Received 23 May 2024, accepted 28 June 2024, date of publication 2 July 2024, date of current version 16 July 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3422313

A Novel Deep Learning Approach for Accurate

Cancer Type and Subtype Identification
JABED OMAR BAPPI1 , MOHAMMAD ABU TAREQ RONY 2 , MOHAMMAD SHARIFUL ISLAM 3,

SAMAH ALSHATHRI 4 , AND WALID EL-SHAFAI 5,6 , (Senior Member, IEEE)

1 Department of Electrical and Electronics Engineering, Port City International University, Chittagong, 4225 Bangladesh
2 Department of Statistics, Noakhali Science and Technology University, Noakhali 3814, Bangladesh
3 Department of Computer Science and Telecommunication Engineering, Noakhali Science and Technology University, Noakhali 3814, Bangladesh
4 Department of Information Technology, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, P.O. Box 84428,

Riyadh 11671, Saudi Arabia

5 Security Engineering Laboratory, Computer Science Department, Prince Sultan University, Riyadh 11586, Saudi Arabia
6 Department of Electronics and Electrical Communication Engineering, Faculty of Electronic Engineering, Menoufia University, Menouf 32952, Egypt

Corresponding authors: Mohammad Abu Tareq Rony (rony1513@[Link]), Samah Alshathri (sealshathry@[Link]), and
Walid El-Shafai ([Link]@[Link])
This work was supported by Princess Nourah Bint Abdulrahman University, Riyadh, Saudi Arabia, through the Researchers Supporting
Project under Grant PNURSP2024R197.

ABSTRACT Cancer is a disease where abnormal cells grow uncontrollably and spread to other body parts.
It can originate anywhere in the human body, which consists of trillions of cells. These cells continually
divide, replenishing the body’s needs. As cells age or sustain damage, they naturally undergo apoptosis,
allowing new cells to take their place. Our research uses a secondary dataset from Kaggle, comprising over
130,000 images representing various cancer types. We have developed a novel Deep-learning model capable
of detecting and classifying cancer at early stages with remarkable accuracy. The model classifies eight
primary cancer types and 26 subtypes, each represented by 5,000 images. Our approach combines various
computational tools, including pre-trained Convolutional Neural Networks, Machine learning, and Deep
learning classifiers such as KNN and SVM, and innovative multimodal architectures of merged CNN-LSTM
hybrids. We applied two distinct classification strategies. In our first approach, the main class and subclass
are classified together. In the second approach, the model first predicts the main eight classes and then
26 subclasses concerning the main class classification, where the KNN model achieved higher accuracy for
the Lymphoma class than CNNs. Finally, the X-OR gate-based fusion technique applied after prediction
significantly reduces misclassifications and enhances the certainty of cancer types. Our findings reveal great
accuracy levels of 99.25% for primary cancer classifications and 97.80% for subclass classifications. The
introduction of novel models, Vception (VGG + Inception) and Vmobilnet (VGG + MobileNet), integrated
with LSTM, further advances diagnostic capabilities. Again, By utilizing an X-OR gate post-prediction from
Vmobilenet and Vception models, we achieved a main class accuracy of 99.95% and a subclass accuracy of
99.13%, significantly boosting model confidence. Moreover, individually, KNN achieved 97.14% accuracy
for the Lymphoma class using PCA. This study not only sets a new benchmark in cancer detection but also
promises to improve patient care and treatment outcomes significantly.

INDEX TERMS Cancer detection, deep learning, cancer imaging, CNNs, misclassifications, PCA.

I. INTRODUCTION regular cells mutate into malignant forms. Such mutations

Cancer stands as a significant challenge in the domain are the consequence of complex interactions between genetic
of global health, responsible for almost 10 million deaths factors inherent to the individual and external environmental
each year. This severe and intricate illness arises when influences. The progression of cancer involves multiple
phases, starting with the emergence of pre-cancerous con-
The associate editor coordinating the review of this manuscript and ditions and eventually leading to the growth of invasive
approving it for publication was Carmelo Militello . tumors [1].
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see [Link]
94116 VOLUME 12, 2024
J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

In developing countries, cancer mortality remains a sig- prompt diagnosis significantly reduces the risk of death
nificant challenge for humanity [2]. Despite the availability from these diseases. Advanced technologies, such as Deep
of numerous preventive measures, some types of cancer Learning(DL) and ML, are being employed to accelerate the
still lack effective treatments. Cancer mortality represents detection of these cancers. These methods enable the medical
a critical concern within the healthcare sector, notably as research community to evaluate a larger number of patients
a leading cause of death among women. Early detection more quickly and cost-effectively [11].
of brain cancer is essential for effective treatment. Brain Cervical lymphadenopathy frequently occurs in children.
tumors are diagnosed through biopsies, a process requiring Currently, there is no established decision model for identi-
invasive brain surgery. However, computational intelligence fying high-grade Lymphoma in pediatric patients with this
techniques offer a non-invasive alternative, assisting physi- condition. While previous research has pinpointed specific
cians in identifying and classifying brain tumors [3]. predictive factors for Lymphoma, and a small number have
Automatic detection of leukemia is a vital process in developed multivariate models, none have achieved the
early medical diagnosis and treatment planning and presents discriminative power necessary for practical clinical use [12].
a formidable challenge in healthcare. Leukemia, a blood Oral cancer stands as one of the leading cancer types
cancer originating in the bone marrow, primarily affects white worldwide, with late detection significantly contributing to
blood cells (WBCs), disrupting normal blood function [4]. increased morbidity and higher mortality rates. Roughly
A cost-effective and minimally invasive method, microscopic 50% of all oral cancer cases are identified in South Asia,
analysis of WBCs is widely recognized for the early and about two-thirds of cases arise in low- and middle-
identification of this hematological disorder. Despite its income nations. The predominant causes of oral cancer are
importance, there has been a scarcity of extensive literature extensive alcohol use and tobacco smoking. The survival rates
reviews on the application of deep and Machine Learn- for oral cancer are notably low in these regions, primarily
ing(ML) techniques specifically for Acute Lymphoblastic because approximately two-thirds of affected individuals
Leukemia (ALL) detection, a gap that underscores the need in low- and middle-income countries present with oral
for more thorough research in this critical area of medical lesions at advanced stages, hindering effective treatment
technology [5]. outcomes [13].
Breast cancer is the most prevalent form among women Molecular imaging allows for the observation and quan-
and particularly affects those with denser breast tissue due titative evaluation of changes in biological processes at the
to physiological characteristics [6]. Early detection of this molecular and/or cellular scale, offering crucial insights
disease is crucial in reducing the mortality rate. According into early cancer detection [14]. In contemporary medical
to Globocan 2018 data, breast cancer accounts for one in imaging analysis, DL has emerged as a prevalent method, sur-
every four cancer diagnoses in women globally and is the fifth passing conventional ML and visual assessment constraints
leading cause of death worldwide [7]. by extracting intricate hierarchical features with robust
Cervical cancer, predominantly affecting women, is a representational capacity. There’s a dynamic surge in cancer
globally recognized health concern. It arises from the molecular image research employing DL methodologies [15].
abnormal proliferation of cells in the cervix and can gradually The following are the main contributions of this research
extend to other body organs [8]. If identified early, cervical work:
cancer is entirely treatable. There are several screening • As all of the previous research focuses on single cancer
techniques for this disease, yet early detection through classes, we proposed a novel method to tackle eight main
cost-effective Pap smear tests is crucial for a successful cancer classes and 26 sub-classes concurrently, offering
cure [9]. a comprehensive understanding of cancer types.
The kidneys function as the body’s filtration system, elim- • We introduce two novel innovative convolutional neural
inating unwanted or harmful substances while recirculating network architectures, VCEPTION (combining VGG
essential nutrients such as vitamins, amino acids, glucose, and Inception) with LSTM and VMOBILENET (com-
and hormones into the bloodstream. Research indicates that bining VGG and MobileNet) with LSTM, to elevate
kidney cancer ranks as the 13th most prevalent cancer model performance in multicancer image classification.
worldwide, and among men, it is the 9th most common. • Finally, a novel fusion approach is applied where we
In 2012, there were 214,000 new cases in men and 124,000 introduce X-OR gate integration in two independently
in women. In the United States, kidney tumors represent trained models post-prediction, we ensure precise dis-
approximately 3.7% of all cancer diagnoses. Annually, over ease classification, advancing diagnostic accuracy and
62,000 Americans are diagnosed with one of the 4,444 types clinical decision-making in cancer imaging.
of kidney cancer [10]. The organization of this paper is as follows: Section II
Colon and lung cancer have emerged as significant reviews the relevant literature. Section III introduces our
contributors to disability and mortality worldwide. The novel method for cancer detection. Section IV evaluates
histopathological examination of these cancers plays a crucial the results of different deep learning techniques. Finally,
role in identifying the best treatment approach. Early and Section V summarizes our research findings.

VOLUME 12, 2024 94117

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

II. LITERATURE REVIEW impressive accuracy levels of 97.2%, sensitivity of 98.3%,

The ongoing research focuses on using DL techniques to and specificity of 96.5%, outperforming existing systems.
diagnose various cancers from image datasets. Recent studies The article in [19] aims to improve cervical cancer screen-
have aimed to develop ML models for accurate and early ing through computer-assisted diagnosis. By developing a
cancer detection. The use of DL techniques to analyze method that combines low- and high-resolution whole slide
radiographic images for cancer detection and classification images (WSIs) for lesion cell recognition and a neural
has become increasingly popular. This section provides network-based model for WSI classification, the research
a detailed review of DL methods used for diagnosing achieves high specificity (93.5%) and sensitivity (95.1%) in
cancers, along with a comparative analysis summarized slide classification. Tested on 1,170 patient-wise WSIs, the
in Table 1. system outperforms three cytopathologists on average. It also
This study [16] tackles early detection by proposing a highlights the top 10 lesion cells with an 88.5% true positive
sophisticated deep-learning algorithm that scrutinizes white rate on 447 positive slides. Post-deployment, the system
blood cells using microscopic blood smear images. ALL is a can recognize one giga-pixel WSI in about 1.5 minutes,
critical cancer predominantly found in children, characterized promising efficient and accurate cervical cancer screening.
by the overproduction of lymphocytes. The novel algorithm The research [20] introduces computational techniques
is based on a Convolutional Neural Networks(CNNs), for kidney segmentation in CT images, crucial for disease
aptly named ALLNET, which distinguishes leukemic cells diagnosis and treatment planning. Manual segmentation is
from non-cancerous ones. The research utilized a publicly time-consuming and subject to variability, prompting the
accessible dataset of microscopic images for training and adoption of deep CNNs. The proposed method combines
validating the ALLNET model. The training was executed on image processing and CNNs to minimize false positives.
Google Collaboratory, harnessing the computational power Using the KiTS19 dataset, our approach achieved outstanding
of Nvidia Tesla P-100 GPUs. The model excelled, attaining results: Dice coefficient of 96.33%, Jaccard index of 93.02%,
an accuracy of 95.54%, specificity of 95.81%, sensitivity of sensitivity of 97.42%, specificity of 99.94%, and accuracy
95.91%, F1-score of 95.43%, and precision of 96%. These of 99.92%. In the KiTS19 challenge, it attained a Dice
results underscore the model’s potential as a pre-screening coefficient of 93.03%. These findings underscore the efficacy
tool for leukemia in blood tests, significantly aiding in the of our method, showcasing its potential for precise kidney
early and accurate detection of ALL. segmentation in CT imaging, aiding in early tumor detection
The study [17] focuses on classifying brain tumor types and clinical decision-making.
from MRI images using an enhanced DL approach, specif- This paper [21] introduces a novel training strategy for
ically employing Residual Networks. The objective is to accurate segmentation of sparsely annotated histopatholog-
develop a Computer Assisted Diagnosis (CAD) system to ical data and proposes a transfer learning scenario for precise
aid doctors and radiologists in accurate tumor diagnosis and classification of colorectal images. Leveraging pre-trained
classification. The benchmark dataset consists of 3064 MRI DL models from computer vision datasets addresses the
images encompassing three brain tumor types: Meningiomas, challenge of limited annotations in whole-slide images
Gliomas, and Pituitary tumors. Through rigorous evaluation, (WSIs). Various state-of-the-art CNNs are reviewed and
the proposed model achieves a remarkable 99% accuracy, compared, including AlexNet, VGG, ResNet, DenseNet, and
surpassing previous methodologies. In addition to accuracy, Inception models. Transfer learning utilizes features learned
the study employs other metrics such as precision, recall, from ImageNet to enhance classification accuracy. Testing on
f1-score, and balanced accuracy to address imbalanced the AiCOLO colon cancer dataset yields impressive results,
dataset challenges, ensuring robust performance in tumor with ResNet achieving up to 96.98% accuracy. Additionally,
classification. pixel-wise segmentation strategies employing UNet and
The research [18] addresses the challenging task of breast SegNet models showcase significant improvements, with
cancer detection, considering its significant impact as a SegNet achieving up to 81.22% accuracy. Evaluation of CRC-
leading cause of female cancer mortality. Introducing the 5000, nct-crc-he-100k, and Warwick datasets demonstrates
DL-assisted Efficient Adaboost Algorithm (DLA-EABA), the robustness of the proposed methods. This study provides
the research employs advanced computational techniques valuable insights into effective network selection and training
to enhance breast cancer classification. It explores the strategies for colon tumor segmentation in histopathological
utilization of deep CNNs for tumor classification across images.
various imaging modalities like MRI, ultrasound, digital This study [22] addresses the challenge of diagnosing
breast tomosynthesis, and mammography. The DL frame- hematopoietic malignancies through diagnostic histopathol-
work integrates convolutional layers, Long Short-Term ogy, requiring labor-intensive slide reading with near-perfect
Memory(LSTM), and max-pooling layers, culminating in a accuracy. Despite AI’s assistance, achieving clinically usable
fully connected layer and softmax layer for classification diagnostic accuracy remains elusive due to dataset size and
and error estimation. By combining ML approaches with variation handling requirements. The research establishes a
feature selection and extraction methods, the study achieves highly accurate DL platform using smaller datasets, focusing

94118 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

on human diffuse large B-cell Lymphoma (DLBCL) and

non-DLBCL images from three hospitals. Multiple classify
pathologic images with near-perfect accuracy (100% for
hospital A, 99.71% for hospital B, and 100% for hospital
C). While technical variability initially affects cross-hospital
performance, eliminating it maintains 100% diagnostic
accuracy. This work demonstrates the clinical practicality of
utilizing DL models for DLBCL diagnosis and potentially
other hematopoietic malignancies.
The objective of this study [23] is to develop a lightweight
deep CNN tailored for binary classification of oral lesions
as either benign or malignant/potentially malignant using
real-time clinical images. A small CNN leveraging a pre-
trained EfficientNet-B0 model is proposed. Training and
testing are conducted using a dataset of 716 clinical images.
FIGURE 1. Dataset selection.
Performance evaluation metrics include accuracy, specificity,
sensitivity, receiver operating characteristics (ROC), and area
under curve (AUC). The proposed CNN model achieves
promising results, with an accuracy of 85.0%, specificity step to leverage the combined strengths of individual
of 84.5%, sensitivity of 86.7%, and AUC of 0.928. These models. Finally, we propose a fusion of CNN, LSTM, and
findings suggest that deep CNNs can effectively support K-Nearest Neighbors(KNN) models including the integration
the development of low-cost embedded vision devices for of Principal Component Analysis(PCA) to improve lym-
oral cancer diagnosis, thereby enhancing the quality and phoma cancer subclass classification.
accessibility of oral cancer screening and early detection.
A. MULTICANCER IMAGE DATASET
A. RESEARCH GAP AND QUESTIONS In this research, we conducted a comprehensive analysis
In the field of early detection of cancer, the literature of two distinct Multicancer image datasets sourced from
review identifies a significant research gap in the domain Kaggle [24]. The first dataset consists of seven main classes
of multi-cancer detection, particularly in multiple classes and 23 subclasses, while the second dataset comprises eight
and subclasses at the same time, classification accuracy, main classes and 26 subclasses shown in Table 2. The
adaptability, and efficiency. Our research addresses two list of the main classes and their subclasses is given in
primary research questions we have identified from literature Table 3. A notable disparity between these datasets lies in
analysis: the resolution of the images, with the first dataset containing
• Our proposed novel model can simultaneously detect lower-resolution images compared to the higher-resolution
multiple types of cancer, enhancing the scope of images present in the second dataset.
diagnosis of multi-cancer and potentially improving the Our approach involved initially focusing on fine-tuning
efficiency of cancer screening processes. our models and optimizing hyperparameters using the
• Our study investigates the most fruitful ML and lower-resolution dataset, as depicted in Figure 1. This
DL approaches for the efficient detection of multiple decision was primarily driven by computational constraints.
cancers. Processing over 100,000 images from the higher-resolution
To bridge this gap, our paper introduces an advanced dataset would have imposed substantial demands on GPU
technique. Our novel method aims to enhance accuracy and resources and time. By starting with the lower-resolution
efficiency, catering to the detection of Cancer. By integrating dataset, we aimed to streamline our experimentation process,
diverse CNNs and DL techniques, the proposed novel method iteratively refining our models and methodologies before
offers an efficient and robust solution in the evolving scaling up to the higher-resolution dataset.
landscape of Cancer detection. To facilitate our analysis, we strategically selected rep-
resentative samples from each dataset to ensure adequate
III. PROPOSED METHODOLOGY coverage of the main classes and subclasses. This sampling
In this section, we first explain the dataset used in this strategy allowed us to maintain a balanced distribution of
study and the image preprocessing methods we applied classes and subclasses while mitigating the computational
to the dataset. We then propose our novel approach by burden associated with processing large datasets.
integrating CNN models and LSTM to enhance multicancer Moving forward, we plan to utilize the insights gained
detection and classification, specifically VCEPTION (VGG from our experimentation with the lower-resolution dataset
+ Inception) and VMOBILENET (VGG + MobileNet) to inform our approach to the higher-resolution dataset.
for our merged CNN with LSTM framework. We also By establishing a robust foundation with the lower-resolution
introduce the Exclusively-OR(X-OR) gate as post-processing data, we aim to optimize our models’ performance and

VOLUME 12, 2024 94119

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

TABLE 1. Summary of recent studies on cancer research.

TABLE 2. Dataset information. complex image data [27]. The architecture of CNNs shown
in Figure 3, characterized by its convolutional layers, pooling
operations, and fully connected layers, forms the bedrock
of our methodological approach [28]. By meticulously
analyzing medical images through multiple filters, CNNs
can detect subtle cues indicative of cancerous changes,
generalizability before transitioning to the more complex thus providing a robust framework for early detection
and computationally intensive task of working with higher- [29]. Beyond spatial analysis, understanding temporal and
resolution images. sequential data is paramount in the medical field. LSTM
networks shown in Figure 4, a sophisticated variant of
recurrent neural networks (RNNs), offer a window into the
B. IMAGE PROCESSING TECHNIQUES
temporal dynamics of cancer progression. LSTMs are adept
To enhance the quality and features of the multicancer at handling sequences of data, making them ideal for ana-
images, a series of image processing techniques were lyzing series of medical images or patient histories to detect
applied. Morphological operations were applied to refine the patterns or changes over time that may indicate the onset of
representation of cancerous regions. Erosion reduced noise cancer.
and fine-tuned object boundaries, while dilation enlarged In our study, we innovatively combine CNNs with LSTM
boundaries and connected broken structures, enhancing the networks. The fusion of these technologies creates a robust
overall image quality. Subsequently, image enhancement framework for evaluating medical images, significantly
techniques, including histogram equalization and contrast enhancing the precision of cancer detection by leveraging the
stretching, were employed to improve visibility and aid distinct advantages of each approach for a holistic analysis.
feature extraction. Gaussian blur is a low-pass filtering tech- This hybrid approach ensures a thorough examination of
nique, that was then applied to reduce noise and emphasize cancer-related imagery by evaluating both the spatial details
important features [25]. The blurring operation contributed and the temporal progression of potential cancerous forma-
to the overall smoothing of the images, preserving essential tions. The integration strategy involves placing the LSTM
structures. Finally, Fourier transforms analysis [26] was units after the convolutional layers, as depicted in Figure 5,
applied to gain insights into the frequency domain, providing thereby facilitating a deeper and more understanding of the
a representation of image features. The segmentation process, data. This figure illustrates the seamless integration of a
utilizing contour detection, highlighted relevant regions of CNN with an LSTM unit, showcasing how this combination
interest, contributing to the overall understanding of the can significantly enhance the model’s ability to identify and
multicancer dataset. Figure 2 shows different types of image classify cancerous lesions by effectively leveraging spatial
processing that are applied in multicancer image datasets. and temporal data dimensions.

C. ENHANCING CANCER DETECTION THROUGH

CNN-LSTM INTEGRATION 1) EXPLORING THE MERGE CNNS
CNNs, drawing inspiration from the hierarchical structure of Our novel approach involves merging CNNs, where multiple
the human brain, have revolutionized the way we process CNNs are combined, and their outputs are merged after

94120 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

TABLE 3. Main cancer classes and subclasses.

FIGURE 2. Combined image processing steps for cancer detection.

convolution and flattening. This strategy enriches the model’s or addition, before proceeding with further convolutional
understanding by providing a holistic view of the patient’s layers. By integrating multiple CNNs within the merge
condition, thereby enhancing the accuracy of cancer detec- framework, our model can effectively capture complementary
tion. In the merge CNN framework, as illustrated in Figure 6, features from various modalities or perspectives, leading
each input image passes through multiple CNNs separately, to more robust and accurate predictions in multiple cancer
and their outputs are merged, typically through concatenation detection tasks.

VOLUME 12, 2024 94121

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

different, we use a different CNN-LSTM model and classify

the subclass accordingly.
Figure 8 represents the second strategy, where the model
classifies the main and subclass together. After this classifica-
tion, we select the two best models. Then, we apply the XOR
gate operation to these two models. If both models predict the
same output, we consider this classification. However, if the
two models predict different outputs, we skip the image. This
approach ensures a more precise classification by considering
the agreement between the two models.
FIGURE 3. Fundamentals of CNN architecture. For tighter classification, we partition the dataset into
three subsets: training (80%), validation, and test data (10%
each). These subsets undergo various image processing
techniques such as normalization, resizing, and Gaussian
blur before being fed into our merged CNN with LSTM
architecture. Through rigorous experimentation, we identify
the optimal models—VCEPTION (VGG + Inception) and
VMOBILENET (VGG + MobileNet)—for our merged CNN
with LSTM framework. We introduce an X-OR gate mecha-
nism between these models to enhance performance, ensuring
consensus in disease prediction and disregarding cases
where the models provide conflicting outputs. In hierarchical
classification, we follow a similar procedure of dataset
splitting and employ merged CNN with LSTM models. If the
main class prediction indicates Lymphoma, we apply PCA
followed by KNN classification. Otherwise, we proceed with
FIGURE 4. Intricacies of LSTM architecture. the standard merged CNN with the LSTM model. This
methodology, characterized by meticulous dataset handling,
innovative model architectures, and strategic classification
approaches, promises heightened accuracy and performance
2) MULTIMODAL FUSION WITH LSTM
in multicancer image classification. Figure 6 illustrates the
In this variant, our exploration extends to a multimodal model architecture. Beginning with dataset selection, the
CNN architecture, integrating different pre-trained CNNs image processing unfolds in two sections: a vertical and a hor-
where the input data consists of images. The incorporation of izontal side. The vertical section represents the simultaneous
LSTM units further augments the model’s ability to grasp the output model, where two CNNs process the image, merge,
temporal dynamics inherent in multimodal information. This and then undergo LSTM processing. Subsequently, a dense
enhancement results in improved performance, particularly layer and dropout are applied before jointly classifying the
for tasks demanding a comprehensive understanding of main class and subclass. Conversely, the horizontal section
diverse data sources. Following the LSTM layer, a dense layer depicts an alternative classification approach. Here, two
and dropout are applied to mitigate overfitting. CNNs merge, pass through an LSTM layer, predict the main
class, and, based on this main class, undergo another two
D. PROPOSED NOVEL MODEL merged CNNs. Another LSTM layer is then applied, enabling
Our proposed methodology for multicancer image classifi- the model to predict specific main class subtypes.
cation, illustrated in Figures 7 and 8, involves a compre-
hensive approach aimed at detecting both main classes and
subclasses. E. APPLIED X-OR GATE AS A POST PROCESSING
Figure 7 represents our first strategy, where we initially Incorporating X-OR logic into a medical imaging classi-
conduct model training on the lower-resolution dataset to fication system that utilizes two different models is an
identify the most effective model. Subsequently, we apply innovative approach aimed at enhancing diagnostic accuracy
this optimized model to the higher-resolution dataset for and reliability. This method addresses a critical aspect of
improved accuracy and performance. In strategy one, the medical diagnostics: ensuring the correct identification of
focus is primarily on classifying the main classes. Once diseases, as misclassification can lead to inappropriate patient
a main class is identified, we further classify the subclass management with potentially grave consequences.
with respect to the main class. For instance, if the main The principle of the X-OR gate [30], which outputs true or
class is classified as LYMPHOMA, we apply PCA and high only when the inputs differ, is central to this approach.
KNN techniques. On the other hand, if the main class is By applying X-OR logic to the predictions from two distinct

94122 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 5. Integration of CNN with LSTM for advanced cancer detection.

FIGURE 6. Merge CNN framework.

FIGURE 7. Strategy 1.

classification models for the same input image, the system Disease 1 and Model B predicts Disease 2 for the same image,
can identify cases of disagreement or confusion between the the X-OR gate would output a signal indicating a discrepancy,
models as shown in Table 4. For instance, if Model A predicts suggesting that the case requires further review or should

VOLUME 12, 2024 94123

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 8. Strategy 2.

TABLE 4. X-OR output table. TABLE 5. Hyperparameter settings.

(when the models disagree), as indicated by the X-OR

gate’s output. This system design significantly increases the
diagnostic system’s overall performance and trustworthiness
by ensuring that only cases with consistent model predictions
are accepted without question, thereby minimizing the risk of
misdiagnosis.

F. REFINING DATA WITH DIMENSIONALITY REDUCTION

VIA PCA
In dealing with high-dimensional image data, PCA serves
as a critical preprocessing step, streamlining the dataset to
FIGURE 9. X-OR logic in medical image classification.
its most informative features. This dimensionality reduction
not only simplifies the computational task but also amplifies
be classified into a confusion category. This methodology is the model’s focus on the most significant aspects of cancer
depicted in Figure 9, which illustrates how predictions from detection. By reducing the dimensionality of the image data,
the two models are input into an X-OR gate. The outcome of PCA helps mitigate the curse of dimensionality, making the
this process is either a confirmation of a consensus diagnosis dataset more manageable for ML algorithms such as KNN,
(when both models agree) or a flag for further investigation Support Vector Machine(SVM), etc.

94124 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 10. 2D t-SNE visualization main cancer and sub cancer.

TABLE 6. Model performance metrics for strategy 1.

Mathematically, PCA identifies the principal components Cancer, Lymphoma, Oral Cancer, Brain Cancer, and Cervical
v1 , v2 , . . . , vd , which are the eigenvectors corresponding to Cancer.
the largest eigenvalues of the covariance matrix 6, capturing
the maximum variance in the data. By projecting the original
image data X onto the subspace defined by the principal G. CNN + LSTM + KNN FUSION
components V using the equation In our secondary classification approach, mirroring our
initial methodology, we embarked on an additional inves-
XPCA = XV (1) tigation to delve deeper into our model’s performance.
We identified a discrepancy in the effectiveness of sub-
PCA retains the essential information while reducing the class classification within the Lymphoma category when
data’s dimensionality, making it particularly beneficial for employing individualized CNN analyses for each main class.
tasks such as skin cancer detection from dermoscopic images. To address this observation comprehensively, we extended
our analysis by employing various ML models separately,
1) IMAGE IN 2D(T-SNE) including the integration of PCA. Our findings revealed
In Figure 10, we present a comprehensive visual exploration that Lymphoma subclass classification exhibited superior
of a multicancer image dataset using t-distributed Stochastic accuracy with KNN compared to CNN-based approaches.
Neighbor Embedding (t-SNE), a method renowned for This nuanced exploration underscores the importance of
its ability to effectively reduce the dimensionality of considering alternative methodologies within specific sub-
high-dimensional data for visualization. This figure artfully classifications, offering valuable insights into optimizing
segregates the dataset into two distinct halves: the left side cancer classification strategies. In this section, we present
meticulously showcases the array of subclasses, while the a novel approach that combines CNN, LSTM networks,
right side brings into focus the primary cancer categories. and KNN to enhance classification performance. Initial
Figure 11 offers a closer look at eight specific cancer experiments involved applying distinct CNN models to each
subtypes, including Breast Cancer, Colon Cancer, Kidney main class independently, revealing consistent success across

VOLUME 12, 2024 94125

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 11. t-SNE visualization of multicancer image data.

most classes, except for Lymphoma, where CNN exhibited for the initial prediction of the main class. If Lymphoma is
suboptimal accuracy. To address this challenge, we propose identified as the primary prediction, we further refine the
a multi-stage approach. We first merge the CNN with LSTM classification using PCA combined with KNN. Conversely,

94126 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 12. Illustration of the integrated CNN-LSTM-KNN architecture.

TABLE 7. Model performance metrics with KNN for strategy 1.

TABLE 8. Model performance metrics - combined CNNs for strategy 2.

TABLE 9. Model performance metrics - combined models for strategy 2.

for classes other than Lymphoma, the input is directed where,

through an alternative CNN with LSTM. In addition to the - ŷq : Predicted class label for data point xq .
CNN-LSTM-KNN architecture, we explore the integration - Nq : Set of k nearest neighbors of xq .
of PCA with various traditional ML models, including DT, - yi : Class label of data point xi in the training dataset D.
SVM, and KNN. Interestingly, the PCA-KNN combination - I(·): Indicator function that returns 1 if the condition is true
exhibits superior performance specifically for Lymphoma and 0 otherwise.
compared to other pre-trained CNN models. To visually
represent our proposed architecture, refer to Figure 12. H. HYPERPARAMETER SETTING
This Figure illustrates the sequential flow of information The custom Adam optimizer extends the standard Adam
through the integrated CNN, LSTM, and KNN model, algorithm to accommodate specific optimization require-
showcasing decision points where Lymphoma triggers the ments [32]. One potential modification involves introducing
secondary PCA-KNN pathway, while other classes proceed an additional regularization term into the parameter update
through an alternative CNN-LSTM route. The mathematical rule to mitigate overfitting [33]. The custom Adam update
representation [31] of KNN shows: rule with regularization is given by:
X
ŷq = argmaxy I(yi = y) (2) mt = β1 mt−1 + (1 − β1 )gt ,
xi ∈Nq vt = β2 vt−1 + (1 − β2 )g2t ,

VOLUME 12, 2024 94127

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 13. Different merge model accuracy and loss for main and sub class.

TABLE 10. Lymphoma and subclass accuracy using various models with TABLE 11. Classification report for main class by Vmobilenet.
PCA.

mt
m̂t = ,
1 − β1t
vt
v̂t = ,
1 − β2t
!
m̂t
θt = θt−1 − α + λR(θt−1 ) . (3) encompass input shape, dense layer configurations, dropout
v̂t + ϵ
p
rate, optimization strategy, and evaluation metrics. These
where α is the learning rate, β1 and β2 are exponential decay hyperparameter choices are pivotal in fine-tuning the models
rates for moment estimates, gt is the gradient of the loss for optimal performance, particularly in medical image
function, ϵ is a small constant for numerical stability, λ is analysis tasks like skin cancer detection, where accuracy and
the regularization parameter, and R(θt−1 ) is a regularization reliability are paramount.
function applied to the model parameters θ.
Table 5 provides a comprehensive overview of the hyper- IV. RESULT AND DISCUSSION
parameter settings utilized in the experimental configuration Table 6 shows how different ML models perform in image
for the KNN and CNN classification methods. For KNN, classification tasks, giving a clear picture of their strengths
parameters such as PCA value, image size, and the number and weaknesses. DL models such as InceptionV3 and
of nearest neighbors are specified, while for CNN, details EfficientNetB0 exhibit superior performance, particularly in

94128 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

TABLE 12. Classification report for main class by Vception.

TABLE 13. Classification report for main class after applying X-OR gate in
two models.

handling sub-class accuracy, which is a testament to their

robust feature extraction and generalization capabilities. The
Table also highlights the challenges faced by custom convolu-
tional networks, such as ConvNetxTiny and ConvNetxLarge,
which lag behind the more sophisticated pre-trained models
in both main and sub-class accuracy metrics. Furthermore,
traditional ML models like KNN, SVM, and DT show
significantly lower performance, underscoring the advantage
of DL models in complex visual recognition tasks. This
comparative analysis emphasizes the importance of model
selection based on the specific requirements of the classifica-
tion task and the nature of the dataset, highlighting the need
for further research into optimization strategies to enhance
model performance and generalizability across diverse tasks.
Table 7 investigates the efficacy of combining KNN with
various CNN models for multicancer classification. The
study introduces hybrid models, notably Inception + KNN
and MobileNet + KNN, to utilize the DL capabilities
of CNNs alongside the simplicity and interpretability
of KNN.
Table 8 showcases the performance metrics of various
combined CNN models for an unspecified classification task. FIGURE 14. Confusion matrices (a)Vmobilenet main class (b)Vception
The combinations include VGG + ResNet, Inception + main class (c)After applying X-OR gate for the main class before and after
applying the X-OR gate.
ResNet, VGG + Xception, Inception + Xception, and
ResNet + Xception. These models are evaluated based on
four key metrics: Main Class Accuracy, Sub Class Accuracy,
Test Main Class Accuracy, and Test Sub Class Accuracy. with Inception + ResNet showing the highest Main Class
The table reveals a general trend where combining different Accuracy on both training and test data. However, the Sub
CNN architectures results in varied performance outcomes, Class Accuracies are notably lower across all combinations,

VOLUME 12, 2024 94129

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

FIGURE 15. Confusion matrices (a)Vmobilenet subclass (b)Vception subclass (c)After applying X-OR gate for the subclass before and after applying the
X-OR gate.

suggesting challenges in capturing the finer distinctions 20 epochs. Notably, the merge models ‘‘VGG-Inception’’ and
within classes. This analysis highlights the potential and ‘‘VGG-MobileNet’’ exhibit particularly impressive accuracy
limitations of architecturally diverse CNN combinations in results. Specifically, for the main class classification, ‘‘VGG-
enhancing model performance for complex classification Inception’’ achieves an outstanding accuracy of 99.83%
tasks. and 99.81% with corresponding loss values of 0.0059 and
Table 9 summarizes the performance of various combi- 0.0068, respectively. These values underscore the exceptional
nations of DL models, which merge LSTM networks with performance of ‘‘VGG-Inception’’ in accurately classifying
popular CNN architectures such as VGG, ResNet, Inception, main class labels.
Xception, and MobileNet. The combination of VGG +
Inception + LSTM, in particular, distinguishes itself by
securing the highest accuracies in test scenarios for both A. ACCURACY OF LYMPHOMA
category levels, achieving 99.13% and 97.80%, respectively. As CNN exhibits lower accuracy in the classification of
This indicates that merging LSTM with CNN models Lymphoma for strategy 1 in Figure7, our study delves
successfully harnesses spatial and sequential data attributes, into the classification accuracies of various ML models
thereby bolstering the models’ predictive performance on applied to both Lymphoma and its subclasses. Emphasizing
new data. the role of PCA, we investigate its impact on model
Figure 13 illustrates the variations in merge model performance using the top 50 pixels extracted from medical
accuracy and loss graphs for both the main class and subclass. images. The results presented in Table 10 highlight the
Additionally, it showcases the maximum accuracy gain and robustness and effectiveness of the KNN model, achieving an
minimum loss values achieved by each merge model within impressive accuracy of 97.14% for Lymphoma classification.

94130 VOLUME 12, 2024

J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

TABLE 14. Classification report for subclass by Vmobilenet. TABLE 15. Classification report for subclass by Vception.

TABLE 16. Classification report for subclass after applying X-OR gate in
This improvement is particularly striking in the KNN two models.
model, surpassing other models across all subclasses. This
underscores the KNN model’s potential for accurate and
reliable Lymphoma classification.

B. EFFECT OF X-OR GATE AT STRATEGY 2

Our research meticulously evaluated the two best CNN
models, Vmobilenet and Vception, for cancer classification.
At first, the two models did not reach the highest accuracy.
But after adding an X-OR gate later, their accuracy and
precision greatly improved. This strategy resulted in notable
enhancements in both main and subclass classifications,
as evidenced by the confusion matrices. The X-OR gate’s
implementation refined accuracy by leveraging the com-
bined strengths of the individual models. This led to a
decrease in misclassifications and improved clarity in class
differentiation. Notably, the X-OR gate had a substantial
impact on precision, recall, and F1 scores, highlighting the
advantage of merging models to address the complexities of
cancer detection tasks. Our study conclusively shows that
integrating an X-OR gate between two CNN models not only
addresses individual model limitations but also outperforms
our proposed model in accuracy. This underscores the
effectiveness of our approach in enhancing the reliability observed for Vmobilenet and Vception, respectively, which,
and precision of medical diagnostics, specifically in cancer after applying the X-OR gate, increased to 99.13%. These
classification. results underscore the efficacy of combining models through
Figures 14 and 15 illustrate the confusion matrices for X-OR logic in enhancing the precision of medical image
both the main and subclass classifications before and after classification.
the application of the X-OR gate. The implementation of the Following this, The Tables accompanying this study
X-OR gate led to a main class accuracy nearing 99.95%, provide a comprehensive insight into the classification
a substantial improvement over the initial accuracies of performance of the Vmobilenet and Vception models, both
99.25% for Vmobilenet and 99.83% for Vception. Similarly, individually and after integrating an X-OR gate. Table 11
for the subclass, accuracies of 97.66% and 97.80% were and Table 14 detail the classification reports for the main and
VOLUME 12, 2024 94131
 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

TABLE 17. State of the art on cancer diagnosis.

methods developed in the past year. As shown in Table 17,

other models primarily focus on a single class of cancer,
whereas our model targets eight different types of cancer
and their subtypes. The accuracy of our proposed model is
99.25% for main classes and 97.80% for sub-classes.

V. CONCLUSION AND FUTURE WORK

In this study, we developed a novel approach for enhancing
the accuracy and reliability of multiple cancer detection
using images. We combined CNN models with LSTM
networks and applied PCA and KNN, finding PCA and
KNN together sometimes outperformed CNNs. The rigorous
evaluation showed our methodology improved accuracy,
FIGURE 16. False images before applying X-OR gate.
precision, recall, and F1-scores for cancer classifications.
Post-processing with an X-OR gate between models also
enhanced performance, even when individual model accuracy
sub-classes using the Vmobilenet model, respectively. Simi- was not high. Our findings highlight the potential of ensemble
larly, Table 12 and Table 15 present the performance metrics learning and logical operations in complex classification
for the Vception model. Notably, the strategic application of tasks. However, challenges remain, such as accurately
an X-OR gate to combine the outputs of these models resulted classifying main and subclass cancer types together due
in marked improvements, as highlighted in Table 13 and 16. to independent nodes, which may lead to contradictory
These Tables showcase enhanced accuracy, precision, recall, classifications. Additionally, using an X-OR gate introduces
and F1 scores, further substantiating the efficacy of model a ‘‘confusion section’’ for difficult-to-classify images, and
ensembling techniques in complex classification tasks. the hierarchical classification process can mislead subclass
In Figure 16, three subfigures illustrate the comparative predictions if the main class is incorrectly predicted.
performance of the Vmobilenet and Vception models in In our study, we applied two different approaches to
image classification tasks. Subfigure 16a shows images cancer classification, revealing significant shortcomings. For
that were misclassified by Vception but correctly identified instance, the model might predict a main class and a subclass
by Vmobilenet, highlighting the instances where Vmo- that are inherently contradictory due to their independent
bilenet outperforms Vception. Conversely, Subfigure 16b classification nodes—such as identifying the main class as
includes images accurately classified by Vmobilenet but not brain cancer but the subclass as belonging to cervical cancer.
by Vception, indicating similar discriminative capabilities In our study, we found a challenging part when working with
between the models. Lastly, Subfigure 16c depicts images an X-OR gate that made it hard to deal with some images.
that both models failed to classify correctly, underscoring Out of the 12,001 images we tested, we couldn’t easily
the challenges or limitations faced by each model in certain classify 100 main categories and 821 subcategories because
scenarios. This arrangement effectively demonstrates the they were unclear. This brings up the issue of how to handle
strengths and weaknesses of each model in processing and these uncertain cases. Also, our method first tries to figure
recognizing various image types. out the main category of an image before identifying the
more specific subcategory related to that main category. This
C. COMPARISON WITH STATE-OF-THE-ART STUDIES FOR approach can lead to incorrect subclass predictions if the main
STRATEGY 2 class is wrongly identified, given the subclass’s dependence
For a fair comparison, we evaluated the performance scores on the main class accuracy. In future work, we aim to address
of our novel proposal against state-of-the-art methods. This these limitations to enhance the reliability and coherence of
review encompassed a broad spectrum of cutting-edge our cancer classification model.
94132 VOLUME 12, 2024
J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

ACKNOWLEDGMENT [17] S. A. Abdelaziz Ismael, A. Mohammed, and H. Hefny, ‘‘An enhanced

The authors would like to acknowledge Princess Nourah deep learning approach for brain cancer MRI images classification using
residual networks,’’ Artif. Intell. Med., vol. 102, Jan. 2020, Art. no. 101779.
bint Abdulrahman University Researchers Supporting Project [18] J. Zheng, D. Lin, Z. Gao, S. Wang, M. He, and J. Fan, ‘‘Deep learning
number (PNURSP2024R197), Princess Nourah bint Abdul- assisted efficient AdaBoost algorithm for breast cancer detection and early
rahman University, Riyadh, Saudi Arabia. The authors would diagnosis,’’ IEEE Access, vol. 8, pp. 96946–96954, 2020.
[19] S. Cheng et al., ‘‘Robust whole slide image analysis for cervical cancer
like to thank Prince Sultan University for their support. screening using deep learning,’’ Nature Commun., vol. 12, no. 1, p. 5639,
Sep. 2021.
[20] L. B. da Cruz, J. D. L. Araujo, J. L. Ferreira, J. O. B. Diniz, A. C. Silva,
REFERENCES J. D. S. de Almeida, A. C. de Paiva, and M. Gattass, ‘‘Kidney segmentation
from computed tomography images using deep neural network,’’ Comput.
[1] M. Nahiduzzaman, L. F. Abdulrazak, M. A. Ayari, A. Khandakar, and Biol. Med., vol. 123, Aug. 2020, Art. no. 103906.
S. M. R. Islam, ‘‘A novel framework for lung cancer classification using [21] A. Ben Hamida, M. Devanne, J. Weber, C. Truntzer, V. Derangère,
lightweight convolutional neural networks and ridge extreme learning F. Ghiringhelli, G. Forestier, and C. Wemmert, ‘‘Deep learning for colon
machine model with Shapley additive explanations (SHAP),’’ Expert Syst. cancer histopathological images analysis,’’ Comput. Biol. Med., vol. 136,
Appl., vol. 248, Aug. 2024, Art. no. 123392. Sep. 2021, Art. no. 104730.
[2] I. Houda, C. Dickhoff, C. A. Uyl-de Groot, R. A. M. Damhuis, N. Reguart, [22] D. Li, J. R. Bledsoe, Y. Zeng, W. Liu, Y. Hu, K. Bi, A. Liang, and S. Li, ‘‘A
M. Provencio, A. Levy, R. Dziadziuszko, C. Pompili, M. D. Maio, deep learning diagnostic platform for diffuse large B-cell lymphoma with
M. Thomas, A. Brunelli, S. Popat, S. Senan, and I. Bahce, ‘‘Challenges high accuracy across multiple hospitals,’’ Nature Commun., vol. 11, no. 1,
and controversies in resectable non-small cell lung cancer: A clinician’s p. 6004, Nov. 2020.
perspective,’’ Lancet Regional Health—Eur., 2024. [23] F. Jubair, O. Al-karadsheh, D. Malamos, S. A. Mahdi, Y. Saad, and
[3] S. Saeedi, S. Rezayi, H. Keshavarz, and S. R. Niakan Kalhori, ‘‘MRI- Y. Hassona, ‘‘A novel lightweight deep convolutional neural network
based brain tumor detection using convolutional deep learning methods for early detection of oral cancer,’’ Oral Diseases, vol. 28, no. 4,
and chosen machine learning techniques,’’ BMC Med. Informat. Decis. pp. 1123–1130, May 2022.
Making, vol. 23, no. 1, p. 16, Jan. 2023. [24] O. S. Naren. (2022). Multi Cancer Dataset: 8 Types of Cancer Images.
[4] K. K. Anilkumar, V. J. Manoj, and T. M. Sagi, ‘‘A review on computer [Online]. Available: [Link]
aided detection and classification of leukemia,’’ Multimedia Tools Appl., multi-cancer
vol. 83, no. 6, pp. 17961–17981, Jul. 2023. [25] L. Fu, J. Zhu, W. Li, Q. Zhu, B. Xu, Y. Xie, Y. Zhang, Y. Hu, J. Lu, P. Dang,
[5] P. K. Das, D. V. A. S. Meher, R. Panda, and A. Abraham, ‘‘A systematic and J. You, ‘‘Tunnel vision optimization method for VR flood scenes
review on recent advancements in deep and machine learning based based on Gaussian blur,’’ Int. J. Digit. Earth, vol. 14, no. 7, pp. 821–835,
detection and classification of acute lymphoblastic leukemia,’’ IEEE Jul. 2021.
Access, vol. 10, pp. 81741–81763, 2022. [26] R. Valand, S. Tanna, G. Lawson, and L. Bengtström, ‘‘A review of Fourier
[6] H. Abolhassani, A. Eskandari, A. Saremi Poor, A. Zarrabi, B. Khodadadi, transform infrared (FTIR) spectroscopy used in food adulteration and
S. Karimifard, H. Sahrayi, M. Bourbour, and M. Tavakkoli Yaraki, authenticity investigations,’’ Food Addit Contaminants, Part A, vol. 37,
‘‘Nanobiotechnological approaches for breast cancer management: Drug no. 1, pp. 19–38, Jan. 2020.
delivery systems and 3D in-vitro models,’’ Coordination Chem. Rev., [27] M. S. Islam, M. A. T. Rony, and T. Sultan, ‘‘GastroVRG: Enhancing early
vol. 508, Jun. 2024, Art. no. 215754. screening in gastrointestinal health via advanced transfer features,’’ Intell.
[7] M. F. Ak, ‘‘A comparative analysis of breast cancer detection and diagnosis Syst. Appl., vol. 23, Sep. 2024, Art. no. 200399.
using data visualization and machine learning applications,’’ Healthcare, [28] J. O. Bappi, M. A. T. Rony, and M. S. Islam, ‘‘BNVGLENET:
vol. 8, no. 2, p. 111, Apr. 2020. Hypercomplex Bangla handwriting character recognition with hierarchical
[8] F. R. Burdier, D.-E.-N. Waheed, B. Nedjai, R. D. M. Steenbergen, class expansion using convolutional neural networks,’’ Natural Lang.
M. Poljak, M. Baay, A. Vorsters, and S. Van Keer, ‘‘DNA methylation as Process. J., vol. 7, Jun. 2024, Art. no. 100068.
a triage tool for cervical cancer screening—A meeting report,’’ Preventive [29] M. S. Islam and M. A. T. Rony, ‘‘CDK: A novel high-performance transfer
Med. Rep., vol. 41, May 2024, Art. no. 102678. feature technique for early detection of osteoarthritis,’’ J. Pathol. Informat.,
[9] M. M. Kalbhor and S. V. Shinde, ‘‘Cervical cancer diagnosis using vol. 15, Dec. 2024, Art. no. 100382.
convolution neural network: Feature learning and transfer learning [30] R. Sardar, A. Nandi, A. Bhowmik, B. Dutta, and S. Ghosh, ‘‘Design of
approaches,’’ Soft Comput., Jul. 2023, doi: 10.1007/s00500-023-08969-1. EX-OR gate with ANN using sigmoid and ReLu functions for artificial
[10] K. Rajkumar, R. T. Sri Ramoju, T. Balelly, S. Ashadapu, C. R. Prasad, intelligence applications in Python,’’ in Proc. Int. Conf. Data Anal.
and Y. Srikanth, ‘‘Kidney cancer detection using deep learning models,’’ Insights, 2023, pp. 779–789.
in Proc. 7th Int. Conf. Trends Electron. Informat. (ICOEI), Apr. 2023, [31] S. Zhang, ‘‘Challenges in KNN classification,’’ IEEE Trans. Knowl. Data
pp. 1197–1203. Eng., vol. 34, no. 10, pp. 4663–4675, Oct. 2022.
[11] H. A. Mengash, M. Alamgeer, M. Maashi, M. Othman, M. A. Hamza, [32] R. B. Sarooraj and S. Prayla Shyry, ‘‘Analysis of traffic flow prediction
S. S. Ibrahim, A. S. Zamani, and I. Yaseen, ‘‘Leveraging marine predators from spatial–temporal data using hybrid GSA-adam optimizer based
algorithm with deep learning for lung and colon cancer diagnosis,’’ LSTM network for intelligent transport system,’’ Multimedia Tools Appl.,
Cancers, vol. 15, no. 5, p. 1591, 2023. vol. 83, no. 6, pp. 16735–16761, Jul. 2023.
[12] E. A. M. Zijtregtop, L. A. Winterswijk, T. P. A. Beishuizen, C. M. Zwaan, [33] A. D. Viniski, J. P. Barddal, A. de Souza Britto, and H. V. A. de Campos,
R. A. J. Nievelstein, F. A. G. Meyer-Wentrup, and A. Beishuizen, ‘‘Incremental specialized and specialized-generalized matrix factorization
‘‘Machine learning logistic regression model for early decision making models based on adaptive learning rate optimizers,’’ Neurocomputing,
in referral of children with cervical lymphadenopathy suspected of vol. 552, Oct. 2023, Art. no. 126515.
lymphoma,’’ Cancers, vol. 15, no. 4, p. 1178, Feb. 2023.
[13] H. Myriam, A. A. Abdelhamid, E. M. El-Kenawy, A. Ibrahim, M. M. Eid,
M. M. Jamjoom, and D. S. Khafaga, ‘‘Advanced meta-heuristic algorithm
based on particle swarm and al-biruni Earth radius optimization methods
for oral cancer detection,’’ IEEE Access, vol. 11, pp. 23681–23700, 2023.
[14] J. Dubroff, D. Johnson, and H. Botha, ‘‘State of art of molecular imaging
JABED OMAR BAPPI received the [Link]. degree
in the diagnosis and treatment of dementia,’’ Amer. J. Geriatric Psychiatry,
in electrical and electronics engineering from the
vol. 32, no. 4, p. 31, Apr. 2024.
[15] Y. Xue, S. Chen, J. Qin, Y. Liu, B. Huang, and H. Chen, ‘‘Application Department of Electrical and Electronic Engi-
of deep learning in automated analysis of molecular images in cancer: A neering (EEE), Port City International University,
survey,’’ Contrast Media Mol. Imag., vol. 2017, pp. 1–10, Jun. 2017. Chittagong, Bangladesh, in 2022. His current
[16] N. Sampathila, K. Chadaga, N. Goswami, R. P. Chadaga, M. Pandya, research interests include data science, artificial
S. Prabhu, M. G. Bairy, S. S. Katta, D. Bhat, and S. P. Upadya, intelligence, data mining, natural language pro-
‘‘Customized deep learning classifier for detection of acute lymphoblastic cessing, ML, DL, and image processing.
leukemia using blood smear images,’’ Healthcare, vol. 10, no. 10, p. 1812,
Sep. 2022.

VOLUME 12, 2024 94133

 J. O. Bappi et al.: Novel DL Approach for Accurate Cancer Type and Subtype Identification

MOHAMMAD ABU TAREQ RONY received SAMAH ALSHATHRI received the bachelor’s
the [Link]. degree in statistics from Noakhali degree in computer science and the master’s
Science and Technology University, Noakhali, degree in computer engineering from King Saud
Bangladesh. In addition, he possesses expertise University, Riyadh, Saudi Arabia, and the Ph.D.
in devising advanced analytics strategies using degree from the Department of Computer and
data. His diverse professional experience includes Mathematics, Plymouth University, Plymouth,
four years of research in areas, such as artificial U.K. She is currently an Assistant Professor
intelligence. He is currently a Research Data Sci- with the Department of Information Technology,
entist at aiQuest Intelligence, Dhaka, Bangladesh. College of Computer and Information Sciences,
Moreover, he actively engages in partnerships with Princess Nourah Bint Abdulrahman University
international researchers, recognizing that research plays an indispensable (PNU), Riyadh. Her research interests include wireless networks, cloud
role in fostering innovation. Overall, he is hardworking and has taught computing, fog computing, the IoT, data mining, machine learning, text
himself various skills, such as NLP, LLM, and Deep Learning. He has analytics, image classification, and deep learning. She was the Chair of the
published articles in refereed journals and conference proceedings, such as Network and Communication Department and participated in organizing
IEEE ACCESS, Elsevier, Springer, MDPI, and International Conferences. many international conferences. She has authored or coauthored many
articles published in well-known journals in the research field.

WALID EL-SHAFAI (Senior Member, IEEE)

was born in Alexandria, Egypt. He received the
[Link]. degree (Hons.) in electronics and electrical
communication engineering from the Faculty of
Electronic Engineering (FEE), Menoufia Univer-
sity, Menouf, Egypt, in 2008, the [Link]. degree
from Egypt–Japan University of Science and
Technology (E-JUST), in 2012, and the Ph.D.
degree from FEE, Menoufia University, in 2019.
Since January 2021, he has been a Postdoctoral
Research Fellow with the Security Engineering Laboratory (SEL), Prince
Sultan University (PSU), Riyadh, Saudi Arabia. He is currently a Senior
Cybersecurity Researcher with SEL and an Assistant Professor with the
College of Computer Science and Information Systems. He is also an
Associate Professor with the Department of Electronics and Communi-
cation Engineering (ECE), Faculty of Electronic Engineering, Menoufia
University. His research interests include wireless mobile and multimedia
communications systems, image and video signal processing, efficient 2D
MOHAMMAD SHARIFUL ISLAM received the video/3D multi-view video coding, multi-view video plus depth coding,
[Link]. degree in computer science and telecom- 3D multi-view video coding and transmission, quality of service and
munication engineering from Noakhali Science experience, digital communication techniques, cognitive radio networks,
and Technology University, Bangladesh, in 2023, adaptive filters design, 3D video watermarking, steganography, encryption,
with a focus on a deep passion for cutting-edge error resilience and concealment algorithms for H.264/AVC, H.264/MVC,
technologies to the research community. His and H.265/HEVC video codecs standards, cognitive cryptography, medical
academic journey, rooted in the confluence of image processing, speech processing, security algorithms, software-defined
computer science and telecommunications, has networks, the Internet of Things, medical diagnoses applications, FPGA
evolved into a fervent pursuit of specialized areas, implementations for signal processing algorithms and communication
including data science, machine learning, natural systems, cancellable biometrics and pattern recognition, image and video
language processing, and image processing. His work in these fields is driven magnification, artificial intelligence for signal processing algorithms and
by a quest to uncover hidden insights within data, develop intelligent learning communication systems, modulation identification and classification, image
algorithms, bridge the communication gap between humans and machines, and video super-resolution and denoising, cybersecurity applications,
and artistically enhance digital imagery. As a researcher, his approach is malware and ransomware detection and analysis, deep learning in signal
characterized by a blend of technical proficiency and creative problem- processing, and communication systems applications. He also serves as a
solving, aiming to contribute significantly to the frontiers of technology and reviewer for several international journals.
its application in understanding and improving their digital world.

94134 VOLUME 12, 2024