Received 19 July 2023, accepted 24 August 2023, date of publication 29 August 2023, date of current version 7 September 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3309711

Biomedical Image Analysis for Colon and Lung

Cancer Detection Using Tuna Swarm Algorithm
With Deep Learning Model
MARWA OBAYYA 1 , MUNYA A. ARASI 2 , NUHA ALRUWAIS 3, RAED ALSINI 4,

ABDULLAH MOHAMED5 , AND ISHFAQ YASEEN 6

1 Department of Biomedical Engineering, College of Engineering, Princess Nourah Bint Abdulrahman University, Riyadh 11671, Saudi Arabia
2 Department of Computer Science, College of Science and Arts at Rijal Almaa, King Khalid University, Abha 62529, Saudi Arabia
3 Department of Computer Science and Engineering, College of Applied Studies and Community Services, King Saud University, Riyadh 11495, Saudi Arabia
4 Department of Information Systems, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia
5 Research Centre, Future University in Egypt, New Cairo 11845, Egypt
6 Department of Computer and Self Development, Preparatory Year Deanship, Prince Sattam Bin Abdulaziz University, Al-Kharj 16278, Saudi Arabia

Corresponding authors: Munya A. Arasi ([email protected]) and Ishfaq Yaseen ([email protected])

The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University for funding this work through large
group Research Project under grant number (RGP2/61/44). Princess Nourah bint Abdulrahman University Researchers Supporting Project
number (PNURSP2023R203), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Research Supporting Project number
(RSPD2023R608), King Saud University, Riyadh, Saudi Arabia. This study is supported via funding from Prince Sattam bin Abdulaziz
University project number (PSAU/2023/R/1444). This study is partially funded by the Future University in Egypt (FUE).

ABSTRACT The domain of Artificial Intelligence (AI) is made important strides recently, leading to devel-
opments in several domains comprising biomedical diagnostics and research. The procedure of AI-based
systems in biomedical analytics takes opened up novel avenues for the progress of disease analysis, drug
discovery, and treatment. Cancer is the second major reason of death worldwide; around one in every six peo-
ple pass away suffering from it. Among several kinds of cancers, the colon and lung variations are the most
frequent and deadliest ones. Initial detection of conditions on both fronts significantly reduces the probability
of mortality. Deep learning (DL) and Machine learning (ML) systems are exploited to speed up such cancer
detection, permitting researchers to analyze a huge count of patients in a lesser time count and at a minimal
cost. This study develops a new Biomedical Image Analysis for Colon and Lung Cancer Detection using
Tuna Swarm Algorithm with Deep Learning (BICLCD-TSADL) model. The presented BICLCD-TSADL
technique examines the biomedical images for the identification and classification of colon and lung cancer.
To accomplish this, the BICLCD-TSADL technique applies Gabor filtering (GF) to preprocess the input
images. In addition, the BICLCD-TSADL technique employs a GhostNet feature extractor to create a
collection of feature vectors. Moreover, AFAO was executed to adjust the hyperparameters of the GhostNet
technique. Furthermore, the TSA with echo state network (ESN) classifier is utilized for detecting lung and
colon cancer. To demonstrate the more incredible outcome of the BICLCD-TSADL system, an extensive
experimental outcome is carried out. The comprehensive comparative analysis highlighted the greater
efficiency of the BICLCD-TSADL technique with other approaches with maximum accuracy of 99.33%.

INDEX TERMS Cancer, biomedical imaging, artificial intelligence, colon cancer, tuna swarm algorithm,
GhostNet.

I. INTRODUCTION In 2020, Colon and Lung cancers were expected to rank in

Cancer takes place because of the uncontrollable growth of the top 3 most common types of cancer, as per a statistical
abnormal cells from the body’s tissues or organs. Cancer study done in the US. Lung tumours may occur concurrently
cells may follow in distinct tissues or organs of the body [1]. with colon cancer; almost 17% of cases of these two tumours
arise at the same time [2]. Also, there is a risk of tumour cells
The associate editor coordinating the review of this manuscript and spreading among 2 organs if an initial diagnosis is lacking.
approving it for publication was Humaira Nisar . Smoking has a harmful effect on the growth of lung tumours,
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME 11, 2023 94705
 M. Obayya et al.: BICLCD-TSADL Model

and it can be considered that an unaware diet leads to the This study develops a new Biomedical Image Anal-
growth of colon cancer [3]. Thus, lung tumour is the second ysis for Colon and Lung Cancer Detection using Tuna
cancer with colon cancer. In simple, a patient can affect by Swarm Algorithm with Deep Learning (BICLCD-TSADL)
both colon and lung cancer simultaneously. Therefore, it is model. The BICLCD-TSADL technique applies Gabor fil-
dynamic to examine both cancer types in patients and to tering (GF) to preprocess the input images. In addition,
detect them in advance [4]. the BICLCD-TSADL technique employs a GhostNet feature
The common symptoms are muscle pain, fatigue, cough, extractor to make a collection of feature vectors. Moreover,
etc., follow by various kinds of syndromes [5]. Some of AFAO can be executed to adjust the hyperparameters of
the frequently used radiographic imaging methods are ultra- the GhostNet technique. Furthermore, the TSA with echo
sound, mammography, histopathological imaging, computed state network (ESN) classifier is utilized for detecting lung
tomography (CT), magnetic resonance imaging (MRI), and and colon cancer. To demonstrate the greater efficiency of
positron emission tomography (PET) for cancer detection [6]. the BICLCD-TSADL algorithm, an extensive experimental
Of these, histopathology images comprising phenotypic data outcome is carried out. In short, the key contributions of the
are vital for the evaluation and diagnosis of cancer diseases. study are listed as follows.
Manual examination of these medical images by profession- • Develop an automated colon and lung cancer detection
als is a difficult and delicate task. Hence, it also necessitates model, comprising GF preprocessing, GhostNet feature
a strong focus and time-consuming task [7]. Likewise, the extraction, AFAO-based hyperparameter tuning, ESN
recognitionofcasesismoredifficultintheevent of initial iden- classification, and TSA-based parameter optimization.
tification; the symptoms are difficult and vague to identify. To the best of our knowledge, the BICLCD-TSADL
Likewise, identifying tumours consumes a lot of time and technique never existed in the literature.
is dependent on different opinions of physicians in the initial • Employ AFAO with the GhostNet model for the feature
stages [8]. A distinct domain of healthcare can sort out these extraction process, which contributes to the accurate
difficulties. Artificial intelligence (AI) methods are utilized in representation of colon and lung cancer image data.
the medical domain, like early identification of health disas- • Employ the ESN model to effectively learn and classify
ters, biomedical image, and disease forecast [9]. DL methods cancerous and non-cancerous patterns in colon and lung
are highly capable of analyzing data from anatomical repre- cancer images, contributing to accurate cancer detection.
sentations, high-dimensional images, and videos. Likewise, • Hyperparameter optimization process using AFAO and
DL methods extract hidden characteristics and features from TSA helps to improve the cancer detection performance
medical images that are invisible to the naked eye for the of the BICLCD-TSADL model.
initial identification of cancers and discrimination among
their phases [10]. In this study, due to the same features of II. RELATED WORKS
abnormal cells in the initial phases, numerous hybrid sys- In [11], CNN methods were used to examine imaging data of
tems have been established with extraction features by mixed colon cells. CNN with average pooling and max pooling layers
techniques. and MobileNetV2 methods have been applied for colon cell
Multi-level hyperparameters, also called hierarchical imageclassification.Wahid et al. [12]introducedaCADmech-
hyperparameters, can be employed for controlling the config- anism through the CNN to find colon, lung and cancer tissues.
uration and behavior of complex DL models. These hyper- Heretheauthor,namelyResNet18,ShuffleNet V2,andGoogle
parameters are organized in a hierarchical manner, where Net,inadditiontoonesimplecustomizedCNNmodel,usesthree
higher-level hyperparameters affects the values or ranges of pre-trained CNN models. Kumar et al. [13] devise 4 CNN,
lower-level hyperparameters. It offers a structured way of namely,3-block,baseline,3-block,and2-blockCNNswithdata
managing and optimizing the settings for each component, augmentation, for classifying colon tissue histopathological
making it easier to maintain, understand, and tune the over- images(HPI).TofourCNN,theHPIwasfedasinput.
all model. When optimizing hyperparameters for a feature Garg and Garg [14] purposes of using and altering the
extractor and classifier, it is important to consider the hier- present pre-training CNN-related technique for finding colon
archical relationship between these components. The feature and lung cancer with the help of HPI, including bet-
extractor is responsible for transforming the input data into ter augmentation approaches. Here, 8 various pre-trained
a meaningful feature representation, while the classifier uses CNN methods like MobileNet, InceptionResNetV2, VGG16,
these features to make predictions. The learning rate is a crit- InceptionV3, DenseNet169, ResNet50, NASNetMobile, and
ical hyperparameter in many feature extractor architectures, Xception were trained on LC25000 data. Adu et al. [15]
such as neural networks. It controls the step size at which present a novel DHS-CapsNet abbreviated as dual horizontal
the model updates its internal parameters during training. For squash capsule network for categorizing the colon and lung
classifiers like neural networks, the number of hidden units in cancers on HPI. DHS-CapsNet is a new horizontal squash
the classifier’s layers significantly impacts the model’s capac- (HSquash) function develop for encoder feature fusion (EFF).
ity to learn complex patterns. Tuning this hyperparameter While a squash function, HSquash is modelled to make sure
using random search or Bayesian optimization can help that vectors can be effectually squashed and makes sparsity
achieve the right balance between model complexity and for higher discriminative capsules for extracting significant
generalization. data from images with different backgrounds.
94706 VOLUME 11, 2023
M. Obayya et al.: BICLCD-TSADL Model

In [16], chooses the gene elements with maximum correla-

tion with tumour by Weighted Gene Co-expression Network
Analysis (WGCNA) derived specific genes to differential
expression outcomes utilizing the least absolute shrinkage
and selection operator (Lasso) technique and effectuated sur-
vival study followed by combining the genes from the compo-
nents with Lasso-extracting feature genes are joined to detect
colon cancer vs healthy controls with the help of DT, RF,
and SVM and colon cancer staging can be identified through
differently expressed genes for all stages. Sakr et al. [17]
present an innovative lightweight DL method that depends
on a CNN for potential colon recognition. In this technique,
the input HPI was normalised before presenting them as
CNN approach and the next colon tumour identification was
executed.
In [18], a classification method for colon cancer has been
generated as 2 classes such as polyps and adenocarcinomas.
This method was created utilizing the CNN approach with FIGURE 1. Workflow of the BICLCD-TSADL approach.
the MobileNet structure. Mohalder et al. [19] presented a DL
approach to forecasting CRC tumours in HPI. CNN system tion, and TSA-based parameter tuning. Fig. 1 represents the
employs to analyze difficult information. By CNN algorithm workflow of the BICLCD-TSADL approach.
it is examined our difficult tumor images to recognize abnor-
mal or suspicious tumor designs. The authors utilized the A. GF BASED PREPROCESSING
ReLU activation function in the hidden state and softmax Primarily, the BICLCD-TSADL technique preprocesses the
function in the resultant layer. Bhattacharya et al. [20] exam- GF system to eliminate the noise. It is dependent upon the
ine a framework in which either DL or meta-heuristic tech- employ of Gabor functions, which are a family of difficult
niques are utilized to predict colon or lung cancer, or both, functions which are utilized for representing local structures
in HPIs with near-perfect precision. In [21], a colon can- from the signal/image. The GF is a kind of linear filtering
cer classification method was established to support medical which utilizes a Gabor function as its kernel. This filter can
workers categorize 2 kinds of cancer colon adenocarcinomas be planned to capture data on the frequency and orientation
and benign colonic tissues. The classification approach uti- of local image features like textures and edges.
lizes CNN with the structure of VGG19, VGG16, ResNet152,
ResNet101, DenseNet201, MobileNetV2, and InceptionV3. B. OPTIMAL GHOSTNET FEATURE EXTRACTOR
In spite of the ML and DL models that existed in earlier For the feature extraction process, the GhostNet model is
studies, it is still needed to enhance lung and colon cancer used. A GhostNet-based DL method has been performed
classification performance. Because of the continual deepen- for extracting the feature from the image [22]. GhostNet
ing of the model, the number of parameters of DL models recommended an innovative Ghost model that produces addi-
also increases quickly which results in model overfitting. tional feature maps through a reasonable operation. This basic
At the same time, different hyperparameters have a signifi- neural network unit makes image feature with fewer compu-
cant impact on the efficiency of the CNN model. Particularly, tations and inputs. There exist 2 features to this implemen-
hyperparameters such as epoch count, batch size, and learning tation of the module. To construct a feature map with other
rate selection are essential to attain effectual outcomes. Since channels, GhostNet first implemented the typical convolu-
the trial and error method for hyperparameter tuning is a tion. Then, it implemented a simple operation to construct
tedious and erroneous process, metaheuristic algorithms can an additional mapping feature. Fig. 2 represents the structure
be applied. Therefore, in this work, we employ IAFO and of the GhostNet approach. Lastly, it concatenated numerous
TSA algorithms for the parameter selection of the GhostNet mapping features to make a novel output.
and ESN models respectively. Ghost bottleneck is the essential part of GhostNet that
comprised 2 ghost models. The procedure of making the M
mapping feature from the ghost module was formulated by
III. THE PROPOSED MODEL Eq. (1):
In this work, we have introduced a novel BICLCD-TSADL γ = X ∗f + b (1)
system for automated colon and lung cancer detec-
tion. The purpose of the BICLCD-TSADL technique is In Eq. (1), w and h denote the width and height of the
to investigate biomedical images for the detection and input, correspondingly; c shows the channel counts, b indi-
classification of colon and lung cancer. To accomplish this, cates the bias term, and ∗ denotes the convolution function;
the BICLCD-TSADL technique comprises GF-based prepro- f ∈Rc×k×k×m , X ∈ Rh×c×w is the convolutional kernel. k ∗ k
cessing, GhostNet feature extraction, AFAO, ESN classifica- indicates the convolution size f .

VOLUME 11, 2023 94707

 M. Obayya et al.: BICLCD-TSADL Model

step, and variable vectors are initialized. Next, every part

of the loop was updated frequently; still, the variable θz
converges. The biased 1st and 2nd moments are evaluated.
Next, the bias-corrected 1st and 2nd-moments can be esti-
mated. The predictive accuracy of GhostNet with the AFAO
algorithm was assessed against the testing set.

C. CANCER DETECTION USING ESN MODEL

For the cancer detection and classification process, the ESN
FIGURE 2. Structure of GhostNet model. model is used. Statistical techniques can be comparatively
easy to model but are insufficient forecasted power. The ML
Firstly, the W × H × C size-based input mapping feature technique needs several computations and proceeds a long-
was downscaled by the typical convolution. Then, a cheap time [24]. The combined learning system generates the model
linear operation is exploited on W ′ × H ′ × C feature maps design further difficulty that could not conducive to updating
with a k × k small kernel convolutional function that directly the model. The ESN can be regarded as the most efficient
produces a considerable quantity of ghost features. Even- approach to training RNNs that are the benefits of an easily
tually, the outcome of both procedures is merged to form trained procedure and short time consumption. It determines
an outcome mapping feature of dimensional W ′ × H ′ × c the u(n) = [u1 (n) · · · uk (n)]T as the input instance at time
that is similar to the original. The regular convolution and n.y(n) refers to the resultant equivalent to (n). An input matrix
linear alteration used in the Ghost Module permit to brilliant Win and the reservoir layer weighted matrix Ŵ are uniformly
preservation of novel features. Finally, the feature maps are distributed among [−1, 1] that maintained constant. In the
converted as the last 1280-dimension feature vector utilizing trained procedure, with the instances input, the reservoir layer
the convolutional and global average pooling layers. The can be upgraded utilizing the subsequent equation.
computation cost is lesser than classical convolution directly.  
To adjust the hyperparameters of the GhostNet model, the x (n + 1) = f Win u (n + 1) + Ŵ x (n) (8)
AFAO is used. The AFAO algorithm is a modified variant
whereas Ŵ refers to the reservoir state connection matrix,
of stochastic gradient descent (SGD) that is becoming ever
f represents the activation function inside the reservoir state
more popular in DL applications [23]. AFAO algorithm is a
that is generally obtained strictly as hyperbolic tangent func-
stochastic optimizer technique used to improve the feature
tions, and Win implies the input connection matrix. Based
for the training procedure of the GhostNet model. AFAO
on the above state of reservoir layers, the resultant ESN is
algorithm increases the machine’s efficacy and accelerates
computed utilizing the subsequent formula.
the learning process by letting the proposed method converge
fast. The hyperparameters of the GhostNet technique were y (n + 1) = fout (Wout x (n + 1)) (9)
repeatedly updated dependent on the training dataset. The
Wout stands for the resultant connection matrix, and fout
significant formula to determine the optimization algorithm
represents the resultant excitation function. In the trained
using the AFAO algorithm is given below.
method, the reservoir state can be gathered as a state matrix
gz = ∇θ fz (θz−1 ) (2) X. The last network resultant weighted Wout is computed
hz = β1 .hz−1 + (1 − β1 ) .gz (3) utilizing the subsequent formula.

vz = β2 .vz−1 + (1 − β2 ) .g2z (4) Wout =(X T X )−1 X T Y (10)

hz whereas The superscript T implies the matrix transpose and
ĥz = (5)
1 − β1z −1 denotes the inverse of a matrix. X defines the matrix
vz procedure of the input reservoir layer, and Y demonstrates the
v̂z = (6)
1 − β2z resultant matrix method. An optimum solution of the resultant
matrix is initiated utilizing least squares or MSE alone.
α · ĥz
θz = θz−1 − p (7)
v̂z + ε D. PARAMETER TUNING USING TSA
where f (θ ) represents the stochastic objective function, α Finally, the TSA is used to optimally select the parameters
denotes the size of the step, β1 and β2 characterize the expo- related to the ESN model. The TSA method begins the opti-
nential decay rate, and z represents the time-step. θ0 and θz mization process by randomly and uniformly generating the
correspondingly signify the first and last variable vectors, and initial population in the search range [25]. The TSA can be
hz and vz represent 1st and 2nd-moment vectors, correspond- mathematically formulated as follows:
ingly. θ̂Hz and v̂z denote bias-corrected moment estimates, and X1int = rand · (ub − lb) + lb, i = 1, 2, . . . , NP (11)
g2z shows the component-wise square.
The objective function, step size, and exponential decay In Eq. (11), Xiint
denotes the initial location of the i − th
rate are confirmed. The 1st and 2nd-moment vectors, time- individual, and Ub and lb show the upper as well as lower

94708 VOLUME 11, 2023

M. Obayya et al.: BICLCD-TSADL Model

boundaries of searching spaces correspondingly. NP indi- TABLE 1. Details of database.

cates the amount of tuna population. Typically, this value
affects the optimization rate of the TSA and is a uniformly
distributed random vector. Spiral foraging is the primary
foraging method of tuna schools that chases the prey by
making tight spirals. Along with chasing prey, schools of tuna
exchange data with all the others. Every tuna is sequenced and
strongly related; hence adjacent tuna share data. The spiral
foraging strategy can be mathematically expressed as:

α1 · Xbest
t t
− Xit +α2 ·Xit ,


 +β · Xbest
tuna are performed based on the probability allocation, and


 i = 1,
X1t+1 = (12) if the selection probability for the two foraging techniques
α t t i
− Xit +α2 ·Xt−1 ,



 1 · X best +β · Xbest is 1/2, then they are concurrently performed, and it can be
i = 2, 3, . . . , NP,


mathematically formulated as follows.
t
α1 = ∂ + (1 − a) · , −X1ι )+TF · 2 ·Xbest
 t t t
(13) X +(rand ·Xbest −Xit ),
tmax  best


T
 ifrand< 0.5,
α2 = (1 − a) − (1 − a) · , X1t+1 =
tmax
(14) 
 TF · p2
· X1ι ,

β = ebl · cos (2πb) ifrsnd≥ 0.5,

(15)
l = e3cos(tmax +1/t)−1)π) (16) (19)
 (t/tmax )
t
where best denotes the existing better individual (food), Xiι+1 p= 1− , (20)
tmax
denotes the i − th individuals of the t + 1 iteration, a indicates
the constant, defines to which extent the tuna follows the where TF denotes the random integer within [1, −1].
better individual and the prior individual at the initial phase, The fitness chosen is a vital feature in the TSA approach.
t and tmax indicates the existing and the maximum amount An encoder solution can be utilised to grow the ability of
of iterations, α1 and α2 denotes the weight coefficient which candidate solutions. Here, the accuracy value is the critical
controls the movement trends of the individual to the better factor utilized to propose a fitness function.
and the prior individuals, and b is a uniformly distributed
Fitness = max (P) (21)
random integer within [0,1]. Once the best individual could
TP
not find food, blindly following the optimum individual for- P= (22)
aging is not advantageous to group foraging. Hence, to assist TP + FP
every individual in having the best spatial search abilities, whereas FP signifies the false positive and TP denotes the
a reference point for spiral search should be given to produce true positive value.
a random coordinate in the search range, thereby allowing
TSA to have the best global exploration abilities, and it can IV. EXPERIMENTAL VALIDATION
be a mathematical equation by Eq. (17): The proposed model is simulated using Python 3.6.5 tool on
PC i5-8600k, GeForce 1050Ti 4GB, 16GB RAM, 250GB
α1 · Xrand
t t
−Xit +α2 ·Xit ,



 +β · Xrand SSD, and 1TB HDD. The parameter settings are given as

t=1′
 i = 1, follows: learning rate: 0.01, dropout: 0.5, batch size: 5, epoch
X1 = (17)
α t t
−Xit +α2 ·Xi−1
t
,



 1 · Xrand +β · Xrand count: 50, and activation: ReLU.
i = 2, 3, . . . , NP The experimental validation of the BICLCD-TSADL


algorithm was tested on the LC25000 database [26] com-
t
where Xrand denotes the random reference point in the prising five classes with 5000 samples under each class,
search ranges, TSA is typically explored extensively globally as depicted in Table 1. Fig. 3 depicts the sample of Lung
at an early stage and then slowly transitioned to accurate and colon images. The 5 classes are colon adenocarcino-
local exploitation. Thus, with the increasing amount of iter- mas, benign colonic tissues, lung adenocarcinomas, lung
ations, TSA slowly changes the reference point of spiral squamous cell carcinomas and benign lung tissues. For
foraging from a random individual at the beginning to an experimental validation, 70:30 of training/testing data.
optimum individual. The spiral foraging strategy can be Fig. 4 establishes the classifier outcomes of the
mathematically expressed by Eq. (18), as shown at the bottom BICLCD-TSADL system under the test database. Figs. 4a-4b
of the next page. represents the confusion matrix offered by the
Tuna chooses to spiral foraging along with parabolic coop- BICLCD-TSADL model on 70:30 of TRP/TSP. The figure
erative foraging. Tuna forms a parabola with the targeted food denoted that the BICLCD-TSADL approach has identified
as a reference to the Z -point. Tuna finds the targeted food by and classified all 5 class labels accurately. Likewise, Fig. 4c
searching around the parabola. Both foraging approaches of represents the PR investigation of the BICLCD-TSADL

VOLUME 11, 2023 94709

 M. Obayya et al.: BICLCD-TSADL Model

TABLE 2. Classifier outcome of BICLCD-TSADL system on 70% of TRP.

FIGURE 3. a) Lung Images b) Colon Images.

FIGURE 5. (a) Classification outcome on 70% of TRP and (b) Average

outcome on 70% of TRP.

The results indicate that the BICLCD-TSADL technique

reaches proficient results in each class. For instance,
on ColAd, the BICLCD-TSADL technique gains accuy ,
precn , recal , Fscore , and AUCscore of 99.15%, 97.31%,
98.44%, 97.87%, and 98.88%, respectively. Followed by,
on LunAd, the BICLCD-TSADL method attains accuy ,
precn , recal , Fscore , and AUCscore of 99.51%, 98.92%,
98.67%, 98.80%, and 99.20% correspondingly. Concur-
rently, on LunSC, the BICLCD-TSADL system reaches
accuy , precn , recal , Fscore , and AUCscore of 99.33%, 98.24%,
98.43%, 98.34%, and 99% correspondingly. Finally, the
BICLCD-TSADL technique reaches average accuy , precn ,
recal , Fscore , and AUCscore of 99.33%, 98.31%, 98.31%,
98.31%, and 98.95%, respectively.
In Table 3 and Fig. 6, the overall outcomes of the
BICLCD-TSADL system are examined on 30% of TSP.
The outcomes stated that the BICLCD-TSADL method-
ology reaches proficient outcomes under each class.
FIGURE 4. Classifier outcome of (a-b) Confusion matrices, (c) PR curve,
(d) ROC curve. For instance, on ColAd, the BICLCD-TSADL algorithm
gains accuy , precn , recal , Fscore , and AUCscore of 99.17%,
approach. The figures reported that the BICLCD-TSADL 97.38%, 98.64%, 98.01%, and 98.98% correspondingly.
approach has obtained higher PR performance under Afterwards, on LunAd, the BICLCD-TSADL technique
5 classes. Eventually, Fig. 4d illustrates the ROC investigation reaches accuy , precn , recal , Fscore , and AUCscore of 99.51%,
of the BICLCD-TSADL model. The figure stated that the 98.90%, 98.56%, 98.73%, and 99.15% correspondingly.
BICLCD-TSADL approach has led to able outcomes with Simultaneously, on LunSC, the BICLCD-TSADL algorithm
maximal ROC values under 5 classes. reaches accuy , precn , recal , Fscore , and AUCscore of 99.25%,
In Table 2 and Fig. 5, the overall outcomes of the 97.93%, 98.32%, 98.12%, and 98.90% correspondingly.
BICLCD-TSADL system are examined on 70% of TRP. At last, the BICLCD-TSADL approach obtains average

(
α1 · t
+ β · Xbest
t
− Xit + α2 · Xit ,


 Xbest i = 1, t
 if rand ≥
 α1 · t
+ β · Xbest
t
− Xit + α2 ·Xi−1
t
, i = 2, 3, . . . , NP,



Xbest tmax
X1t+1 ′ = ( (18)
α1 · t
+ β · Xrand
t
− Xit + α2 · Xit ,


 Xrand i = 1, t
if rand <


α1 · Xrand + β · Xrand − Xit + α2 ·Xi−1 , i = 2, 3, . . . , NP,
 t t

t tmax


94710 VOLUME 11, 2023

M. Obayya et al.: BICLCD-TSADL Model

TABLE 3. Classifier outcome of BICLCD-TSADL algorithm on 30% of TSP.

FIGURE 8. Loss curve of the BICLCD-TSADL approach.

TABLE 4. Comparative outcome of BICLCD-TSADL approach with recent

algorithms.

FIGURE 6. (a) Classification outcome on 30% of TSP and (b) Average

outcome on 30% of TSP.

RCNN, and DAELGNN models accomplish closer outcomes.

However, the BICLCD-TSADL technique gains effectual
outcomes with a maximum accuy of 99.33%, precn of
98.31%, recal of 98.31%, and Fscore of 98.31%. These results
show the improvement of the BICLCD-TSADL system in
medical image analysis. The proposed IAFO and TSA mod-
els choose the optimal values for the hyperparameters of
FIGURE 7. Accuracy curve of the BICLCD-TSADL approach. a given GhostNet and ESN models. Hyperparameters are
accuy , precn , recal , Fscore , and AUCscore of 99.31%, 98.28%, settings that are not learned during training but must be set
98.26%, 98.27%, and 98.92% correspondingly. before training. They can have a significant impact on the
Fig. 7 examines the accuracy of the BICLCD-TSADL performance of the model, and selecting the optimal values
technique under the training and validation process on the test can lead to better accuracy. By exploiting IAFO and TSA
dataset. The figure notifies that the BICLCD-TSADL method algorithms, the BICLCD-TSADL model can achieve even
reaches maximal accuracy values over increasing epochs. better results by focusing on the selection of the optimal
In addition, the increasing validation accuracy over training settings for the algorithm. These results ensured the improved
accuracy exhibits that the BICLCD-TSADL technique learns performance of the BICLCD-TSADL technique over other
efficiently on the test dataset. existing techniques.
The loss investigation of the BICLCD-TSADL algorithm V. CONCLUSION
at the time of training and validation is depicted on the In this article, we have developed a novel BICLCD-TSADL
test database in Fig. 8. The outcomes indicate that the algorithm for automated colon and lung cancer detection.
BICLCD-TSADL system reaches closer values of training The purpose of the BICLCD-TSADL technique is to inves-
and validation loss. It can be clear that the BICLCD-TSADL tigate biomedical images for the identification and clas-
algorithm learns efficiently on the test dataset. sification of colon and lung cancer. To accomplish this,
A comparative classification outcome of the the BICLCD-TSADL technique comprises GF-based prepro-
BICLCD-TSADL technique is made with recent approaches cessing, GhostNet feature extraction, AFAO, ESN classifica-
in Table 4 and Fig. 9 [27]. The outcomes depicted the tion, and TSA-based parameter tuning. The design of AFAO
ineffectual outcomes of the mSRC model, whereas the and TSA for parameter tuning techniques helps to accom-
ResNet50, CNN, and DL models obtain slightly improved plish enhanced classification performance. To demonstrate
performance. At the same time, the MPADL-LC3, Faster the greater performance of the BICLCD-TSADL algorithm,

VOLUME 11, 2023 94711

 M. Obayya et al.: BICLCD-TSADL Model

[8] A. H. Chehade, N. Abdallah, J.-M. Marion, M. Oueidat, and P. Chauvet,

‘‘Lung and colon cancer classification using medical imaging: A feature
engineering approach,’’ Phys. Eng. Sci. Med., vol. 45, no. 3, pp. 729–746,
Sep. 2022.
[9] M. Toğaçar, ‘‘Disease type detection in lung and colon cancer images
using the complement approach of inefficient sets,’’ Comput. Biol. Med.,
vol. 137, Oct. 2021, Art. no. 104827.
[10] I. Pacal, D. Karaboga, A. Basturk, B. Akay, and U. Nalbantoglu, ‘‘A com-
prehensive review of deep learning in colon cancer,’’ Comput. Biol. Med.,
vol. 126, Nov. 2020, Art. no. 104003.
[11] Z. Tasnim, S. Chakraborty, F. M. J. M. Shamrat, A. N. Chowdhury,
H. A. Nuha, A. Karim, S. B. Zahir, and M. M. Billah, ‘‘Deep learning
predictive model for colon cancer patient using CNN-based classification,’’
Int. J. Adv. Comput. Sci. Appl., vol. 12, no. 8, pp. 687–696, 2021.
[12] R. R. Wahid, C. Nisa’, R. P. Amaliyah, and E. Y. Puspaningrum, ‘‘Lung and
colon cancer detection with convolutional neural networks on histopatho-
logical images,’’ AIP Conf. Proc., vol. 2654, Feb. 2023, Art. no. 020020.
[13] A. Kumar, A. Vishwakarma, V. Bajaj, A. Sharma, and C. Thakur, ‘‘Colon
cancer classification of histopathological images using data augmenta-
FIGURE 9. Comparative outcome of BICLCD-TSADL approach with recent tion,’’ in Proc. Int. Conf. Control, Autom., Power Signal Process. (CAPS),
algorithms. Dec. 2021, pp. 1–5.
[14] S. Garg and S. Garg, ‘‘Prediction of lung and colon cancer through analysis
an extensive experimental outcome is carried out. A detailed of histopathological images by utilizing pre-trained CNN models with
comparative outcome highlighted the greater outcome of the visualization of class activation and saliency maps,’’ in Proc. 3rd Artif.
Intell. Cloud Comput. Conf., Dec. 2020, pp. 38–45.
BICLCD-TSADL technique with state-of-the-art approaches [15] K. Adu, Y. Yu, J. Cai, K. Owusu-Agyemang, B. A. Twumasi, and X.
with maximum accuracy of 99.33%. In future, the perfor- Wang, ‘‘DHS-CapsNet: Dual horizontal squash capsule networks for
mance of the proposed model can be improved by feature lung and colon cancer classification from whole slide histopathological
images,’’ Int. J. Imag. Syst. Technol., vol. 31, no. 4, pp. 2075–2092,
fusion and ensemble learning process. In addition, compu- Dec. 2021.
tation complexity of the proposed model can be investigated [16] Y. Su, X. Tian, R. Gao, W. Guo, C. Chen, C. Chen, D. Jia, H. Li, and X. Lv,
in future. ‘‘Colon cancer diagnosis and staging classification based on machine
learning and bioinformatics analysis,’’ Comput. Biol. Med., vol. 145,
ACKNOWLEDGMENT Jun. 2022, Art. no. 105409.
[17] A. S. Sakr, N. F. Soliman, M. S. Al-Gaashani, P. Pławiak, A. A. Ateya,
The authors extend their appreciation to the Deanship of
and M. Hammad, ‘‘An efficient deep learning approach for colon cancer
Scientific Research at King Khalid University for fund- detection,’’ Appl. Sci., vol. 12, no. 17, p. 8450, Aug. 2022.
ing this work through large group Research Project under [18] Y. G. C. Kelana, S. Rizal, and S. Saidah, ‘‘Classification of histopatholog-
grant number (RGP2/61/44). Princess Nourah bint Abdul- ical images of colon cancer using convolutional neural network method,’’
in Proc. Int. Conf. Comput. Sci., Inf. Technol. Eng. (ICCoSITE), Feb. 2023,
rahman University Researchers Supporting Project number pp. 821–826.
(PNURSP2023R203), Princess Nourah bint Abdulrahman [19] R. D. Mohalder, F. Bin Ali, L. Paul, and K. Hasan Talukder, ‘‘Deep
University, Riyadh, Saudi Arabia. Research Supporting learning-based colon cancer tumor prediction using histopathological
images,’’ in Proc. 25th Int. Conf. Comput. Inf. Technol. (ICCIT), Dec. 2022,
Project number (RSPD2023R608), King Saud University, pp. 629–634.
Riyadh, Saudi Arabia. This study is supported via funding [20] A. Bhattacharya, B. Saha, S. Chattopadhyay, and R. Sarkar, ‘‘Deep fea-
from Prince Sattam bin Abdulaziz University project number ture selection using adaptive β-Hill climbing aided whale optimization
algorithm for lung and colon cancer detection,’’ Biomed. Signal Process.
(PSAU/2023/R/1444). Control, vol. 83, May 2023, Art. no. 104692.
REFERENCES [21] I. D. Irawati, I. Andrea Larasaty, and S. Hadiyoso, ‘‘Comparison of con-
volution neural network architecture for colon cancer classification,’’ Int.
[1] M. S. N. Raju and B. S. Rao, ‘‘Lung and colon cancer classification using J. Online Biomed. Eng. (iJOE), vol. 18, no. 3, pp. 164–172, Mar. 2022.
hybrid principle component analysis network-extreme learning machine,’’ [22] S. S. R. Bairaboina and S. R. Battula, ‘‘Ghost-ResNeXt: An effective deep
Concurrency Comput., Pract. Exper., vol. 35, no. 1, Jan. 2023. learning based on mature and immature WBC classification,’’ Appl. Sci.,
[2] M. A. Talukder, M. M. Islam, M. A. Uddin, A. Akhter, K. F. Hasan, and
vol. 13, no. 6, p. 4054, Mar. 2023.
M. A. Moni, ‘‘Machine learning-based lung and colon cancer detection
[23] S. Sorguli and H. Rjoub, ‘‘A novel energy accounting model using fuzzy
using deep feature extraction and ensemble learning,’’ Expert Syst. Appl.,
restricted Boltzmann machine—Recurrent neural network,’’ Energies,
vol. 205, Nov. 2022, Art. no. 117695.
[3] O. Attallah, M. F. Aslan, and K. Sabanci, ‘‘A framework for lung and colon vol. 16, no. 6, p. 2844, Mar. 2023.
cancer diagnosis via lightweight deep learning models and transformation [24] Y. Li and Y. Li, ‘‘Predicting chaotic time series and replicating chaotic
methods,’’ Diagnostics, vol. 12, no. 12, p. 2926, Nov. 2022. attractors based on two novel echo state network models,’’ Neurocomput-
[4] M. A. Fahami, M. Roshanzamir, N. H. Izadi, V. Keyvani, and ing, vol. 491, pp. 321–332, Jun. 2022.
R. Alizadehsani, ‘‘Detection of effective genes in colon cancer: A machine [25] S.-M. Guo, J.-K. Guo, Y.-G. Gao, P.-Y. Guo, F.-J.-A. Huang, S.-C. Wang,
learning approach,’’ Informat. Med. Unlocked, vol. 24, 2021, Z.-C. Lou, and X. Zhang, ‘‘Research on engine speed control based on tuna
Art. no. 100605. swarm optimization,’’ J. Eng. Res. Rep., pp. 272–280, Dec. 2022.
[5] H. C. Reis and V. Turk, ‘‘Transfer learning approach and nucleus segmen- [26] A. A. Borkowski, M. M. Bui, L. B. Thomas, C. P. Wilson, L. A. DeLand,
tation with MedCLNet colon cancer database,’’ J. Digit. Imag., vol. 36, and S. M. Mastorides. Lung and Colon Cancer Histopathological Image
no. 1, pp. 306–325, Sep. 2022. Dataset (LC25000). Accessed: Mar. 14, 2023. [Online]. Available:
[6] N. Kumar, M. Sharma, V. P. Singh, C. Madan, and S. Mehandia, https://www.kaggle.com/datasets/andrewmvd/lung-and-colon-cancerhisto
‘‘An empirical study of handcrafted and dense feature extraction techniques pathological-images?resource=download
for lung and colon cancer classification from histopathological images,’’ [27] H. A. Mengash, M. Alamgeer, M. Maashi, M. Othman, M. A. Hamza,
Biomed. Signal Process. Control, vol. 75, May 2022, Art. no. 103596. S. S. Ibrahim, A. S. Zamani, and I. Yaseen, ‘‘Leveraging marine preda-
[7] M. Masud, N. Sikder, A.-A. Nahid, A. K. Bairagi, and M. A. AlZain, tors algorithm with deep learning for lung and colon cancer diag-
‘‘A machine learning approach to diagnosing lung and colon cancer using nosis,’’ Cancers, vol. 15, no. 5, p. 1591, 2023, doi: doi.org/10.3390/
a deep learning-based classification framework,’’ Sensors, vol. 21, no. 3, cancers15051591.
p. 748, Jan. 2021.

94712 VOLUME 11, 2023