upsilon}^{s}$ symbolizes present location of the white sharks with respective to $W_{q\,\textit{best}_{\upsilon}}$ .

$\displaystyle O_{O}=\left|{\,1-e^{\left({-\textit{pq}_{2}\times\upsilon/V}% \right)}\,}\right|$ (22)

where, $\textit{pq}_{2}$ denotes positive constant employed for controlling exploration as well as exploitation characteristics.

Step 6:Fish school characteristics

To imitate characteristics of white shark school mathematically, an initial optimal solutions are conserved and location of other sharks are updated according to these optimum locations. A fish school attribute of the white sharks can be computed by,

$\displaystyle W_{\upsilon+1}^{s}=\frac{W_{\upsilon}^{s}+{W^{\prime}}^{s}_{% \upsilon+1}}{2\times R}$ (23)

Here, $R$ denotes randomly chosen number that is distributed uniformly within an interval 0 and 1.

Step 7:Termination

WSO is terminated after attaining optimum resultant by continually performing above steps. Algorithm 1 exposes pseudo code of WSO.

Algorithm 1: Pseudo code of WSO
SL. No	Pseudo code of WSO
1	Input: $W_{\upsilon}^{s},t,m_{1},m_{2},\psi,\textit{pq},\textit{kr},W_{0},\textit{fe},% \textit{gt},O_{O}$
2	Output: $W_{\upsilon+1}^{s}$
3	Begin
4	Initialization of problem parameters
5	Initialize WSO parameters
6	Generate initial locations of WSO randomly
7	Compute objective function utilizing Eq. (7)
8	Initialize initial population velocity
9	while $\left({\upsilon<V}\right)$ do
10	Update $t,m_{1},m_{2},\psi,\textit{pq},\textit{kr},W_{0},\textit{fe},\textit{gt},O_{O}$ parameters
11	for $s=1$ to $h$ do
12	Update newer velocity vector utilizing Eq. (9)
13	end for
14	for $s=1$ to $h$ do
15	if $R<\textit{gt}$ then
16	Update new location vector utilizing Eq. (14)
17	else
18	end if
19	end for
20	for $s=1$ to $h$ do
21	if $R<O_{O}$ then
22	Determine $\vec{B}_{W}$ utilizing Eq. (21)
23	for $s==1$ to $h$ then
24	Update location utilizing Eq. (20)
25	else
26	Compute fish school attribute utilizing Eq. (23)
27	end if
28	end if
29	end for
30	Modify location of white shark that proceeds beyond a boundary
31	Estimate and update newer locations
32	$\upsilon=\upsilon+1$
33	end while
34	Return optimum outcome attained thus far
35	Terminate

3.4 FE

FE is a crucial phase in an image processing, wherein features of image are extracted. Then, it is numbered and arranged into specific classes. FE is a reduction in the feature values of an image for obtaining superior outcomes and rapid speed while performing classification procedure. Here, features considered for extraction are like DCT with ALDP, shape features, statistical features as well as GLCM texture features. An input taken to process FE is bacilli segmented image $S_{t}$ .

3.4.1 DCT with ALDP

DCT [16] comprises group of base vectors that are created cosine operations. DCT has a privilege to concentrate on major helpful data within specific coefficients for the ordinary image. DCT matrix $N\!\!\!\!/_{\textit{ee}}\left({\textit{gs},\textit{zb}}\right)$ on $M\times L$ gray scale can be modeled by,

$\displaystyle X\!\!\!\!/\,\left({\textit{ij},\textit{pa}}\right)=\sum\limits_{% \textit{{gs}}=0}^{M-1}{\sum\limits_{\textit{zb}=0}^{L-1}{N\!\!\!\!/_{\textit{% ee}}\left({\textit{gs},\textit{zb}}\right)\,\mathchar 22\mskip-10.0mu \lambda% \,\left(\textit{ij}\right)\,\mathchar 22\mskip-10.0mu \lambda\,\left({\textit{% pa}}\right)}}\times\cos\left[{\frac{\left({2\textit{gs}+1}\right)\,\textit{ij}% \pi}{2M}}\right]\,\cos\left[{\frac{\left({2\textit{zb}+1}\right)\,\textit{pa}% \pi}{2L}}\right]$ (24)

ALDP [14]is an enhanced form of local directional pattern (LDP). These features are consisted of 8-bit binary code generated for individual pixel. ALDP code implied by $\textit{ALDP}_{\textit{kb},\textit{he}}\,\left({\textit{DC}_{0},\textit{DC}_{1% },\ldots,\textit{DC}_{7}}\right)$ of pixel at location $\left({\textit{kb},\textit{he}}\right)$ with angle vector values, $\left({\textit{DC}_{0},\textit{DC}_{1},\ldots,\textit{DC}_{7}}\right)$ can be computed by,

$\displaystyle\textit{ALDP}_{\textit{kb},\textit{he}}\,\left({\textit{DC}_{0},% \textit{DC}_{1},\ldots,\textit{DC}_{7}}\right)=\sum\limits_{lr}^{7}{\textit{DC% }_{\textit{lr}}}\times 2^{\textit{lr}}$ (25)

Figure 3.

DCT with ALDP feature.

A newer feature presented in this research is DCT with ALDP. At first, ALDP feature is applied to DCT. Therefore, few bands such as, high low (HL), high-high (HH), low-low (LL) and low high (LH) are generated. Here, HH band is ignored as it contains much noise. Then, bands other than HH are concatenated to acquire feature vector. $P_{1}$ denotes an obtained feature vector from DCT with ALDP. Figure 3 indicates designed DCT with ALDP feature.

3.4.2 Shape features

Density, length, coverage and area [1] are considered shape features that are interpreted in beneath sub-sections.

(i)
Density

A procedure employed to compute density is termed as zoning. Firstly, segments are categorized to zones having unchanging dimensions and extract the feature for individual zone. Density features can be calculated as,

$\displaystyle K_{1}=\frac{\left|{\textit{NZ}_{\textit{to}}^{\textit{fg}}}% \right|}{\left|{\textit{NZ}_{\textit{to}}}\right|}$ (26)

Here, $\left|{NZ_{\textit{to}}^{\textit{fg}}}\right|$ represents overall foreground pixels comprised in $\textit{to}^{\text{th}}$ segment whereas $\left|{\textit{NZ}_{\textit{to}}}\right|$ refers to overall count of pixels in $\textit{to}^{\text{th}}$ segment and $K_{1}$ denotes density feature.
(ii)
Length

A length feature illustrates to total length of segmented area (bacilli) of considered image. The length feature can be termed as $K_{2}$ .
(iii)
Coverage

Coverage feature is employed to categorize persons utilizing image or shape coverage. To obtain this feature, average of extracted crucial points is evaluated. This average value is employed for signifying object location and coverage factor is exploited to evaluate width as well as length of bacilli. $K_{3}$ represents coverage feature concerned for FE.
(iv)
Area

An area feature is specified as a region covered by a bacilli in segments and this covered region is extracted for individual segment. The area considered for extraction is denoted as $K_{4}$ .

The extracted shape features can be presented as $K_{t}$ , such that,

$\displaystyle K_{t}=\left\{{K_{1},K_{2},K_{3},K_{4}}\right\}$ (27)

3.4.3 Statistical features

The statistical features taken into concern for extraction are skewness, variance, kurtosis and mean.

(i)
Mean

Mean [15] signifies to a texture feature, which indicates brightness level of an image. Moreover, mean estimates average value of intensity values that is given by,

$\displaystyle E_{1}=\sum\limits_{\textit{bn}=1}^{\textit{LV}}{\textit{bn}\,% \textit{PR}\,\left({\textit{bn}}\right)}$ (28)

where, $\textit{PR}\,\left({\textit{bn}}\right)$ mentions to probability distribution of the bin bn with LV levels.
(ii)
Variance

Variance [27] presents deviation values of image gray levels corresponding to mean gray level. It can be modeled as,

$\displaystyle E_{2}=\sum\limits_{\textit{ry}=0}^{\textit{GA}-1}{\left({\textit% {ry}-E_{1}}\right)^{2}}*\textit{PR}\,\left({\textit{ry}}\right)$ (29)

Here, $\textit{PR}\,\left({\textit{ry}}\right)$ indicates probability of ry that denotes individual gray level in an image whereas GA symbolizes overall gray levels.
(iii)
Skewness

Skewness [15] signifies to an evaluation of dissimilarity of intensity level distribution regarding mean, which is computed by,

$\displaystyle E_{3}=\frac{1}{\left(\textit{SD}\right)^{3}}\sum\limits_{\textit% {bn}=1}^{\textit{LV}}{\left({\textit{bn}-E_{1}}\right)^{3}}\textit{PR}\,\left(% \textit{bn}\right)$ (30)

Here, SD implies standard deviation and $E_{1}$ reveals mean.
(iv)
Kurtosis

Kurtosis [15] implies to peak of distribution of intensity value across mean that can be calculated as follows.

$\displaystyle E_{4}=\frac{1}{\left(\textit{SD}\right)^{4}}\sum\limits_{\textit% {bn}=1}^{\textit{LV}}{\left({\textit{bn}-E_{1}}\right)^{4}}\textit{PR}\,\left(% \textit{bn}\right)$ (31)

The considered statistical features for extraction is specified as $E_{t}$ , which is given by,

$\displaystyle E_{t}=\left\{{E_{1},E_{2},E_{3},E_{4}}\right\}$ (32)

The extracted shape and statistical features are jointly termed as $P_{2}$ , such that,

$\displaystyle P_{2}=\left\{{K_{t},E_{t}}\right\}$ (33)

3.4.4 GLCM texture features

Energy, dissimilarity, homogeneity, entropy and autocorrelation [13] are concerned GLCM features for FE.

(i)
Energy

Energy signifies to sum of the squared components in GLCM, which can be estimated utilizing following expression.

$\displaystyle H_{1}=\sum\limits_{p,\vartheta=0}^{{\rm K}-1}{\left({G_{p,% \vartheta}}\right)^{2}}$ (34)

Here, $G_{p,\vartheta}$ reveals component $p,\vartheta$ of normalized symmetrical GLCM.
(ii)
Dissimilarity

Dissimilarity evaluation enables for the principle similarities amongst segmentations developed by distinctive approaches and segmentations on diverse images. Dissimilarity can be formulated as follows,

$\displaystyle H_{2}=\textit{Dissimilarity}\,\,\left({\textit{qh},\textit{ds}}% \right)=\sum\nolimits_{\textit{ax}=1}^{\textit{AX}}{\textit{ff}_{\textit{qh},% \textit{ax}}\,\log_{\textit{ff}_{\textit{ds},\textit{ax}}}^{\textit{ff}_{% \textit{qh},\textit{ax}}}}$ (35)
(iii)
Homogeneity

Homogeneity illustrates to a measure that develops in an image with lesser contrast that is estimated by,

$\displaystyle H_{3}=\sum\nolimits_{p}{\sum\nolimits_{\vartheta}{G_{p,\vartheta% }}}/1+\,\left|{p-\vartheta}\right|$ (36)
(iv)
Entropy

Entropy is a measure utilized for ranking textures and it can be modeled employing below explained expression.

$\displaystyle H_{4}=\sum\limits_{p,\vartheta=0}^{{\rm K}-1}{-\ln\,\left({G_{p,% \vartheta}}\right)}\,G_{p,\vartheta}$ (37)
(v)
Autocorrelation

Autocorrelation is computed through digital image of the magnitude, which is formulated as shown below.

$\displaystyle H_{5}=\textit{Autocorrelation}\,\left({\textit{mm},\textit{nn}}% \right)=\sum\limits_{\textit{xx}}^{\rm E}{\sum\limits_{\textit{yy}}^{\rm K}{p% \,\left(\textit{xx,yy}\right)}}*p\,\left({\textit{xx}-\textit{mm},\textit{yy}-% \textit{nn}}\right)$ (38)

The extracted GLCM features are signified by $P_{3}$ , such a way that,

$\displaystyle P_{3}=\left\{{H_{1},H_{2},H_{3},H_{4},H_{5}}\right\}$ (39)

Thus, a feature vector is attained from FE stage that can be symbolized by $P_{t}$ , such that,

$\displaystyle P_{t}=P_{1}||\,P_{2}\,||\,P_{3}$ (40)

3.5 Infection level identification of TB utilizing Xception T-Cascade NNet

TB is a deadly disease in the developing countries, with an illness spreading by means of direct relationship or air. In spite of its severity, earlier infection level identification of TB through reliable approaches can enhance life of patients. Here, Xception T-Cascade NNet is introduced for infection level identification of TB. However, Xception T-Cascade NNet is an incorporation of Xception and cascade NFN.

Figure 4.

General model of Xception T-Cascade NNet for infection level identification of TB.

Figure 4 shows the general model of Xception T-Cascade NNet for infection level identification of TB. At first input sputum image $U_{t}$ is passed to Xception model. Next, outcome from Xception model as well as feature vector are fed to Xception T-Cascade NNet layer, wherein fusion and regression are executed. Finally, outcome from Xception T-Cascade NNet layer and input sputum image $U_{t}$ are subjected to Cascade NFN model, wherein infection level identified output is obtained.

3.5.1 Xception model

Figure 5.

Structural illustration of Xception.

Xception [17] is modeled as a concept based upon Inception component that develops associations of the cross-channels as well as spatial relationships amongst feature maps of CNN capable to be entirely decoupled. This Xception model is much robust and powerful than Inception modules. A considered sputum image $U_{t}$ is given as an input to Xception model. Figure 5 reveals structural illustration of Xception. The crucial compositions within Xception structure is expounded beneath.

(i)

Conv layer

Thereafter input layer, conv layer is applied within Xception structure to generate conv kernels for calculating diverse feature maps in order to reveal features of input image. The newer feature map is gathered by initial conv function with detection outcomes from conv kernels that are fed to compute activation operation. To generate individual feature map, conv kernels are classified into every regions of an input image. The varied conv kernels develop exact outcomes of feature maps. The location $\left({\varepsilon,\alpha}\right)$ upon a feature value in feature map as $u^{\text{th}}$ layer identifies $z^{\text{th}}$ as follows,

$\displaystyle X_{t}=\Upsilon\varsigma_{u}^{z}\,U_{t}+{\rm P}_{u}^{z}$ (41)

Here, $\Upsilon\varsigma_{u}^{z}$ specifies weight vector whereas ${\rm P}_{u}^{z}$ indicates bias value of $u^{\text{th}}$ filter of $z^{\text{th}}$ layer and $U_{t}$ implies input sputum image. All conv layers of Xception model are inserted along with the BN and ReLU activation function that can be modeled by,

$\displaystyle\textit{ReLU}\,\left({U_{t}}\right)=\max\,\left({U_{t},0}\right)$ (42)

where, $U_{t}$ symbolizes considered input sputum image. ReLU is linear for every positive values and it is zero for every negative values. Moreover, it is not a complicated math with non-linearity of network that is significant in CNN to identify non-linear features, which make quick convergence and excellent detections, with lesser overfitting.

(ii)

Depth-Wise Separable Conv Layer

The vital layer of Xception is depth-wise separable conv layer, which can decrease computation as well as model parameters that are arranged in depth and spatial sizes of the color channels. A depth-wise separable conv layer develops the filter to individual channel of an input image fixed to ${\rm N}$ and generates feature map for determining $W\!\!\!\!/_{F\!\!\!\!/}\times W\!\!\!\!/_{F\!\!\!\!/}\times{\rm N}$ and depth-wise conv utilizing single filter of an input channel evaluated by beneath equation.

$\displaystyle\hat{D\!\!\!\!/}_{u,\rho,o}=\sum\limits_{\varepsilon,\alpha,o}{% \hat{L\!\!\!\!/}_{\varepsilon,\alpha,o}}\times F\!\!\!\!/_{u+\varepsilon-1,% \rho+\alpha-1,o}$ (43)

Here, $\hat{D\!\!\!\!/}$ substitutes outcome of the feature map produced by $F\!\!\!\!/$ that is an input of feature map, $\hat{L\!\!\!\!/}$ mentions depth-wise conv kernel with a dimension of $W\!\!\!\!/_{u}\times{W\!\!\!\!/}_{u}\times{\rm N}$ . The $o^{\text{th}}$ filter in $\hat{L\!\!\!\!/}$ is employed to channel $o^{\text{th}}$ in $\hat{L\!\!\!/}$ for an evaluation of feature map’s outcome. A pixel location of conv kernel is imposes to $\varepsilon,\alpha$ and feature map’s pixel location is illustrated as $u,\rho$ .

Afterwards depth-wise conv layer, Xception employs BN, the following layers utilize max-pooling layer for reducing computation cost and assist to understand invariance as,

$\displaystyle F\!\!\!\!/_{o}=\max\textit{pooling}\,\left({F\!\!\!\!/_{% \varepsilon},\varsigma}\right)$ (44)

Here, $\varsigma$ allocates max-pooling filter. A feature map output $F\!\!\!\!/_{o}$ that is organized in shape dimension, wherein individual $F\!\!\!\!/_{o}$ stores a maximum value of $F\!\!\!\!/_{\varepsilon}$ in an input feature map.

(iii)

Residual connection

A residual connection is developed based on another CNN structure termed ResNet, wherein interior network uses identity shortcut associations directly to newest layers. A residual block identifies parameters like ${\rm M}_{\varepsilon}$ that is formulated by,

$\displaystyle I\!\!\!\!/\varsigma=f\,\left({I\!\!\!\!/\varsigma\,\left\{{{\rm M% }\varepsilon}\right\}}\right)+G\!\!\!\!/\varsigma$ (45)

Here, $G\!\!\!\!/\varsigma$ indicates input vector, $I\!\!\!/\varsigma$ reveals output vector of layers. An evaluation of $f\,\left({I\!\!\!/\varsigma\,\left\{{{\rm M}\varepsilon}\right\}}\right)$ represents residual mapping that is to be trained. The residual association is highly rapid to train a model.

(iv)

Swish activation function

In current years, development of swish activation operation has an effect in collaborated approaches among exhaustive as well as searching methods of the reinforcement learning. The swish is capable to enhance classification process over ReLU as expressed below.

$\displaystyle\textit{swish}\,\left({U_{t}}\right)=U_{t}\,.\,\textit{sigmoid}\,% \left({\varphi\,U_{t}}\right)$ (46)

Here, $\varphi$ implies per-channel learnable parameter, $U_{t}$ specifies input sputum image whereas $\textit{sigmoid}\,\left({\varphi\,U_{t}}\right)$ denotes computation of sigmoid operation. An outcome from Xception model is signified by $X_{t}$ .

3.5.2 Xception T-Cascade NNet layer

In Xception T-Cascade NNet layer, fusion process and regression process are accomplished by acquiring ${\rm A}_{t}$ as input, such that,

$\displaystyle{\rm A}_{t}=\left\{{X_{t},P_{t}}\right\}$ (47)

Here, $X_{t}$ indicates outcome from Xception model whereas $P_{t}$ mentions feature vector.

By Taylor concept [28], the expression can be specified as,

$\displaystyle a\,\left({\gamma+1}\right)=a\,\left(\gamma\right)+\frac{a^{% \prime}\,\left(\gamma\right)}{1\,!}$ (48)

$\displaystyle a^{\prime}\,\left(\gamma\right)=\frac{a\,\left(\gamma\right)-a\,% \left({\gamma-\varpi}\right)}{\varpi}$ (49)

Assume, $\varpi=1$ and substitute above formulation in Eq. (48), thus the equation becomes,

$\displaystyle a\,\left({\gamma+1}\right)=a\,\left(\gamma\right)+\frac{a\,\left% (\gamma\right)-a\,\left({\gamma-1}\right)}{1\,!\,}$ (50)

$\displaystyle a\,\left({\gamma+1}\right)=2\,.\,a\,\left(\gamma\right)-a\,\left% ({\gamma-1}\right)$ (51)

$\displaystyle Z_{t}=2\,.\,P_{t}-X_{t}$ (52)

Here, $P_{t}$ specifies feature vector and $X_{t}$ reveals outcome from prior layer.

$\displaystyle Z_{t}=2\,.\,P_{t}-\Upsilon\varsigma_{u}^{z}\,U_{t}+{\rm P}_{u}^{z}$ (53)

$\displaystyle Z_{t}=2\,.\,\sum\limits_{x=1}^{\nu}{T_{x}*P_{t\,\left(x\right)}}% -\Upsilon\varsigma_{u}^{z}\,U_{t}+{\rm P}_{u}^{z}$ (54)

where, $\nu$ denotes number of features and $Z_{t}$ symbolizes outcome from Xception T-Cascade NNet layer.

3.5.3 Cascade NFN model

Figure 6.

Framework of cascade NFN.

Cascade NFN [18] is about definite variety of the neo-fuzzy parts and most modern atypical adaptive training rules. A foremost attribute of developed system is its ability to maintain cascade intensification till desirable accuracy is acquired. Figure 6 illustrates framework of Cascade NFN. ${\rm H}_{t}$ is subjected as input to the Cascade NFN model, such that,

$\displaystyle{\rm H}_{t}=\left\{{U_{t},Z_{t}}\right\}$ (55)

Here, $U_{t}$ reveals input sputum image and $Z_{t}$ specifies outcome from Xception T-Cascade NNet layer

Neo-fuzzy neurons refer to training scheme of nonlinearity state with varied incoming images and resulting value to execute conversion.

$\displaystyle\hat{\chi}=\sum\limits_{v=1}^{i}{\Im_{v}\,\left({r_{v}}\right)}$ (56)

Here, $r_{v}$ implies part $v$ on an input vector $r=\left({r_{1},\ldots,r_{v},\ldots,r_{i}}\right)^{Q}\in R\!\!\!\!/^{i}$ whereas $\hat{\chi}$ depicts scalar outcome of NFN. NFN contain synapses mentioned as ${\rm Z}_{v}$ and major goal is to alter $v^{\text{th}}$ part entry of $r_{v}$ into,

$\displaystyle\Im_{v}\,\left({r_{v}}\right)=\sum\limits_{v=1}^{\lambda}{n_{\eta% \,v}\phi_{\eta\,v}}\left({r_{v}}\right)$ (57)

where, $n_{\eta\,v}$ describes synaptic weight signified by $\eta$ in nonlinearity synapse $v$ and $\eta=1,2,\ldots,\lambda$ , $v=1,2,\ldots,i;\,\phi_{\eta\,v}\,\left({r_{v}}\right)$ indicates membership connection $\eta$ in a nonlinearity synapse $v$ , which outcomes fuzzification functioning of the crisp part $r_{v}$ . With concern to this manner, NFN conducted conversion is modeled by,

$\displaystyle\hat{\chi}=\sum\limits_{v=1}^{i}{\sum\limits_{\eta=1}^{\gamma}{n_% {\eta\,v}\phi_{\eta\,v}}\left({r_{v}}\right)}$ (58)

A fuzzy inference criterion fulfilled by the same NFN has a categorization, when $r_{v}$ is $J_{\eta\,v}$ , therefore resultant is signified as $n_{\eta\,v}$ , $\eta=1,2,\ldots,\lambda$ that connect synapse carries out zero-order fuzzy inference. It is essentially presented to exploit triangular membership designs as a substitute to the activation mechanisms amid neo-fuzzy neurons. Utilizing cubic splines, a disclosed fault is eliminated as indicated beneath.

$\displaystyle\phi_{v\omega}\,\left({r_{v}}\right)=\left\{{\begin{array}[]{l}0.% 25\,\left({2+3\frac{2r_{v}-k_{v\omega}-k_{v,\omega-1}}{k_{v\omega}-k_{v,\omega% -1}}-\left({\frac{2r_{v}-k_{v\omega}-k_{v,\omega-1}}{k_{v\omega}-k_{v,\omega-1% }}}\right)^{3}}\right)\,,\,\,\,\,\,\,\,\,\,\,\,\,\,r_{v}\in\left[{k_{v,\omega-% 1};k_{v\omega}}\right]\,,\\ 0.25\,\left({2-3\frac{2r_{v}-k_{v,\omega+1}-k_{v\omega}}{k_{v,\omega+1}-k_{v% \omega}}-\left({\frac{2r_{v}-k_{v,\omega+1}-k_{v\omega}}{k_{v,\omega+1}-k_{v% \omega}}}\right)^{3}}\right)\,,\,\,\,\,\,\,\,\,\,\,\,\,\,r_{v}\in\left[{k_{v% \omega};k_{v,\omega+1}}\right]\\ \end{array}}\right.$ (59)

During particular time period, quantity of an input requests two neighboring operational relationships instantly. Regarding this concept, membership associations are consistently spaced, $v^{\text{th}}$ part of input $r_{v}$ make active entirely two neighbor operation processing, but their total sum is equals to one that can be expressed by,

$\displaystyle\phi_{\eta\,v}\,\left({r_{v}}\right)+\phi_{\eta+1,v}\,\left({r_{v% }}\right)=1$ (60)

$\displaystyle\Im_{v}\,\left({r_{v}}\right)=n_{\eta\,v}\phi_{\eta\,v}\,\left({r% _{v}}\right)+n_{\eta+1,v}\phi_{\eta+1,v}\,\left({r_{v}}\right)$ (61)

It take place satisfactorily to enhance fitting features of computation model with an advantage of distinct structural item mentioned to be expanded nonlinearity synapse ${\rm B}{\rm Z}_{v}$ . Also, it creates complete neo-fuzzy neurons, which consist of ${\rm B}{\rm Z}_{v}$ items as a substitution of true synapses ${\rm Z}_{v}$ . The expression of extra variables can be expressed by,

$\displaystyle I_{\eta\,v}\left({r_{v}}\right)=\phi_{\eta\,v}\left({r_{v}}% \right)\,\left({n_{\eta\,v}^{0}+n_{\eta\,v}^{1}\,r_{v}+n_{\eta\,v}^{2}r_{v}^{2% }+\ldots+n_{\eta\,v}^{\partial}r_{v}^{\partial}}\right)$ (62)

$\displaystyle\Im_{v}\,\left({r_{v}}\right)=\sum\limits_{\eta=1}^{\lambda}{\phi% _{r_{v}}\,\left({r_{v}}\right)}\,\left({n_{\eta\,v}^{0}+n_{\eta\,v}^{1}\,r_{v}% +n_{\eta\,v}^{2}r_{v}^{2}+\ldots+n_{\eta\,v}^{\partial}r_{v}^{\partial}}\right)$ (63)

$\displaystyle\Im_{v}\,\left({r_{v}}\right)=n_{1v}^{0}\phi_{1v}\left({r_{v}}% \right)+n_{1v}^{1}\phi_{1v}\,\left({r_{v}}\right)+\ldots+n_{1v}^{\partial}o_{v% }^{\partial}\partial_{1v}\,\left({r_{v}}\right)+n_{2v}^{0}\phi_{2v}\left({r_{v% }}\right)+\ldots+n_{2v}^{\partial}r_{v}^{\partial}\phi_{2v}\,\left({r_{v}}% \right)+\ldots+n_{\lambda v}^{\partial}r_{v}^{\partial}\phi_{\lambda v}\,\left% ({r_{v}}\right)$ (64)

$\displaystyle n_{v}=\left({n_{1v}^{0},n_{1v}^{1},\ldots,n_{1v}^{\partial},n_{2% v}^{0},\ldots,n_{2v}^{0},\ldots,n_{\lambda v}^{\partial}}\right)^{Q}$ (65)

$\displaystyle\tilde{\phi}_{v}\,\left({r_{v}}\right)=\left({\phi_{1v}\left({r_{% v}}\right),r_{v}\phi_{1v}\left({r_{v}}\right),\ldots,r_{v}^{\partial}\phi_{1v}% \left({r_{v}}\right),\phi_{2v}\left({r_{v}}\right),r_{v}^{\partial}\phi_{2v}% \left({r_{v}}\right),r_{v}^{\partial}\phi_{\lambda v}\left({r_{v}}\right)}% \right)^{Q}$ (66)

It can be formulated by,

$\displaystyle\Im_{v}\,\left({r_{v}}\right)=n_{v}^{Q}\tilde{\phi}_{v}\left({r_{% v}}\right)$ (67)

$\displaystyle\hat{\xi}=\sum\limits_{v=1}^{\aleph}{\Im_{v}\,\left({r_{v}}\right% )=}\sum\limits_{v=1}^{\aleph}{n_{v}\tilde{\phi}_{v}\,\left({r_{v}}\right)=}n^{% Q}\tilde{\phi}\left(r\right)$ (68)

Here, $\tilde{\phi}\left(r\right)=\left({\tilde{\phi}_{1}^{Q}\left({r_{1}}\right),% \ldots,\tilde{\phi}_{1}^{Q}\left({r_{v}}\right),\ldots,\tilde{\phi}_{i}^{Q}% \left({r_{i}}\right)}\right)^{Q}$ whereas $\tilde{n}^{Q}=\left({n_{1}^{Q},\ldots,n_{v}^{Q},\ldots,n_{i}^{Q}}\right)^{Q}$ . The resultant of Cascade NFN model can be presented as,

$\displaystyle N_{t}=n_{\eta\,v}^{0}+n_{\eta\,v}^{1}r_{v}+\ldots+n_{\eta\,v}^{% \partial}r_{v}^{\partial},\,\,\,\,\,\eta=1,2,\ldots,\lambda$ (69)

where, $n_{\eta}$ refers to an output from prior layer, such that $n_{\eta}=2\,.\,\sum\limits_{x=1}^{\nu}{T_{x}*P_{t\,\left(x\right)}}-\Upsilon% \varsigma_{u}^{z}\,U_{t}+{\rm P}_{u}^{z}$ and $N_{t}$ indicates infection level identified output from the Cascade NFN model.

4. Results and discussion

Xception T-Cascade NNet presented for infection level identification of TB attained good results for performed analysis that are mentioned in this part.

4.1 Experiment setup

Xception T-Cascade NNet modeled for infection level identification of TB is implemented in PYTHON tool.

4.2 Dataset description

Tuberculosis Image Dataset [9] contains overall information regarding TB. It comprises of 928 sputum images along with bacilli about 3734 in the bounding boxes and the file is about 478 MB.

4.3 Experimentation outcomes

Figure 7.

Experimental resultants, a) Input image, b) Gray image, c) Pre-processed image, d) Segmented image.

The experimental resultants of Xception T-Cascade NNet are demonstrated in Fig. 7. Figure 7(a) exhibits input image, Fig. 7(b) reveals gray image, Fig. 7(c) delineates pre-processed image, Fig. 7(d) specifies segmented image.

4.4 Evaluation metrics

Accuracy, FNR, FPR, TNR and TPR are the estimation measures taken into consideration to assess Xception T-Cascade NNet.

4.4.1 Accuracy

Accuracy is a metric employed for evaluating classifiers that calculate overall samples classified exactly as modeled beneath.

$\displaystyle\textit{ACC}=\frac{\textit{TP}_{\textit{tp}}+\textit{TN}_{\textit% {tn}}}{\textit{TP}_{\textit{tp}}+\textit{TN}_{\textit{tn}}+\textit{FP}_{% \textit{fp}}+\textit{FN}_{\textit{fn}}}$ (70)

Here, $\textit{TP}_{\textit{tp}}$ specifies true positive, $\textit{FP}_{\textit{fp}}$ implies false positive, $\textit{TN}_{\textit{tn}}$ denotes true negative and $\textit{FN}_{\textit{fn}}$ mentions false negative.

4.4.2 FNR

FNR illustrates to a percentage of positive cases that are inexactly predicted as negative that is computed by,

$\displaystyle\textit{FNR}=\frac{\textit{FN}_{\textit{fn}}}{\textit{TP}_{% \textit{tp}}+\textit{FN}_{\textit{fn}}}$ (71)

4.4.3 FPR

FPR is computed as a ratio amid count of negative instances wrongly classified as positive, which is estimated by,

$\displaystyle\textit{FPR}=\frac{\textit{FP}_{\textit{fp}}}{\textit{FP}_{% \textit{fp}}+\textit{TN}_{\textit{tn}}}$ (72)

4.4.4 TNR

TNR is a probability of negative examination outcome, trained on individual actually being negative and it is evaluated as,

$\displaystyle\textit{TNR}=\frac{\textit{TN}_{\textit{tn}}}{\textit{TN}_{% \textit{tn}}+\textit{FP}_{\textit{fp}}}$ (73)

4.4.5 TPR

TPR defines as a probability of positive examination outcome, trained on individual exactly being positive, which is given as,

$\displaystyle\textit{TPR}=\frac{\textit{TP}_{\textit{tp}}}{\textit{TP}_{% \textit{tp}}+\textit{FN}_{\textit{fn}}}$ (74)

4.5 Performance analysis

Figure 8.

Performance analysis of Xception T-Cascade NNet, a) Accuracy, b) FNR, c) FPR, d) TNR, e) TPR.

Figure 8 mentions performance estimation of Xception T-Cascade NNet by altering percentages of training data with diverse epochs. When training data $=$ 90%, acquired performance values by Xception T-Cascade NNet are described as follows. Assessment of Xception T-Cascade NNet according to accuracy is demonstrated in Fig. 8(a). Xception T-Cascade NNet obtained 0.734, 0.773, 0.804, 0.853 and 0.888 of accuracy with 20, 40, 60, 80 and 100 epochs. Figure 8(b) specifies evaluation of Xception T-Cascade NNet regarding FNR. FNR achieved by Xception T-Cascade NNet is 0.262 with 20 epoch, 0.218 with 40 epoch, 0.197 with 60 epoch, 0.146 with 80 epoch and 0.094 with 100 epoch. Analysis of Xception T-Cascade NNet concerning FPR is expounded in Fig. 8(c). With epochs $=$ 20, 40, 60, 80 and 100, Xception T-Cascade NNet attained FPR of 0.263, 0.218, 0.196, 0.147 and 0.101. Figure 8(d) illustrates estimation of Xception T-Cascade NNet in terms of TNR. Xception T-Cascade NNet acquired TNR about 0.738, 0.782, 0.803, 0.854 and 0.906 with epochs 20, 40, 60, 80 and 100. Evaluation of Xception T-Cascade NNet in respective of TPR is revealed in Fig. 8(e). TPR acquired by Xception T-Cascade NNet with epochs 20, 40, 60, 80 and 100 is 0.737, 0.782, 0.804, 0.853 and 0.899.

4.6 Comparative techniques

AFC-Deep CNN [1], GDT-DBN [2], GFNN [4] and CNN [6] are compared with Xception T-Cascade NNet to show its effectiveness for infection level identification of TB.

4.7 Comparative evaluation

The estimation of Xception T-Cascade NNet is executed with relation to estimation metrics by changing percentages of training data and values of k fold.

4.7.1 Analysis in relation to training data

Figure 9.

Comparative analysis regarding training data, a) Accuracy, b) FNR, c) FPR, d) TNR, e) TPR.

Evaluation of Xception T-Cascade NNet based upon concerned metrics by changing training data is explicated in Fig. 9. Here, values achieved for training data $=$ 90% by considered schemes are illustrated. Figure 9(a) presents an estimation of Xception T-Cascade NNet with relation to accuracy. Xception T-Cascade NNet acquired 0.885 of accuracy whereas AFC-Deep CNN, GDT-DBN, GFNN and CNN attained 0.728, 0.785, 0.805 and 0.853, revealing enhancement of performance about 17.740%, 11.307%, 9.045% and 3.518%. Estimation of Xception T-Cascade NNet in accordance to FNR is explained in Fig. 9(b). FNR obtained by Xception T-Cascade NNet is 0.092 while FNR acquired by AFC-Deep CNN is 0.262, GDT-DBN is 0.216, GFNN is 0.196 and CNN is 0.143. Figure 9(c) delineates assessment of Xception T-Cascade NNet in accordance to FPR. AFC-Deep CNN, GDT-DBN, GFNN and CNN obtained FPR about 0.261, 0.216, 0.191 and 0.147 whereas Xception T-Cascade NNet acquired 0.106. Evaluation of Xception T-Cascade NNet concerning TPR is depicted in Fig. 9(d). Xception T-Cascade NNet achieved TPR of 0.908 while TPR acquired by AFC-Deep CNN, GDT-DBN, GFNN and CNN are 0.738, 0.784, 0.804 and 0.857. It mentions improvement in performance about 18.711%, 13.718%, 11.446% and 5.610%. Figure 9(e) signifies analysis of Xception T-Cascade NNet by means of TPR. AFC-Deep CNN, GDT-DBN, GFNN and CNN attained TPR value about 0.739, 0.784, 0.809 and 0.853 whereas TPR achieved by Xception T-Cascade NNet is 0.894. This shows enhancing of performance about 17.365%, 12.318%, 9.524% and 4.521%.

4.7.2 Analysis in relation to k fold

Figure 10.

Comparative analysis regarding k fold, a) Accuracy, b) FNR, c) FPR, d) TNR, e) TPR.

Figure 10 interprets assessment of Xception T-Cascade NNet in regard to concerned metrics by varying k fold. Here, values acquired by techniques while k fold is 9 are discussed. Estimation of Xception T-Cascade NNet with respective of accuracy is indicated in Fig. 10(a). Xception T-Cascade NNet attained accuracy of 0.880 while accuracy achieved by AFC-Deep CNN, GDT-DBN, GFNN and CNN are 0.734, 0.781, 0.801 and 0.858. This specifies enhancement in performance by 16.578%, 11.242%, 8.958% and 2.483%. Figure 10(b) presents evaluation of Xception T-Cascade NNet considering FNR. AFC-Deep CNN, GDT-DBN, GFNN and CNN attained FNR of 0.253, 0.218, 0.195 and 0.146 whereas Xception T-Cascade NNet obtained 0.101. Analysis of Xception T-Cascade NNet based upon FPR is described in Fig. 10(c). FPR Cquired by Xception T-Cascade NNet is 0.117 while FPR obtained by AFC-Deep CNN is 0.264, GDT-DBN is 0.219, GFNN is 0.199 and CNN is 0.147. Figure 10(d) reveals analysis of Xception T-Cascade NNet in accordance to TNR. AFC-Deep CNN, GDT-DBN, GFNN and CNN achieved TNR value of 0.747, 0.782, 0.805 and 0.854 while TNR obtained by Xception T-Cascade NNet is 0.899. This demonstrates performance enhancement about 16.897%, 13.009%, 10.461% and 4.998 %. Figure 10(e) implies assessment of Xception T-Cascade NNet with regarding to TPR. Xception T-Cascade NNet attained 0.883 of TPR whereas AFC-Deep CNN, GDT-DBN, GFNN and CNN acquired 0.736, 0.781, 0.801 and 0.853, elucidating improvement in performance about 16.660%, 11.574%, 9.274% and 3.393%.

4.8 Comparative discussion

Table 1
Comparative discussion of Xception T-Cascade NNet

Setups	Metrics/ methods	AFC-deep CNN	GDT-DBN	GFNN	CNN	Proposed Xception T-Cascade NNet
Training data $=$ 90%	Accuracy	72.8%	78.5%	80.5%	85.3%	88.5%
	TNR	73.8%	78.4%	80.4%	85.7%	90.8%
	TPR	73.9%	78.4%	80.9%	85.3%	89.4%
	FNR	0.262	0.216	0.196	0.143	0.092
	FPR	0.261	0.216	0.191	0.147	0.106
K fold $=$ 9	Accuracy	73.4%	78.1%	80.1%	85.8%	88%
	TNR	74.7%	78.2%	80.5%	85.4%	89.9%
	TPR	73.6%	78.1%	80.1%	85.3%	88.3%
	FNR	0.253	0.218	0.195	0.146	0.101
	FPR	0.264	0.219	0.199	0.147	0.117

Xception T-Cascade NNet achieved finest resultants while executing assessments regarding metrics by comparing with AFC-Deep CNN, GDT-DBN, GFNN and CNN. Table 1 exhibits discussion table of acquired values by considered approaches. It can be accepted that, Xception T-Cascade NNet attained maximum accuracy, TNR and TPR about 88.5%, 90.8% and 89.4% as well as minimum FNR and FPR about 0.092 and 0.106 for training data $=$ 90%. Thus, Xception T-Cascade NNet is superiorly performed for infection level identification of TB.

5. Conclusion

TB is a severe illness that requires earlier diagnosing for controlling this disease. The sputum image assessment is usual manual approach utilized for TN identification, though present sample evaluation methods are highly tedious, worse specificity, time-consuming and need more skilled personnel. In this research, Xception T-Cascade NNet is devised for infection level identification of TB utilizing sputum images. At first, input sputum image is obtained from dataset. Then, input sputum image is pre-processed by denoising and histogram equalization utilizing CLAHE. After that, SegNet is used for bacilli segmentation and it is tuned by WSO. Thereafter, suitable features like DCT with ALDP, shape features such as area, coverage, length and density, statistical features specifically variance, kurtosis, mean and skewness as well as GLCM texture features namely autocorrelation, energy, homogeneity, dissimilarity and entropy are extracted. At last, infection level identification of TB is carried out employing Xception T-Cascade NNet. The Xception T-Cascade NNet is introduced by combining Xception and cascade NFN with Taylor concept. Furthermore, Xception T-Cascade NNet obtained high accuracy, TNR and TPR of 88.5%, 90.8% and 89.4% as well as low FNR and FPR about 0.092 and 0.106 while training data is 90%. As a future task, microbial factors will be identified to acquire more absolute understanding of TB emergence.

References

Chithra

Jagatheeswari

. Severity detection and infection level identification of tuberculosis using deep learning. International Journal of Imaging Systems and Technology. 2020; 30(4): 994-1011.

Mithra

Emmanuel

. Gaussian model based hybrid technique for infection level identification in TB diagnosis. Journal of King Saud University-Computer and Information Sciences. 2021; 33(8): 988-998.

Samuel

Baskaran

. Design to automate the detection and counting of Tuberculosis (TB) bacilli. arXiv preprint arXiv:2105.11432. 2021.

Mithra

Emmanuel

. GFNN: Gaussian-Fuzzy-Neural network for diagnosis of tuberculosis using sputum smear microscopic images. Journal of King Saud University-Computer and Information Sciences. 2021; 33(9): 1084-1095.

Kurmi

Chaurasia

Goel

Joshi

Kapoor

. Tuberculosis bacteria analysis in acid fast stained images of sputum smear. Signal, Image and Video Processing. 2021; 15: 175-183.

Panicker

Pawan

Rajan

Sabu

. A lightweight convolutional neural network model for tuberculosis bacilli detection from microscopic sputum smear images. Machine learning for healthcare applications. 2021: 343-351.

Aung

Faksri

Sangka

Tomanakan

Namwat

. Detection of mycobacterium tuberculosis complex in sputum samples using droplet digital PCR targeting Mpt64. Pathogens. 2023; 12(2): 345.

Sirohi

Lall

Yenishetti

Panat

Kumar

. Development of a Machine learning image segmentation-based algorithm for the determination of the adequacy of Gram-stained sputum smear images. Medical Journal Armed Forces India. 2022; 78(3): 339-344.

Tuberculosis Image Dataset is taken from https//www.kaggle.com/datasets/saife245/tuberculosis-image-datasets. accessed on February 2024.

10.

Hiroyasu

Hayashinuma

Ichikawa

Yagi

. Preprocessing with image denoising and histogram equalization for endoscopy image analysis using texture analysis. In proceedings of 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), IEEE. 2015 August; 789-792.

11.

Badrinarayanan

Kendall

Cipolla

. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence. 2017; 39(12): 2481-2495.

12.

Braik

Hammouri

Atwan

Al-Betar

Awadallah

. White shark optimizer: A novel bio-inspired meta-heuristic algorithm for global optimization problems. Knowledge-Based Systems. 2022; 243: 108457.

13.

Dinesh

. Texture Image Classification Using Hybridization of Glcm And Glwm Features. International Journal of Scientific and Technology Research. 2020; 9(4).

14.

Shabat

Tapamo

. Angled local directional pattern for texture analysis with an application to facial expression recognition. IET Computer Vision. 2018; 12(5): 603-608.

15.

Malik

Baharudin

. The statistical quantized histogram texture features analysis for image retrieval based on median and laplacian filters in the dct domain. The International Arab Journal of Information Technology. 2013; 10(6): 1-9.

16.

Chilakala

Kishore

. Optimal deep belief network with opposition-based hybrid grasshopper and honeybee optimization algorithm for lung cancer classification: A DBNGHHB approach. International Journal of Imaging Systems and Technology. 2021; 31(3): 1404-1423.

17.

Jinsakul

Tsai

. Enhancement of deep learning in image classification performance using xception with the swish activation function for colorectal polyp preliminary screening. Mathematics. 2019; 7(12): 1170.

18.

Bodyanskiy

Tyshchenko

. A hybrid cascade neuro-fuzzy network with pools of extended neo-fuzzy neurons and its deep learning. International Journal of Applied Mathematics and Computer Science. 2019; 29(3).

19.

Evora

LHRA

Seixas

Kritski

. Artificial neural network models for diagnosis support of drug and multidrug resistant tuberculosis. In proceedings of 2015 Latin America Congress on Computational Intelligence (LA-CCI), IEEE. 2015 October; 1-5.

20.

Radzi

RUKRM

Mansor

Johari

. Review of mycobacterium tuberculosis detection. In proceedings of 2011 IEEE Control and System Graduate Research Colloquium, IEEE. 2011 June; 189-192.

21.

Kant

Srivastava

. Towards automated tuberculosis detection using deep learning. In proceedings of 2018 IEEE Symposium Series on Computational Intelligence (SSCI), IEEE. 2018 November; 1250-1253.

22.

Sheeba

Thamburaj

Mammen

Nithish

Karthick

. Detection of overlapping tuberculosis bacilli in sputum smear images. In proceedings of 7th WACBE World Congress on Bioengineering 2015: 6th to 8th July, 2015, Singapore, Springer International Publishing. 2015; 54-56.

23.

Makkapati

Agrawal

Acharya

. Segmentation and classification of tuberculosis bacilli from ZN-stained sputum smear images. In proceedings of 2009 IEEE International Conference on Automation Science and Engineering, IEEE. 2009 August; 217-220.

24.

Nayak

Shenoy

Galigekere

. A new algorithm for automatic assessment of the degree of TB-infection using images of ZN-stained sputum smear. In proceedings of 2010 International Conference on Systems in Medicine and Biology, IEEE. 2010 December; 294-299.

25.

Patel

Douglas

. Creating a virtual slide map from sputum smear images for region-of-interest localisation in automated microscopy. Computer Methods and Programs in Biomedicine. 2012; 108(1): 38-52.

26.

Priest

Goodfellow

Todd

. A numerical classification of the genus Bacillus. Microbiology. 1988; 134(7): 1847-1882.

27.

Lessa

Marengoni

. Applying artificial neural network for the classification of breast cancer using infrared thermographic images. In Computer Vision and Graphics: International Conference, ICCVG 2016; Warsaw, Poland, September 19–21, 2016, Proceedings, Springer International Publishing. 2016; 8: 429-438.

28.

Mangai

Sankar

Alagarsamy

. Taylor series prediction of time series data with error propagated by artificial neural network. International Journal of Computer Applications. 2014; 89(1).

29.

Tang

. Segnet-based gland segmentation from colon cancer histology images. In proceedings of 2018 33rd Youth Academic Annual Conference of Chinese Association of Automation (YAC), IEEE. 2018 May; 1078-1082.