Received 18 July 2023, accepted 24 September 2023, date of current version 16 October 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3322376

Sparrow Search Algorithm With Stacked Deep

Learning Based Medical Image Analysis for
Pancreatic Cancer Detection and Classification
JANJHYAM VENKATA NAGA RAMESH 1 , T. ABIRAMI2 ,
T. GOPALAKRISHNAN 3 , (Member, IEEE), KANAGARAJ NARAYANASAMY4 , (Member, IEEE),
MOHAMAD KHAIRI ISHAK 5,6 , FATEN KHALID KARIM 7 , SAMIH M. MOSTAFA 8,9 ,
AND ALAA ALLAKANY 10
1 Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Vijayawada, Andhra Pradesh 522502, India
2 Department of Information Technology, Kongu Engineering College, Erode 638060, India
3 Department of Information Technology, Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal 576104, India
4 Department of Computer Science, Karpagam Academy of Higher Education, Coimbatore 641021, India
5 School of Electrical and Electronic Engineering, Engineering Campus, Universiti Sains Malaysia, Nibong Tebal 14300, Malaysia
6 Department of Electrical and Computer Engineering, Ajman University, Ajman, United Arab Emirates
7 Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh 11671, Saudi Arabia
8 Computer Science Department, Faculty of Computers and Information, South Valley University, Qena 83523, Egypt
9 Faculty of Industry and Energy Technology, New Assiut Technological University (NATU), New Asyut 71515, Egypt
10 Faculty of Computers and Information, Kafrelsheikh University, Kafr El-Shaikh 6860404, Egypt

Corresponding author: Faten Khalid Karim ([email protected])

This work was supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2023R300),
Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

ABSTRACT Medical image analysis for pancreatic cancer (PC) classification and recognition is a vital
domain of research and medical practices. PC is challenging to diagnose and treat; medical imaging
approaches aid early diagnosis to analyse and treat, and employ of medical imaging approaches are support
early diagnosis, correct analysis, and treatment planning. Computed Tomography (CT) scans are generally
utilized to detect and classify PCs. Deep learning (DL) approaches have demonstrated the ability to support
the diagnosis and detection of several medical conditions, containing PC. Convolutional Neural Networks
(CNNs) are a kind of DL approach generally employed for image analysis that is trained to automatically
learn and extract features in medical images. So, this study purposes a new Sparrow Search Algorithm with
Stacked Deep Learning based Medical Image Analysis for Pancreatic Cancer Detection and Classification
(SSASDL-PCDC) technique on CT images. The purpose of the study is to design an SSASDL-PCDC
technique to achieve improved pancreatic cancer detection performance. In addition, the SSASDL-PCDC
technique applies Harris Hawks Optimization (HHO) with a densely connected networks (DenseNet) model
for the feature extraction process. Moreover, convolutional neural network with bi-directional long short-
term memory (CNN-BiLSTM) approach was utilized for PC detection and classification. Furthermore,
Sparrow Search Algorithm (SSA) is used to adjust the hyperparameter values of the CNN-BiLSTM
technique. To evaluate the effectiveness of the SSASDL-PCDC technique, extensive experiments were
executed on a comprehensive database of pancreatic CT images. The simulation outcome value depicted that
the SSASDL-PCDC technique with maximum sensitivity of 99.26%, specificity of 99.26%, and accuracy
of 99.26%.

INDEX TERMS Pancreatic cancer, computed tomography images, sparrow search algorithm, medical image
analysis, cancer diagnosis.

I. INTRODUCTION
The associate editor coordinating the review of this manuscript and Pancreatic cancer (PC) is a major cause of death around
approving it for publication was Eduardo Rosa-Molinar . the world, with growing incidence. Earlier detection of PC
2023 The Authors. 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 111927
 J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

is the most important challenge, as the disease is usually This study proposes a novel Sparrow Search Algo-
diagnosed at later stages [1]. Patients with earlier-stage dis- rithm with Stacked DL based Medical Image Analysis for
ease are cured with the help of radiotherapy, surgery, and Pancreatic Cancer Detection and Classification (SSASDL-
chemotherapy. Therefore, an optimum understanding of the PCDC) technique on CT images. The SSASDL-PCDC tech-
risk factor for PC and diagnosis at an earlier stage has nique applies Harris Hawks Optimization (HHO) with a
tremendous potential to reduce overall mortality and improve densely connected networks (DenseNet) model for the fea-
patient survival. Compared to other cancer types, the inci- ture extraction process. Moreover, a convolutional neural
dence of PC is considerably lower, for example colorectal, network with bi-directional long short-term memory (CNN-
lung, and breast cancer [2]. The risk of PC was calculated for BiLSTM) approach was utilized for PC detection and classi-
a long time based on clinical risk, family history, genetic pre- fication. Furthermore, SSA was utilized to alter the hyperpa-
disposition, circulating biomarkers, and behavioral factors. rameter values of the CNN-BiLSTM approach. To evaluate
Presently, patient with high risk due to cystic lesion of the the effectiveness of the SSASDL-PCDC system, extensive
pancreas or family history or rare inherited pathogenic vari- experiments were performed on a comprehensive dataset of
ant undergoes serial pancreas-directed imaging to diagnose pancreatic CT images.
earlier PC [3]. But this patient only accounts for a smaller
fraction of those progressing PC, and data on genetic risk II. RELATED WORKS
factors or family history are not frequently accessible in the Alves et al. [11] employed a recent DL approach for develop-
overall population. CT is the imaging modality used mainly ing an automatic structure for PDAC recognition, concentrat-
for the assessment and detection of PC, but the method’s ing on smaller lesions. The PDAC is employed for training
diagnostic performance is based on the experience of the a nnUnet for automatic cancer recognition and segmentation
radiologist [4]. Moreover, about 40% of tumours that are (nnUnet+T). The 2 extra nnUnet are trained for investigat-
lesser than 2cm evade detection by CT, underscoring a cru- ing the effect of anatomy combination: (i) segmentation of
cial necessity for novel approaches to improve radiologist the tumor, pancreas, and multiple surrounding anatomical
interpretation in increasing the sensitivity for the analysis of structures (nnUnet+MS), and (ii) segmentation of the tumor
PC [5]. and pancreas (nnUnet+TP). Chaithanyadas and King [12]
Artificial intelligence (AI) is a field of computer science concentrations on the effectual classification and segmenta-
offered to produce a novel type of intelligent machine, which tion of PC utilizing DL approach with the subsequent steps
respond just like human intelligence [6]. Currently, several of the procedure: (i) The MRI images gathered in reputed
research workers attempt to exploit AI in the healthcare sec- clinics were utilized for the analysis method, (ii) primarily,
tor, involving healthcare, cardiology, oncology, etc. In com- the raw images can be pre-processing by employing boosted
parison to conventional biometric techniques, AI has great anisotropic diffusion filter (BADF) and CLAHE approaches,
scalability and flexibility that enables it to be deployed for (iii) additional, the segmentation method isolates the nodules
various tasks [7]. An additional benefit is its capability to utilizing FCM clustering (DMFCM) approach, (iv) the fea-
join in a wide range of data varieties and understands diffi- tures exist before removed and chosen by employing HHO-
cult connections among variables in trainable, flexible man- related SVM.
ners. AI shows more benefits in processing big data as the In [13], the DL approach termed as CNN approach was
computer computing power continues to enhance and scale utilized for predicting the pancreas cancer images that are
of medical data continues to expand [8]. In medical prac- embedded by the process of Gaussian Mixture model (GMM)
tice, AI can consistently carry out routine tasks, which free with the EM method for forecasting the major features in
up medical doctor time to resolve medical problems. Deep the CT scan and forecasts the percentage of cancer spread
learning (DL) with convolutional neural networks (CNN) from the pancreas with threshold factors obtained as markers.
showed great promise in the study of medical images. The The authors [14] establish a DL-based Hierarchical CNN
construction of a neural network (NN) is based on the stack (H-CNN) for PC detection. Besides, the H-CNN training and
of neurons encompassed of activation function and param- configurations purposes are established for the efficiency of
eter to combine and feature extraction in the image and the PC image segment.
established a mechanism that takes the complex connections The authors [15] executed a CNN approach utilizing
among diagnosis and images [9]. CNN accomplishes a high ResNet-50 to differentiate among MCN and SCN.
accuracy in the imaging diagnosis of different circumstances Data augmentation is employed for improving the quality
including liver masses, skin cancer, and diabetic retinopathy. and size of trained datasets. Fine-tuned training methods
But, the probable usefulness of CNN for the recognition can be employed by executing the pre-training method in
and diagnoses of PC is not been extensively analyzed [10]. transfer learning (TL) but trained chosen layers. In [16], the
PC frequently proposes with ill-defined margins and irregular authors present a new Multi-modal Fusion and Calibration
contours on CT and therefore are often obscure at a previous Network (MFCNet) for tumor segmentation dependent upon
stage that possesses significant challenges even for the most 3D PET-CT images. Primarily, a Multimodal Fusion Down-
experienced radiologist. sampling Block (MFDB) with the remaining design was

111928 VOLUME 11, 2023

J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

FIGURE 1. Workflow of SSASDL-PCDC algorithm.

presented. The presented MFDB fuse complementary fea- The authors utilized a square-window-related CNN structure
ture of multimodal images but retained the single fea- with 3 convolution layer blocks.
tures of various modal images. Secondarily, the Multimodal
Mutual Calibration Block (MMCB) dependent method of the III. THE PROPOSED MODEL
inception infrastructure was planned. In this study, we have established a novel SSASDL-
Abbas and Obied [17] examine a CAD approach employ- PCDC approach for PC detection on CT images. The study
ing Synergic Inception ResNet-V2, Deep CNN structure for aims to design an SSASDL-PCDC technique to achieve
detection of PC cases in openly Usable CT images, which improved PC detection performance. The SSASDL-PCDC
is removed PC graphical utility to contain medical analysis technique comprises several subprocesses such as DenseNet-
previously the pathogenic analysis, saving valued time for based feature extraction, HHO-based hyperparameter tuning,
disease prevention. Liang et al. [18] purposes to establish a CNN-BiLSTM-based classification, and SSA-based param-
process enabling automatic segmentation of pancreatic GTV eter optimization. Fig. 1 represents the workflow of the
dependent upon multi-parametric MRI employing DNNs. SSASDL-PCDC algorithm.

VOLUME 11, 2023 111929

 J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

A. FEATURE EXTRACTION MODULE their location. Thus, the algorithm could converge towards
In this work, the DenseNet121 approach was employed to a potential solution using this exploitation process. These
derive features from the CT images. Huang et al. presented 3 kinds of hunting behaviors can be detected in Harris
the DenseNet method, and this method improves feature hawks:
reuse abilities dependent upon ResNet in their structure [19]. • Exploration: Harris hawks randomly fly to determine
It takes L(L+1)/2 direct connections, but typical CNN takes novel regions where they could be searched.
L layers with L connections. In DenseNet, mapping features • Exploitation: Harris Hawks attempts to enhance the
were integrated utilizing concatenation rather than summing solution given by the individual.
previously passing as a layer, and every preceding layer map- • Intensification: Harris hawks co-ordinate their effort to
ping feature was employed as input to some particular layers. find their prey.
The Dense Block is the key infrastructure of DenseNet, con- The HHO will be used to enhance the network hyperpa-
taining convolution layers. DenseNet-121 is the most varia- rameter after the initialization of the pre-trained module. The
tions of the DenseNet structure containing a 121-connected HHO exploits the following hunting behaviors to search for
convolution layer with the last resultant layer. an optimum solution:
DenseNet-121 comprises 4 dense blocks, and there exists  
a transition layer among all the dense blocks. This network Xit+1 = Xit + rand(). Xjt − Xkt , (1)
dense connectivity for x0, x1, xl − 1 inputs, whereas the l th
In Eq. (1), rand() denotes the random integer within [0, 1],
layer obtains mapping features in every prior layer demon-
j and k are two random selection hawks, and t shows the exist-
strated as Xl = Hl([x0,x1, . . . ,xl−1]). Hl denotes the com-
ing iteration. The Harris hawks guide the optimizer technique
posite function which comprises 3 functions 3 × 3 convolu-
by updating the position according to the other hawks:
tional, BN, and ReLU. All the dense blocks of DenseNet121
take 2 convolutional layers, 1 × 1 and 3 × 3 that are repeated
 
Xit + Xjt  
differently from all the blocks. A transition layer comprises f +1
Xi = + rand () · Xjt − Xkt , (2)
1 × 1 convolution sand average pooling layers with a stride 2
of 2. Previously sending every mapping feature to the FC B. DETECTION MODULE
layer for classification, this network carries out a 7 × 7 global
For the detection and classification of PC, the CNN-
average pooling layer.
BiLSTM model is used. The CNN, a conventional DNN,
Data-driven AI model is represented by 3 basic ele-
has fully connected, convolutional, and pooling layers [21].
ments: the data use, the parameter utilized for training
It dynamically identifies the higher-level feature demon-
them, and the structure of network use them. Any DL algo-
stration of time sequences in the multivariate time series
rithm should achieve a good balance amongst these com-
prediction to gather considerable data and implements a
ponents to perform tasks with maximal efficiency. A group
local connection and global sharing for time sequences
of primary parameters is utilized in different trials to con-
by the convolutional function. CNN is capable of analyzing
firm the design of the model, and the model performance
the relationship between the variables accurately. Usually,
was measured within every trial. Furthermore, there exists
the 1D-CNN layer simplifies and improves the predictive
optimization concern related to enhancing network depth.
model.
Additionally, improper selection of hyperparameters might
Here, a stacked BLSTM network was used as a basic
lead to the suboptimum result. The manual hyperparameter
element of the model that provides the advantage of capturing
tuning for DenseNet is a laborious process. Subsequently,
temporal data from different sequences. The data from the
the hyperparameter optimization algorithm should be imple-
past and future time series are crucial to making forecasts
mented to iteratively find the optimum value dependent upon
on the PV ′ s power. Consequently, the bi-directional data in
the trained data and upgrade them. The HHO is a meta-
the PV correlation datasets was maximized by the BLSTM to
heuristic optimizer algorithm utilized for the DenseHHO
discover the relationships between past, present, and future
architecture. HHO is stimulated by the hunting behaviors
time steps. BLSTM is a development of the LSTM, an RNN
of Harris hawks [20]. The HHO technique can resolve the
variant which could resolve the vanishing gradient problem
problems of different optimization techniques namely param-
by integrating a pass-selective model. The context data can
eter optimization, function optimization, and feature selec-
be attained from the previous statement of whole inputs; the
tion for ML algorithms. The subsequent steps need to be
BLSTM incorporates the past weather data in both ways and
taken for implementing the algorithm. The most important
operated as follows:
variance among HHO and other optimizer techniques lies
in the establishment of the concept of elite and non-elite ft = σ Wf × [ht−1 xt ] + bf


hawks rather than following a severe hierarchical system. it = σ Wj × [ht−1 xt ] + bi


Consequently, elite hawk takes the lead in the search tech-
nique, which represents the best solution. It is mimic the C̃t = tanh (WC × [ht−1 xt ] + bC )
behaviors of optimum hawks from the population to maintain Ct = ft ×Ct−1 + it × C̃t

111930 VOLUME 11, 2023

J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

0t = σ (W0 × [ht−1 xt ] + b0 ) separate sparrows, and the value of all the rows in Gx shows
ht = 0t × tanh (C) (3) the fitness value (FV) of individuals.
The joiners need a massive predation range to search for the
where ht−1 indicates the implicit layer state at the prior best food, hence in all the iterations, the location of joiners is
time; xt denotes the input parameter at t moment , C̃t shows upgraded as:
the temporary memory state at t time, Ct−1 shows the unit  

memory layer from the prior instant; W and b embodies i
 X · exp −
i,j R2 < ST
the weighted and biased utilized during the trained process iter+1
Xi,j = α · iiermax (6)
and, and ft , it , 0t represent the manages the amount of data 
Xi,j + QL R2 ≥ ST
generated by the forget gate, update gate, and memory units.
−−−→ In Eq. (6), itermax is the constant representing the maximal
h⃗t = LSTM (ht−1 , xt , ct−1 ) , t ∈ [1,T ]
amount of iterations and iter shows the existing iteration.
← −−−→
h t = LSTM (ht+1 , xt , ct+1 ) , t ∈ [T , 1] α represents the random integer in the [0, 1]. Xi,j refers to
h ←i the value of jth parameter of the ith sparrows at t th iterations.
Ht = h⃗t , h t (4) R2 and ST denote the alarm value and safety threshold,
correspondingly, L shows the matrix of dimension 1×d, but
In Eq. (4), h⃗t shows the hidden layer of the forward layer at
← all the elements are fixed as 1. If R2 ≥ST , there exists danger;
t time , h t denotes the hidden layer of backward at t moment, if R2 < ST , the surroundings are safe and the finders search,
and Ht represents the hidden state in the BLSTM for the t and the population leaves flies and food toward other safer
moment . Fig. 2 illustrates the framework of BiLSTM. regions.
Joiners search for food by following and monitoring the
modified finders, and once the lesser changed joiner could
not obtain food, it creates a position alter to obtain additional
food:
 !
 iter
xworst − xi,i
 Q · exp i> 2n

iter+1
Xi,j = i2
 X iter+1 + Xi,j − X iter+1 A+ L otherwise


p p

(7)

In Eq. (7), Xworsi signifies the worst location. Xp represents

the fittest place presently occupied by the producer, and A
refers to the matrix of 1×d size, where the element is assigned
randomly both 1 and −1, and A+ = AT (AAT )−1 .i > n/2
FIGURE 2. Architecture of BiLSTM. shows that ith accession with lesser FVs is hungry and not
able to search.
The initial position was generated randomly when the spar-
C. HYPERPARAMETER TUNING MODULE row proportions that recognize danger reached 10-20% of the
Finally, the SSA adjusts the hyperparameter values of the overall population. Then, the initialization joiners and finders
CNN-BiLSTM model. The high search efficiency and simple involve in competition for food and upgrade the location still
implementation of the SSA technique make it an intelligent the maximal iteration counts are attained. At last, the sparrow
optimization algorithm [22]. This approach splits the sparrow with maximum global FV can be recognized as a global
population as to finders and joiners, with finders seeking food optimum performance:
and if food places for another joiner and groups searching 
 Xbest + β Xi,j − Xbest fi < fg
iter iter iter
by subsequent finders. Consider that the sparrow population 

comprises the size of parameters that enhanced d and n iter+1
Xi,j = iter iter
xi,j − Xworst (8)
sparrows. The location matrix X characterizes the n sparrows iter
 X + K f i = f g
i,j
(f − f ) + ε


in d-dimension space and is represented as follows: i w

x1 x1 . . . x1d g x1 x1 . . . x1d
 1 2    1 2 

In Eq. (8), Xbest shows the present optimum location. K
 x1 x2 . . . xd   g x 1 x 2 . . . x d 
 takes values within [1, 1]. In addition ε denotes a smaller

 2 2 2 2 2 2
X =  . . . .  Gx =  ..  (5) constant utilized to avoid zero-division error. β denotes the
 
 .. .. .. ..   . 
step control parameter that is a randomly generated value
xn xn . . . xn g xn xn . . . xn
1 2 d
 1 2 d


with standard distribution with a mean of 0 and variance of
In Eq. (5), n shows the sparrow counts, d denotes the 1.fg and fw indicate the present global optimum and worse
dimensional of the variable being optimizer, χ represents the FVs, correspondingly.

VOLUME 11, 2023 111931

 J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

The SSA method progresses a fitness function (FF) for TABLE 2. Pancreatic cancer recognition outcome of SSASDL-PCDC
approach on 70:30 TRP/TSP.
accomplishing greater classifier outcomes. It expresses a pos-
itive integer to archetypal the good solution for candidate
performances. During this work, the minimized classifier
error rate can be supposed to be FF, as depicted in Eq. (9).
fitness (xi ) = ClassifierErrorRate (xi )
no.of misclassified instances
= ∗100 (9)
Total no.of instances

IV. RESULTS AND DISCUSSION

The pancreatic cancer classification results of the SSASDL-
PCDC technique are tested using a CT image dataset, com-
prising 250 pancreatic tumor images and 250 non-pancreatic
tumor images as defined in Table 1. Fig. 3 depicts the sample
images.

TABLE 1. Details of database.

TABLE 3. Comparative outcome of SSASDL-PCDC algorithm with other

methods.

TABLE 4. CT outcome of SSASDL-PCDC algorithm with other methods.

FIGURE 3. Sample images.

Fig. 4 showcases the classifier outcome of the SSASDL-

PCDC approach under the test dataset. Fig. 4a portrays the
confusion matrix presented by the SSASDL-PCDC system
on 70% of TRP. The result implied that the SSASDL-PCDC
algorithm has recognized 177 instances of Pancreatic Tumors In Table 2 and Fig. 5, an overall pancreatic cancer recog-
and 166 instances of Non-Pancreatic Tumors. In addition, nition result of the SSASDL-PCDC technique is displayed.
Fig. 4b represents the confusion matrix provided by the The experimental values stated that the SSASDL-PCDC tech-
SSASDL-PCDC methodology on 30% of TSP. The outcome nique attains effectual recognition of pancreatic and non-
inferred that the SSASDL-PCDC algorithm has recognized pancreatic tumors. On 70% of TRP, the SSASDL-PCDC
67 instances of Pancreatic Tumors and 82 instances of Non- technique offers average accuy , precn , sensy , specy , and
Pancreatic Tumors. Followed by, Fig. 4c reveals the PR curve Fscore of 98.03%, 97.98%, 98.03%, 98.03%, and 98% respec-
of the SSASDL-PCDC approach. The outcome stated that the tively. Meanwhile, on 30% of TSP, the SSASDL-PCDC
SSASDL-PCDC system has achieved higher PR outcomes in algorithm provides average accuy , precn , sensy , specy , and
2 class labels. Lastly, Fig. 4d exemplifies the ROC curve of Fscore of 99.26%, 99.40%, 99.26%, 99.26%, and 99.33%
the SSASDL-PCDC algorithm. The result exhibited that the correspondingly.
SSASDL-PCDC approach has resulted in capable outcomes Fig. 6 displays the training accuracy TR_accuy and
with superior ROC values under 2 classes. VL_accuy of the SSASDL-PCDC approach. The TL_accuy

111932 VOLUME 11, 2023

J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

FIGURE 4. Performance of (a-b) Confusion matrices, (c) PR-curve, and (d) ROC curve.

FIGURE 6. Accuy curve of the SSASDL-PCDC approach.

FIGURE 5. Average outcome of SSASDL-PCDC approach on 70:30 TRP/TSP.

In Fig. 7, the TR_loss and VR_loss outcomes of the

SSASDL-PCDC algorithm are exposed. The TR_loss demon-
strates the error among the predictive outcome and original
is defined by the estimate of the SSASDL-PCDC algorithm values on the TR data. The VR_loss signifies the measure of
on the TR dataset whereas the VL_accuy is computed by the performance of the SSASDL-PCDC system on individual
estimating the performance on a distinct testing dataset. The validation data. The outcomes inferred that the TR_loss and
outcomes depict that TR_accuy and VL_accuy increase with VR_loss incline to lesser with increasing epochs. It exposed
an upsurge in epochs. Therefore, the performance of the the greater performance of the SSASDL-PCDC system and
SSASDL-PCDC approach obtains enhancement on the TR its capability to create a correct classification. The minimal
and TS dataset with an increase in the number of epochs. value of TR_loss and VR_loss illustrates the greater outcome

VOLUME 11, 2023 111933

 J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

of the SSASDL-PCDC approach to capturing patterns and

relationships.

FIGURE 9. CT outcome of SSASDL-PCDC algorithm with other methods.

FIGURE 7. Loss curve of the SSASDL-PCDC approach.

SSASDL-PCDC methodology displays lesser performance

Extensive comparative results of the SSASDL-PCDC
with a CT of 0.88s. Thus, the SSASDL-PCDC technique can
technique are defined in Table 3 and Fig. 8 [23], [24],
be employed for accurate and automated pancreatic cancer
[25]. The result demonstrated that the CNN model reaches
classification.
worse results whereas the WELM, KELM, and ELM mod-
els obtain slightly boosted performance. Along with that,
V. CONCLUSION
the ODL-PTNTC model attains considerable performance.
Next, the IDLDMS-PTC technique gains reasonable per- In this study, we have established a novel SSASDL-PCDC
formance with asensy of 99.05%, specy of 98.84%, and system for PC detection on CT images. The study aims to
accuy of 99.15%. Nevertheless, the SSASDL-PCDC tech- design an SSASDL-PCDC technique to achieve improved
nique demonstrates maximum performance with sensy of PC detection performance. The SSASDL-PCDC technique
99.26%, specy of 99.26%, and accuy of 99.26%. comprises several subprocesses such as DenseNet-based fea-
ture extraction, HHO-based hyperparameter tuning, CNN-
BiLSTM-based classification, and SSA-based parameter
optimization. The SSA was utilized to adjust the hyperpa-
rameter values of the CNN-BiLSTM approach. To evaluate
the effectiveness of the SSASDL-PCDC technique, extensive
experiments were performed on a comprehensive dataset of
pancreatic CT images. The experimental outcomes exhibit
that the SSASDL-PCDC technique via integration of the SSA
with stacked deep learning models significantly increases the
accuracy of PC detection and classification. In future, the
performance of the SSASDL-PCDC algorithm was increased
by the employ of fusion-based approaches.

ACKNOWLEDGMENT
Princess Nourah bint Abdulrahman University Researchers
Supporting Project number (PNURSP2023R300), Princess
Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
FIGURE 8. Comparative outcome of SSASDL-PCDC algorithm with other
methods.
REFERENCES
[1] S. Li, J. Xiao, L. He, X. Peng, and X. Yuan, ‘‘The tumor target seg-
A brief CT analysis of the SSASDL-PCDC approach mentation of nasopharyngeal cancer in CT images based on deep learn-
is described in Table 4 and Fig. 9. The outcome that ing methods,’’ Technol. Cancer Res. Treatment, vol. 18, Jan. 2019,
outperformed the CNN approach obtains the least outcomes Art. no. 153303381988456.
[2] A. Gupta, A. Koul, and Y. Kumar, ‘‘Pancreatic cancer detection using
whereas the WELM, KELM, and ELM systems acquire machine and deep learning techniques,’’ in Proc. 2nd Int. Conf. Innov.
somewhat boosted performances. Practices Technol. Manage. (ICIPTM), vol. 2, Feb. 2022, pp. 151–155.
Followed by, the ODL-PTNTC method gains consider- [3] S.-L. Liu, S. Li, Y.-T. Guo, Y.-P. Zhou, Z.-D. Zhang, S. Li, and Y. Lu,
‘‘Establishment and application of an artificial intelligence diagnosis sys-
able results. Afterwards, the IDLDMS-PTC approach attains tem for pancreatic cancer with a faster region-based convolutional neural
reasonable performance with a CT of 1.02s. Nevertheless, the network,’’ Chin. Med. J., vol. 132, no. 23, pp. 2795–2803, Dec. 2019.

111934 VOLUME 11, 2023

J. V. N. Ramesh et al.: SSA With Stacked DL Based Medical Image Analysis

[4] C. Song, M. Wang, Y. Luo, J. Chen, Z. Peng, Y. Wang, H. Zhang, [14] W. Xuan and G. You, ‘‘Detection and diagnosis of pancreatic tumor
Z.-P. Li, J. Shen, B. Huang, and S.-T. Feng, ‘‘Predicting the using deep learning-based hierarchical convolutional neural network on
recurrence risk of pancreatic neuroendocrine neoplasms after radical the Internet of Medical Things platform,’’ Future Gener. Comput. Syst.,
resection using deep learning radiomics with preoperative computed vol. 111, pp. 132–142, Oct. 2020.
tomography images,’’ Ann. Transl. Med., vol. 9, no. 10, p. 833, [15] L. S. Nguon, K. Seo, J.-H. Lim, T.-J. Song, S.-H. Cho, J.-S. Park, and
May 2021. S. Park, ‘‘Deep learning-based differentiation between mucinous cystic
[5] M. N. M. Sehmi, M. F. A. Fauzi, W. S. H. M. W. Ahmad, and neoplasm and serous cystic neoplasm in the pancreas using endoscopic
E. W. L. Chan, ‘‘Pancreatic cancer grading in pathological images using ultrasonography,’’ Diagnostics, vol. 11, no. 6, p. 1052, Jun. 2021.
deep learning convolutional neural networks,’’ FResearch, vol. 10, p. 1057, [16] F. Wang, C. Cheng, W. Cao, Z. Wu, H. Wang, W. Wei, Z. Yan, and Z. Liu,
Nov. 2022. ‘‘MFCNet: A multi-modal fusion and calibration networks for 3D pancreas
[6] S.-H. Zhen, M. Cheng, Y.-B. Tao, Y.-F. Wang, S. Juengpanich, Z.-Y. Jiang, tumor segmentation on PET-CT images,’’ Comput. Biol. Med., vol. 155,
Y.-K. Jiang, Y.-Y. Yan, W. Lu, J.-M. Lue, J.-H. Qian, Z.-Y. Wu, J.-H. Sun, Mar. 2023, Art. no. 106657.
H. Lin, and X.-J. Cai, ‘‘Deep learning for accurate diagnosis of liver tumor [17] S. K. Abbas and R. S. Obied, ‘‘Novel computer aided diagnostic system
based on magnetic resonance imaging and clinical data,’’ Frontiers Oncol., using synergic deep learning technique for early detection of pancreatic
vol. 10, p. 680, May 2020. cancer,’’ Webology, vol. 18, no. 2, pp. 367–379, Oct. 2021.
[7] R. Yang, Y. Chen, G. Sa, K. Li, H. Hu, J. Zhou, Q. Guan, and F. Chen, [18] Y. Liang, D. Schott, Y. Zhang, Z. Wang, H. Nasief, E. Paulson, W. Hall,
‘‘CT classification model of pancreatic serous cystic neoplasms and muci- P. Knechtges, B. Erickson, and X. A. Li, ‘‘Auto-segmentation of pan-
nous cystic neoplasms based on a deep neural network,’’ Abdominal creatic tumor in multi-parametric MRI using deep convolutional neural
Radiol., vol. 47, pp. 232–241, Oct. 2021. networks,’’ Radiotherapy Oncol., vol. 145, pp. 193–200, Apr. 2020.
[8] M. Kriegsmann, K. Kriegsmann, G. Steinbuss, C. Zgorzelski, A. Kraft, [19] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, ‘‘Densely
and M. M. Gaida, ‘‘Deep learning in pancreatic tissue: Identifica- connected convolutional networks,’’ in Proc. IEEE Conf. Comput. Vis.
tion of anatomical structures, pancreatic intraepithelial neoplasia, and Pattern Recognit. (CVPR), Jul. 2017, pp. 4700–4708.
ductal adenocarcinoma,’’ Int. J. Mol. Sci., vol. 22, no. 10, p. 5385, [20] P. P. F. Rajeena, W. N. Ismail, and M. A. S. Ali, ‘‘A metaheuristic Harris
May 2021. hawks optimization algorithm for weed detection using drone images,’’
[9] A. B. Bagheri, M. D. Rouzi, N. A. Koohbanani, M. H. Mahoor, Appl. Sci., vol. 13, no. 12, p. 7083, Jun. 2023.
M. G. Finco, M. Lee, B. Najafi, and J. Chung, ‘‘Potential applications of [21] L. Liu, K. Guo, J. Chen, L. Guo, C. Ke, J. Liang, and D. He, ‘‘A pho-
artificial intelligence and machine learning on diagnosis, treatment, and tovoltaic power prediction approach based on data decomposition and
outcome prediction to address health care disparities of chronic limb- stacked deep learning model,’’ Electronics, vol. 12, no. 13, p. 2764,
threatening ischemia,’’ Seminars Vascular Surg., to be published, doi: Jun. 2023.
10.1053/j.semvascsurg.2023.06.003. [22] J. Madiniyeti, Y. Chao, T. Li, H. Qi, and F. Wang, ‘‘Concrete dam deforma-
[10] C. Park, M. D. Rouzi, M. M. U. Atique, M. G. Finco, R. K. Mishra, tion prediction model research based on SSA–LSTM,’’ Appl. Sci., vol. 13,
G. Barba-Villalobos, E. Crossman, C. Amushie, J. Nguyen, C. Calarge, and no. 13, p. 7375, Jun. 2023.
B. Najafi, ‘‘Machine learning-based aggression detection in children with [23] T. Vaiyapuri, A. K. Dutta, I. S. H. Punithavathi, P. Duraipandy,
ADHD using sensor-based physical activity monitoring,’’ Sensors, vol. 23, S. S. Alotaibi, H. Alsolai, A. Mohamed, and H. Mahgoub, ‘‘Intelligent
no. 10, p. 4949, May 2023. deep-learning-enabled decision-making medical system for pancreatic
[11] N. Alves, M. Schuurmans, G. Litjens, J. S. Bosma, J. Hermans, and tumor classification on CT images,’’ Healthcare, vol. 10, no. 4, p. 677,
H. Huisman, ‘‘Fully automatic deep learning framework for pancreatic 2022, doi: 10.3390/healthcare10040677.
ductal adenocarcinoma detection on computed tomography,’’ Cancers, [24] M. M. Althobaiti, A. Almulihi, A. A. Ashour, R. F. Mansour, and D. Gupta,
vol. 14, no. 2, p. 376, Jan. 2022. ‘‘Design of optimal deep learning-based pancreatic tumor and nontumor
[12] K. V. Chaithanyadas and G. R. G. King, ‘‘Detection of pancreatic tumor classification model using computed tomography scans,’’ J. Healthcare
from computer tomography images using 3D convolutional neural net- Eng., vol. 2022, pp. 1–15, Jan. 2022.
work,’’ in Computational Vision and Bio-Inspired Computing Singapore: [25] K.-L. Liu, T. Wu, P.-T. Chen, Y. M. Tsai, H. Roth, M.-S. Wu, W.-C. Liao,
Springer, 2023, pp. 289–303. and W. Wang, ‘‘Deep learning to distinguish pancreatic cancer tissue from
[13] K. Sekaran, P. Chandana, N. M. Krishna, and S. Kadry, ‘‘Deep learning non-cancerous pancreatic tissue: A retrospective study with cross-racial
convolutional neural network (CNN) with Gaussian mixture model for external validation,’’ Lancet Digit. Health, vol. 2, no. 6, pp. e303–e313,
predicting pancreatic cancer,’’ Multimedia Tools Appl., vol. 79, nos. 15–16, Jun. 2020.
pp. 10233–10247, Apr. 2020.

VOLUME 11, 2023 111935