Received 30 July 2024, accepted 5 September 2024, date of publication 16 September 2024, date of current version 27 September 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3461878

Multimodal Non-Small Cell Lung Cancer

Classification Using Convolutional
Neural Networks
MARIAN MAGDY AMIN , AHMED S. ISMAIL , AND MASOUD E. SHAHEEN
Faculty of Computers and Artificial Intelligence, Fayoum University, Faiyum 2933110, Egypt
Corresponding author: Marian Magdy Amin ([email protected])

ABSTRACT Lung cancer is the leading cause of death worldwide. Early detection of lung cancer is
a hard mission. New Small Cell Lung Cancer (NSCLC) is the most prevalent sub-type of lung cancer.
Differentiating between several NSCLC subtypes is important for making the right decision of treatment
plan for the patient. Despite the focus of recent researchers on single modality approach, multi-omics
modalities have many underlying influences and discoveries in the cancer detection and classification
area. Through this research multi-omics modalities are used. Previous efforts have been focused either on
multimodality using traditional machine learning classifiers or single modality using deep learning. Also,
for the molecular sources (RNA-seq and miRNA-Seq) traditional machine learning approaches are usually
used. For this work, deep learning using Convolutional Neural Networks (CNNs) is used and applied on
the above-mentioned multimodalities. The classification accuracy results obtained for RNA-Seq, miRNA-
Seq, WSIs are 96.79%, 98.59%, 89.73% respectively. The F1 scores obtained for RNA-Seq, miRNA-Seq,
WSIs are 95.238%,99.67%,89.76% respectively. Moreover, the Area Under Curve obtained for RNA-Seq,
miRNA-Seq, WSIs are 100%, 99.41%,97,54% respectively. These results improves the results obtained by
other related works as will be explained. According to these improvements in the results, the lung cancer
classification could be better and the disease would be discovered at early stages which is the goal for the
research field efforts.

INDEX TERMS Lung cancer, convolutional neural networks, multimodality, molecular data, whole slide
images, deep learning, multiomics.

I. INTRODUCTION be early detected and received the appropriate treatment [3],

Lung cancer ranks as the primary cause of death among men [5]. However, cancer is often detected at late stages where
and the second leading cause among women globally [1], [2]. effective treatment and cures is unachievable [3]. The com-
It is widely acknowledged as one of the most lethal forms of plex nature of cancer molecular biology made it difficult for
cancer, surpassing breast, prostate, colorectal, stomach, and professionals to detect early signs and symptoms using tra-
liver cancer in terms of severity [3]. ditional diagnosis and screening techniques. Moreover, one
Efforts to detect lung cancer early have become paramount major challenge when using chemotherapy is maximizing the
due to its aggressive nature. Non-Small Cell Lung Cancer drug efficiency while minimizing the toxic effects of healthy
(NSCLC) accounts for the majority of cases, comprising cells. So, for obtaining successful treatment, accurate classifi-
approximately 80-85% of diagnoses [2], [4]. Despite the high cation for the tumor is needed. But classification of laboratory
mortality rates and the aggressive nature of cancer, many screenings depends on the insights of experts and morpho-
patients could be saved. As declared by the World Health logical appearance of the tumor. But this approach inherits
Organization 30-50% of cases can be avoided if they can serious limitations because of the similar histopathological
appearances of the tumors which have significant different
The associate editor coordinating the review of this manuscript and courses and responses to therapy [3]. As a result, deeper
approving it for publication was Chuan Li. approaches should be followed for accurate classification.
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
134770 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

Studies have been focused recently on the genetic molecular Deep learning has been massively used using single modal-
level for early diagnosis and early treatment plans. ity (images in most of cases) for detecting different types of
Within NSCLC, two predominant subtypes exist: Lung cancer. In this paper, multi-omics modalities are used for lung
Adenocarcinoma (LUAD) and Lung Squamous Cell Carci- cancer classification such as: RNA-Seq [2], [17]and miRNA-
noma (LUSC). LUAD typically develops in the peripheral Seq [2], [18]. Moreover, Whole Slide Images (WSIs) [2], [19]
lung tissue, whereas LUSC tends to manifest in cen- are used. So, in this paper multimodalities are used RNA-Seq,
tral locations [2]. Accurate differentiation between these miRNA-Seq, and WSIs.
subtypes is crucial as they necessitate distinct treatment Despite the focus of this paper on lung cancer, the
approaches [3], [6]. idea itself is effective when being applied on multiple
Next Generation Sequencing methods such as whole- large number of tumors (multi-scale, multi-modalities). The
genome DNA sequencing and total RNA sequencing, are architecture of the deep learning model requires minimal pre-
considered revolutionary technologies for studying genetic processing for multi-layer networks [20]. It utilizes the spatial
changes in cancer [7]. These technologies are promising in relationships within the data to decrease the dimensions that
accurate classification of tumor cells because of their ability need to be learned [15]. This approach speeds up the learning
of sequencing thousands of genes and detecting genomic process while yielding more precise results.
and transcriptomic alterations [3]. NGS achieves that through So, NGS using deep learning techniques is the suitable
comparing DNA and RNA in normal and cancer cells. It dis- solution for the problem of complexity and high dimensional-
covers the genetic changes that leads to abnormal activity ity. Convolutional Neural Networks are used in this work for
that leads to the presence of different levels of genes and both slide images and molecular sources (RNA-Seq (mRNA),
consequently proteins within the cells [3], [8]. This is known miRNA-seq) for dealing with high dimensionality problem.
as gene expression. Differentially expressed genes within the Through this research several past efforts will be shown
cell gives great insight of the motive of tumor growth [3], [8]. through the related works to highlight the importance of
Gene expression has been the focus of recent researches reaching an accurate classification for this type of cancer.
in the field of cancer classification [3], [9]. Cell functions Also, the dataset used and the preprocessing done will be
are determined by individual proteins and the synthesis of discussed in the methodology section. Then, a detailed expla-
these proteins depends on which genes are expressed by the nation of the approach used of the convolutional neural
cell. Therefore, the gene expression gives some insight about networks will be illustrated in the implementation and discus-
the cell function [3]. Gene expression is the process of trans- sion section. After that, the results will be shown to compare
lating DNA into proteins and non-coding RNA. Microarray with the previously obtained results by other researches. Also,
shows limitation because of the incomplete snapshot of the future work points that could be made will be shown in the
transcriptome. Also, it cannot detect previously unidentified future work section.
genes or transcripts [3], [9]. Gene expression quantification,
therefore, is an effective alternative. It identifies which genes
are preferentially expressed in different tissues. So, in this II. RELATED WORKS
paper STAR counts data provided by TCGA program is used. In recent years, there has been significant interest in using
Through this paper two types of gene expression data machine learning models with biological data to aid in the
are used mRNA-Seq and miRNA-Seq. mRNA is involved diagnosis and prognosis of cancer patients, particularly in the
in conveying genetic information from DNA to ribosome, context of lung cancer. Gene expression data has been widely
where it is translated into proteins. While miRNA is a type explored for lung cancer classification, with studies achieving
of non-coding RNA which plays a role in gene regulation high accuracy rates. For instance, Smolander et al. achieved
by binding to target mRNAs and inhibiting their transla- a 95.97% accuracy in distinguishing LUAD from control
tion or promoting their degradation. Any abnormalities in samples using deep learning models applied to coding RNA
these processes of mRNA or miRNA leads to different gene data [21]. Similarly, Fan et al. used support vector machines
expressions which accordingly activates tumor growth [3], (SVMs) with a 12-gene signature to achieve a 91% accuracy
[10]. However, cancer classification using gene expression is in the same classification task [22].
very challenging given the complexity and massive amount Multiclass classification of lung cancer subtypes has also
of genetic data that is produced [3], [9], [11], [12]. The been addressed in the literature. Gonzales et al. developed a
magnitude of variant obtained from RNA-Sequencing for model for classifying small-cell lung cancer (SCLC), LUAD,
example is exponential which makes it difficult for tradi- LUSC, and large-cell lung carcinoma (LCLC) using dif-
tional machine learning methods and bioinformatics tools ferentially expressed genes (DEGs) as input, achieving an
to approach genetic variants for disease prediction [3], [13], accuracy of 88.23% with the random forest (RF) algorithm
[14]. Gene expression is known by high dimensionality, with [23]. Additionally, Castillo-Secilla et al. attained a 95.7%
w very large number of features representing genes and very accuracy in subtype classification of non-small cell lung
small number of training data representing patient samples cancer (NSCLC) using random forest [24]. Studies utilizing
[3], [13], [15]. Deep learning has been used recently for miRNA-Seq data for lung cancer classification have also been
dealing with this problem of high dimensionality [16]. conducted. Ye et al. identified a 10-miRNA signature for

VOLUME 12, 2024 134771

 M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

TABLE 1. Summary of the related works.

distinguishing LUSC from control samples, achieving an A. DATA ACQUISITION AND PRE-PROCESSING
F1 score of 99.4% [18]. In this work two molecular modalities and one imaging
Deep learning approaches, particularly convolutional neu- modality are considered: RNA-Seq, miRNA-Seq and Whole
ral networks (CNNs), have shown promise in combination Slide Images (WSIs). The data were collected from the Can-
with whole-slide imaging (WSI) for NSCLC subtype clas- cer Genome Atlas (TCGA) program [27] which is easily
sification. Coudray et al. utilized CNNs with tiles extracted accessible from the Genomic Data Commons (GDC) portal.
from WSI to classify LUAD, LUSC, and control samples, GDC is the National Cancer Institute’s (NCI) data shar-
achieving an impressive AUC score of 0.978 [19]. Further- ing platform. The GDC contains NCI-generated data from
more, Kanavati et al. used transfer learning with CNNs on some of the largest and most comprehensive cancer genomic
manually labeled images to distinguish lung carcinoma from datasets, including the Cancer Genome Atlas (TCGA) pro-
control samples, obtaining an AUC score of 0.988 [25]. gram. The data provided covers real-world scenarios which
At last, hybrid approaches combining deep learning with is important while dealing with such fatal disease.
traditional statistics have been explored. Graham et al. WSI images were downloaded as a zip file and SVS
utilized tiles extracted from images along with summary files were extracted for the pre-processing phase. For the
statistics to classify LUAD, control, and LUSC samples, molecular data (RNA-Seq and miRNA-Seq), TCGAbiolinks
achieving an accuracy of 81% [26]. Bioconductor R backage was used for downloading the data
III. METHODOLOGY and extracting the cases with the corresponding expressed
Through this section data preparation and pre-processing genes for pre-processing. STAR Counts from TCGA project
methodology will be discussed. Data is collected from are used for analysis.
the Genomic Data Commons (GDC) portal and different For avoiding small test set or data imbalance no anomaly
preprocessing techniques are used for preparing genomic and detection is used, but stratified kfold cross validation is
slide images for the training phase used to obtain unbiased results. Stratified kfold cross

134772 VOLUME 12, 2024

M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

validation ensures gaining the same proportion of data across performed using ‘LabelEncoder’ from scikit-learn to convert
the splits. categorical labels into numerical values. The data is split
The molecular models follows a binary classification either into separate folders for training and validation to make sure
the patient sample is Tumor or Normal while the WSI model that our model is never trained and tested on the same set
follows 3 classes classification (LUAD, LUSC, Control). The of tiles. Normalization for the pixel values is performed
data sample of each patient can be fed to these individual between 1 and 0. This normalization helps the neural network
models for accurate diagnosis. First, the sample is tested converge faster and perform better during training. No data
on RNA-Seq (mRNA) model and miRNA-Seq model to be augmentation is performed.
classified as normal or tumor. For Whole Slide Images, it goes WSI images are known for their large size that they cannot
for further classification as LUAD, LUSC or Normal for be fed directly to any neural networks. So, a magnification
deciding the appropriate treatment plan. factor 20x was used for obtaining 512 × 512 non-overlapping
pixel-tiles. Depending on the original WSI image, tens
to thousands of tiles are generated as in the literature
review [19].

TABLE 2. Number of tiles generated per class.

Table2 shows the number of tiles generated for each

FIGURE 1. The multimodal architecture for NSCLC classification.
class that are used for the training. For implementing the
classification models, the Python packages Tensorflow and
The multimodal architecture is shown as in figure1. For Scikit-Learn were used.
the whole slide images the tiles are first generated, then
feature extraction is performed using the pre-trained model 2) OMICS DATA PRE-PROCESSING
VGG16, after that the training phase is processed through As shown in Figure3, gene expression quantification STAR
a CNN model. Also for RNA-Seq DESeq2 is used for the Counts data is used from the GDC portal for both mRNA and
preprocessing and for the feature reduction and then a CNN miRNA samples. first preprocessing and feature reduction
model is used for training. For miRNA-Seq, normalization is for RNA-Seq is performed using DESeq2 and then a CNN
used for the preprocessing and no feature reduction is needed, model is used for training. For miRNA-Seq normalization is
and then a CNN model is used for the training phase. performed as the preprocessing and no feature reduction is
needed and then a CNN model is used for training. Classifi-
1) WSI PRE-PROCESSING cation is binary meaning it is either Normal or Tumor.
As shown in figure2, tiles are first generated and separated
in different folders for training, validation and testing. Then
a pre-trained model VGG16 is used as a feature reduc-
tion. After that a CNN model is used for the training and
classification.

FIGURE 3. Deep learning system architecture for the molecular samples.

Choosing the right pre-processing tools and packages was

important for accurate extraction of differentially expressed
genes. Noise and missing Values are two common challenges
FIGURE 2. Deep learning system architecture for the whole slide images.
faced when dealing with gene expression from next gener-
ation sequencing. Choosing the right subset of genes that
For the pre-processing stage of WSI, python package affects the most in tumor growth is essential in the prepro-
openslide was used. Images from directories for training and cessing phase. According to the fact that RNA-Seq does not
validation are loaded using OpenCV (‘CV2’). Labels for each follow Normal Distribution but follows Negative Binomial
image is extracted from the directory path. Label encoding is Distribution makes DESeq2 the best choice for RNA-Seq

VOLUME 12, 2024 134773

 M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

pre-processing. DESeq2 is an R package that is used for

the analysis of high-throughput RNA-Seq data to extract the
Differentially Expressed Genes (DEGs) which are usually
responsible for causing cancer.
One essential step in RNA-Seq data analysis involves
normalizing the data, with DESeq2 employing the ‘‘size
factor’’ method to account for variants in sequencing depth
FIGURE 4. Gene expression 3D reshaping for feeding into CNNs.
among samples. This normalization ensures gene expression
values are comparable across samples, facilitating accu-
rate identification of DEGs. Alongside factor normalization, Here in Table3 the number of samples for each data modal-
DESeq2 utilizes a variance-stabilizing transformation for ity are indicated.
further enhance data quality by stabilizing variance across
expression levels. This combined approach minimizes bias TABLE 3. Number of samples per cass for each data modality.
and enhances the accuracy of differentially expressed genes.
Moreover, DESeq2 offers negative binomial distribution
models to address the over-dispersion often observed in
RNA-Seq data, acknowledging variability not adequately
explained by a simple Poisson distribution. By incorporating
the negative binomial distribution, DESeq2 effectively mod-
els gene expression count dispersion, yielding more reliable
estimates of differential expression.
DESeq2 was used over 60660 genes to extract the most
important differentially expressed genes. As parameters, IV. IMPLEMENTATION AND DISCUSSION
a Log2 Fold Chain (LFC) value of 1.2 and a p-value of Through this section a detailed explanation for the training
0.05were set. 33 DEGs were extracted for the training and phase is given. Also, figures of the obtained results are shown.
classification phase.
For the case of miRNA, there was no need for fea- A. MODEL SELECTION AND TRAINING
ture reduction because TCGA provides information for only For the training phase different methodologies are used for
1881 miRNAs. Only normalization was applied using Min- the image and genetic data. For the whole slide images, pre-
MaxScaler from scikit-learn library in python. For both of trained model is first used as a transfer learning for feature
the genetic data the categorical labels were encoded using reduction and then a CNN model is used for training and
‘LabelEncoder’ to convert them into numeric format. classification. RNA-Seq and miRNA samples are trained and
Figure4 shows the gene expression data reshaping pro- classified through a CNN model.
cess for feeding into a CNN model. The gene expression
can be presented in 2D format where rows represent genes 1) WSI TRAINING
and columns represent expression profiles of patient tumor The VGG16 architecture was used for the WSIs. Different
samples. The value in cell Xij represents the gene (i) in patient pre-trained models were tried for training such as Incepri-
sample (j). Then, in order to feed the neural networks with this onV3 and Resnet-50, but VGG16 gives better results. VGG16
data, it has to be reshaped in a format suitable for the network network is a specific implementation within the VGG family
to be understood, analyzed and classified. A common 3D of convolutional neural networks, designed to enhance the
shape for convolutional neural network (number of samples, depth of the architecture while maintaining simplicity in the
sequence length, number of channels) is used for the molecu- design by using small 3 × 3 convolutional filters. These
lar data. The number of samples represents the data points or small filters ensures that the network captures intricate pat-
samples in the dataset. Each sample represents an individual terns while keeping the number of parameters manageable.
instance or observation in the dataset. The sequence length Moreover, the increased depth of the VGG16 network, with
represents the number of genes measured (length of gene 16 weight layers in total, allows it to learn complex features
sequence) in each sample. The number of channels represents and achieve high performance in image classification tasks.
the number of features or gene expression levels measured for In addition to that, the consistent use of the same filter size
each gene. So, reshaping input features array to (number of and doubling the number of filters after each max-pooling
samples, sequence length, number of channels) means orga- layer provides a systematic approach to increasing the net-
nizing your data such that each sample consists of a sequence work capacity and depth.
of gene expression values (sequence length) across different The pre-trained weights on ImageNet are used as the
genes (number of channels). This shape allows the CNN to starting point. Transfer learning is also applied through freez-
effectively learn spatial patterns in the gene expression data ing the layers of VGG16 and set them to non-trainable for
across different genes while considering the sequential nature preventing weights from being applied during training. All
of the data. 18 convolutional and pooling layers in the VGG16 model

134774 VOLUME 12, 2024

M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

FIGURE 5. WSI training and validation accuracy. FIGURE 7. WSI F1 score.

FIGURE 6. WSI confusion matrix. FIGURE 8. WSI training and validation loss curve.

were frozen during training. The training images are passed

through the VGG16 model to extract features and convert
them to feature vectors.
Stratified k-fold cross validation is used to split the training
data into 5 folds while ensuring that each fold has approxi-
mately the same proportion of samples for each class to avoid
small test set and imbalance to obtain unbiased results.
A custom neural networks model is defined and compiled
for classification. It consists of a Flatten layer to flatten the
feature maps, followed by dense layers with ReLU activa-
tion, 0.4 dropout for regularization, and a softmax output
layer. Adam optimizer was used with its default parameters.
Categorical cross entropy as a loss function and accuracy as
evaluation metric.
The network was trained during 8 epochs for each fold. The FIGURE 9. WSI RMSE for each fold.
following metrics were calculated during each fold (accuracy,
F1 score, AUC, RMSE, and confusion matrix) and the aver- features (gene expression data) and labels are loaded and the
age of each matrix was calculated at the end of the training dataset are split into 80% training and 20% testing sets.
and validation phase. A CNN model was defined using keras ‘sequential’ API
which consists of:
2) MOLECUAR SOURCES TRAINING • One dimensional convolutional layer with 64 filters,
The implementation of molecular data training was per- kernel size of 3, and ReLU activation function.
formed using Keras and scikit-learn and tensorflow. The input • Max pooling layer with pool size of 2.

VOLUME 12, 2024 134775

 M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

FIGURE 10. mRNA model accuracy. FIGURE 12. mRNA model confusion matrix.

FIGURE 11. mRNA model ROC curve.

FIGURE 13. mRNA model loss for the training and validation.

• Flattening layer to flatten the output of the output of the

convolutional layer.
• Fully connected dense layer with 128 neurons and RelU
activation function.
• Output layer (‘Dense’) with 2 neurons and softmax acti-
vation function for binary classification.
Then the model was compiled with default parameters of
Adam optimizer, sparse-categorical cross-entropy loss func-
tion and accuracy metric.
For RNA-Seq (mRNA) process runs for stratified 10 fold
cross validation and 10 epochs with batch size of 32, while
for miRNA-Seq the process runs for 5 folds and 10 epochs
with batch size of 32.

V. RESULTS FIGURE 14. miRNA model accuracy for the training and validation.
The results obtained by each modality can be observed in
Table4. As observed, the best results are achieved according
to the following order: miRNA-Seq, RNA-Seq, WSI. Binary results. The WSI assures the result if it is normal and goes
classification is performed on both of the molecular data. for further classification. Otherwise, if classification is tumor
Classification is based on being tumor or normal. Based on the samples are to be classified either for LUAD or LUSC.
the results of the binary classification, the samples go for The followed method helps in determining the right treatment
one more classification step on WSI images for more precise plan.

134776 VOLUME 12, 2024

M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

TABLE 4. Results obtained by each data modality.

previous related works obtained by Ye et al. [18]. Moreover,

WSI classification accuracy of 89.73% exceeds the ones
obtained by Graham et al. [26], and Carrillo-Perez et al. [2].
All efforts exerted in the research field are for improving
the results obtained which reflects in faster discovery and
FIGURE 15. miRNA model ROC curve.
treatment of the disease. So, through better results gained in
this research, the hope for better treatment could be possible.
Also, applying the methodology performed on other types of
cancer could be beneficial.

VII. CONCLUSION
To summarize the previous work, deep learning using convo-
lutional neural networks has already proven its great impact
on the bioinformatics field and will lead the next revolution
of detecting rare diseases at very early stages. Convolutional
neural networks will confront the growing problem of high
dimensionality especially in the omics data. Moreover, it is
already known for its accurate classification of high resolu-
tion slide images. Throughout this research we demonstrated
the great benefit of using convolutional neural networks in
the classification of non-small cell lung cancer using omics
data such as mRNA-Seq and miRNA-Seq and images like
FIGURE 16. miRNA model confusion matrix. Whole Slide Images achieving higher result than obtained by
previous works.

VIII. FUTURE WORK

Different omics modalities could be used for new insights
and new biomarkers to be discovered such as DNA methy-
lation (metDNA), copy number vartaion (CNV). Moreover,
different pre-trained models could be used to test the best
models for higher results such as InceptionV3, Resnet18,
Resnet50. Also, differet types of cancer data are available
in the GDC portal. This methodology could be applied on
other 32 types of cancer as a unified tool for classification.
In addition to that, late fusion is proved to be a great tool for
more insights and biological meanings. It may be applied for
the classification of these different types for more accurate
treatment.
FIGURE 17. miRNA model loss for the training and validation.
REFERENCES
VI. DISCUSSION [1] H. Sung, J. Ferlay, R. L. Siegel, M. Laversanne, I. Soerjomataram,
The results observed in table exceeds those gained by other A. Jemal, and F. Bray, ‘‘Global cancer statistics 2020: GLOBOCAN
estimates of incidence and mortality worldwide for 36 cancers in 185 coun-
previous studies as previously mentioned in Table 1. The tries,’’ CA, Cancer J. Clinicians, vol. 71, no. 3, pp. 209–249, May 2021,
mRNA-Seq classification accuracy is 96.79% which exceeds doi: 10.3322/caac.21660.
the results obtained by Smolander et al. [21], Fan et al. [22], [2] F. Carrillo-Perez, J. C. Morales, D. Castillo-Secilla, O. Gevaert, I. Rojas,
and L. J. Herrera, ‘‘Machine-learning-based late fusion on multi-omics and
and Castillo-Secilla et al. [24]. Also, for the miRNA-seq multi-scale data for non-small-cell lung cancer diagnosis,’’ J. Personalized
classification, F1score of 99.67% is improved comparing to Med., vol. 12, no. 4, p. 601, Apr. 2022, doi: 10.3390/jpm12040601.

VOLUME 12, 2024 134777

 M. M. Amin et al.: Multimodal Non-Small Cell Lung Cancer Classification Using CNNs

[3] T. Khorshed, M. N. Moustafa, and A. Rafea, ‘‘Deep learning [22] Z. Fan, W. Xue, L. Li, C. Zhang, J. Lu, Y. Zhai, Z. Suo, and J. Zhao,
for multi-tissue cancer classification of gene expressions ‘‘Identification of an early diagnostic biomarker of lung adenocarcinoma
(GeneXNet),’’ IEEE Access, vol. 8, pp. 90615–90629, 2020, doi: based on co-expression similarity and construction of a diagnostic model,’’
10.1109/ACCESS.2020.2992907. J. Transl. Med., vol. 16, no. 1, Jul. 2018, Art. no. 205, doi: 10.1186/s12967-
[4] Types of Lung Cancer | Cancer Research U.K. Accessed: May 23, 2024. 018-1577-5.
[Online]. Available: https://www.cancerresearchuk.org/about- [23] S. González, D. Castillo, J. M. Galvez, I. Rojas, and L. J. Herrera, ‘‘Feature
cancer/lung-cancer/stages-types-grades/types selection and assessment of lung cancer sub-types by applying predictive
[5] A. M. Bode and Z. Dong, ‘‘Cancer prevention research—Then and models,’’ in Proc. Int. Work-Conf. Artif. Neural Netw., in Lecture Notes in
now,’’ Nature Rev. Cancer, vol. 9, no. 7, pp. 508–516, Jun. 2009, doi: Computer Science: Including Subseries Lecture Notes in Artificial Intelli-
10.1038/nrc2646. gence and Lecture Notes in Bioinformatics, vol. 11507, 2019, pp. 883–894,
[6] N. Hanna, D. Johnson, S. Temin, S. Baker, J. Brahmer, P. M. Ellis, doi: 10.1007/978-3-030-20518-8_73.
G. Giaccone, P. J. Hesketh, I. Jaiyesimi, N. B. Leighl, G. J. Riely, [24] D. Castillo-Secilla, J. M. Gálvez, F. Carrillo-Perez, M. Verona-Almeida,
J. H. Schiller, B. J. Schneider, T. J. Smith, J. Tashbar, W. A. Biermann, D. Redondo-Sánchez, F. M. Ortuno, L. J. Herrera, and I. Rojas, ‘‘KnowSeq
and G. Masters, ‘‘Systemic therapy for stage IV non–small-cell lung R-Bioc package: The automatic smart gene expression tool for retrieving
cancer: American society of clinical oncology clinical practice guideline relevant biological knowledge,’’ Comput. Biol. Med., vol. 133, Jun. 2021,
update,’’ J. Clin. Oncol., vol. 35, no. 30, pp. 3484–3515, Oct. 2017, doi: Art. no. 104387, doi: 10.1016/j.compbiomed.2021.104387.
10.1200/jco.2017.74.6065. [25] F. Kanavati, G. Toyokawa, S. Momosaki, M. Rambeau, Y. Kozuma,
F. Shoji, K. Yamazaki, S. Takeo, O. Iizuka, and M. Tsuneki, ‘‘Weakly-
[7] Z. Wang, M. Gerstein, and M. Snyder, ‘‘RNA-seq: A revolutionary tool for
supervised learning for lung carcinoma classification using deep learning,’’
transcriptomics,’’ Nature Rev. Genet., vol. 10, no. 1, pp. 57–63, Jan. 2009,
Sci. Rep., vol. 10, no. 1, Jun. 2020, Art. no. 9297, doi: 10.1038/s41598-
doi: 10.1038/nrg2484.
020-66333-x.
[8] J. M. Rizzo and M. J. Buck, ‘‘Key principles and clinical appli- [26] S. Graham, M. Shaban, T. Qaiser, N. A. Koohbanani, S. A. Khurram, and
cations of ‘next-generation’ DNA sequencing,’’ Cancer Prevention N. Rajpoot, ‘‘Classification of lung cancer histology images using patch-
Res., vol. 5, no. 7, pp. 887–900, Jul. 2012, doi: 10.1158/1940-6207. level summary statistics,’’ Proc. SPIE, vol. 10581, pp. 327–334, Mar. 2018,
capr-11-0432. doi: 10.1117/12.2293855.
[9] J. M. Knight, I. Ivanov, K. Triff, R. S. Chapkin, and E. R. Dougherty, [27] J. N. Weinstein, E. A. Collisson, G. B. Mills, K. R. M. Shaw,
‘‘Detecting multivariate gene interactions in RNA-seq data using opti- B. A. Ozenberger, K. Ellrott, I. Shmulevich, C. Sander, and J. M. Stuar,
mal Bayesian classification,’’ IEEE/ACM Trans. Comput. Biol. Bioinf., ‘‘The cancer genome atlas pan-cancer analysis project,’’ Nature Genet.,
vol. 15, no. 2, pp. 484–493, Mar. 2018, doi: 10.1109/TCBB.2015. vol. 45, no. 10, pp. 1113–1120, 2013, doi: 10.1038/ng.2764.
2485223.
[10] S. Sleijfer, J. Bogaerts, and L. L. Siu, ‘‘Designing transformative clin-
ical trials in the cancer genome era,’’ J. Clin. Oncol., vol. 31, no. 15,
pp. 1834–1841, May 2013, doi: 10.1200/jco.2012.45.3639. MARIAN MAGDY AMIN received the B.Sc.
degree in computer science from The American
[11] L. E. MacConaill, ‘‘Existing and emerging technologies for tumor genomic
profiling,’’ J. Clin. Oncol., vol. 31, no. 15, pp. 1815–1824, May 2013, doi: University in Cairo, in 2014. She is a currently
10.1200/jco.2012.46.5948. pursuing the master’s degree (by research) with
[12] The Cancer Genome Atlas Program (TCGA)—NCI. Accessed: Fayoum University, Egypt.
May 23, 2024. [Online]. Available: https://www.cancer.gov/ccg/
research/genome-sequencing/tcga
[13] K. R. Kukurba and S. B. Montgomery, ‘‘RNA sequencing and analysis,’’
Cold Spring Harbor Protocols, vol. 2015, no. 11, Nov. 2015, Art. no. pdb-
top084970, doi: 10.1101/pdb.top084970.
[14] J. E. Dancey, P. L. Bedard, N. Onetto, and T. J. Hudson, ‘‘The genetic
basis for cancer treatment decisions,’’ Cell, vol. 148, no. 3, pp. 409–420, AHMED S. ISMAIL received the master’s degree
Feb. 2012, doi: 10.1016/j.cell.2012.01.014. in computer science and information systems from
[15] C. Peng, X. Wu, W. Yuan, X. Zhang, Y. Zhang, and Y. Li, the Department of Information Systems, Faculty
‘‘MGRFE: Multilayer recursive feature elimination based on an embed- of Computers and Information, Fayoum, Egypt,
ded genetic algorithm for cancer classification,’’ IEEE/ACM Trans. in 2012, and the Ph.D. degree from the Depart-
Comput. Biol. Bioinf., vol. 18, no. 2, pp. 621–632, Mar. 2021, doi: ment of Information Systems, Faculty of Com-
10.1109/TCBB.2019.2921961. puter Science and Information Systems, Fayoum
[16] Y. Lecun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521, University, in January 2021. He is an Assistant
pp. 436–444, May 2015, doi: 10.1038/nature14539. Professor. He has authored/co-authored several
[17] D. Castillo, J. M. Galvez, L. J. Herrera, F. Rojas, O. Valenzuela, O. Caba, scientific researches in various technical fields,
J. Prados, and I. Rojas, ‘‘Leukemia multiclass assessment and classifi- such as semantic web, data science, big data, the Internet of Things, and
cation from microarray and RNA-seq technologies integration at gene blockchain.
expression level,’’ PLoS ONE, vol. 14, no. 2, Feb. 2019, Art. no. e0212127,
doi: 10.1371/journal.pone.0212127.
[18] Z. Ye, B. Sun, and Z. Xiao, ‘‘Machine learning identifies 10 feature
MASOUD E. SHAHEEN received the B.Sc.
miRNAs for lung squamous cell carcinoma,’’ Gene, vol. 749, Jul. 2020,
degree in science from the Department of Math-
Art. no. 144669, doi: 10.1016/j.gene.2020.144669.
ematics and Computer Science, Minia University,
[19] N. Coudray, P. S. Ocampo, T. Sakellaropoulos, N. Narula, M. Snuderl,
in 1996, the M.S. degree in computer science from
D. Fenyö, A. L. Moreira, N. Razavian, and A. Tsirigos, ‘‘Classification
and mutation prediction from non–small cell lung cancer histopathology the Faculty of Science, Fayoum University, Egypt,
images using deep learning,’’ Nature Med., vol. 24, no. 10, pp. 1559–1567, in 2005, and the Ph.D. degree in computer science
Oct. 2018, doi: 10.1038/s41591-018-0177-5. from The University of Southern Mississippi, Hat-
[20] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image tiesburg, MS, USA, in 2013. He was the Vice-Dean
recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), of postgraduate studies and research with the Fac-
Jun. 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90. ulty of Computers and AI, Fayoum University,
[21] J. Smolander, A. Stupnikov, G. Glazko, M. Dehmer, and F. Emmert- from May 2021 to December 2022, where he is an Associate Professor with
Streib, ‘‘Comparing biological information contained in mRNA and the Computer Science Department. He is the Project Portal Manager with
non-coding RNAs for classification of lung cancer patients,’’ BMC Can- Fayoum University. He is also the Vice-Dean for Community Service and
cer, vol. 19, no. 1, Dec. 2019, Art. no. 1176, doi: 10.1186/s12885-019- Environmental Development Affairs.
6338-1.

134778 VOLUME 12, 2024