Results in Engineering 18 (2023) 101111

Contents lists available at ScienceDirect

Results in Engineering
journal homepage: www.sciencedirect.com/journal/results-in-engineering

Lung cancer disease prediction with CT scan and histopathological images

feature analysis using deep learning techniques
Vani Rajasekar a, M.P. Vaishnnave b, S. Premkumar c, Velliangiri Sarveshwaran d, *,
V. Rangaraaj a
a
Department of Computer Science and Engineering, Kongu Engineering College, Erode, India
b
Department of Computer Science and Engineering, Sathyabama Institute of Science and Technology, Chennai, India
c
Department of Computer Science and Engineering, Galgotias University, Uttar Pradesh, India
d
Department of Computational Intelligence, SRM Institute of Science and Technology, Kattankulathur Campus, Chennai, India

A R T I C L E I N F O A B S T R A C T

Keywords: Lung cancer is characterized by the uncontrollable growth of cells in the lung tissues. Early diagnosis of ma­
Lung cancer lignant cells in the lungs, which provide oxygen to the human body and excrete carbon dioxide because of
Histopathological images important processes, is critical. Because of its potential importance in patient diagnosis and treatment, the use of
CT Scan images
deep learning for the identification of lymph node involvement on histopathological slides has attracted wide­
Deep learning
Diagnosis
spread attention. The existing algorithm performs considerably less in recognition accuracy, precision, sensi­
tivity, F-Score, Specificity, etc. The proposed methodology shows enhanced performance in the metrics with six
different deep learning algorithms like Convolution Neural Network (CNN), CNN Gradient Descent (CNN GD),
VGG-16, VGG-19, Inception V3 and Resnet-50. The proposed algorithm is analyzed based on CT scan images and
histopathological images. The result analysis shows that the detection accuracy is better when histopathological
tissues are considered for analysis.

1. Introduction tumour radiographic changes with time [2] (see Figs. 1–4).
Researchers recently established predictive algorithms based on
Cancer has long been recognised as a hazardous disease that can lead differentially expressed genes to categorise lung cancer patients with
to death. Lung cancer is found to be one among frequent malignancies different health outcomes, such as relapse and total survival4. Bio­
worldwide, and it is the important reason of mortality from cancer in markers have been proven to be important in the treatment of NSCLC in
both industrialised and developing countries. The majority of these in­ previous research. Radiographic tumour characteristics may now be
dividuals have non–small cell lung cancer (NSCLC), which has a 5-year assessed quantitatively rather than qualitatively, thanks to artificial in­
mortality rate of only 18%. Despite current medical advances resulting telligence (AI), a technique known as “radiomics” [3]. Several studies
in a significant raise in overall cancer survival rates, this progress is have shown that noninvasively describing tumour characteristics has
lower significant in lung cancer because the majority of symptoms and greater predictive power than standard clinical assessments. Patholo­
recognised patients have advanced illness [1]. The cancer cells migrate gists have been using traditional microscopes and glass slides to make an
from the lungs to the lymph glands and then to the bloodstream. Due to accurate diagnosis since the mid-nineteenth century. The traditional
the obvious natural movement of lymph, cancer cells usually spread procedure has the pathologist manually reviewing several glass slides,
towards the middle of the chest. If this cancer had spread to other or­ which takes a great amount of time and energy. The introduction of slide
gans, metastasis will occur, and early identification can help avoid this. scanning equipment that can increase production digital slides has
Nonsurgical methods, including as radiation, chemotherapy, surgical ushered classical pathology into the digital world, bringing various
intervention, or monoclonal antibodies, are frequently used to treat benefits to the histopathology process [4]. The major advantage is the
late-stage lesions. This emphasises the critical importance of employing capability to utilize computer simulations, such as automated image
follow-up radiography to monitor therapeutic response and track analysis, to assist health care professionals in the investigation and

* Corresponding author.
E-mail addresses: [email protected] (V. Rajasekar), [email protected] (M.P. Vaishnnave), [email protected] (S. Premkumar),
[email protected] (V. Sarveshwaran), [email protected] (V. Rangaraaj).

https://doi.org/10.1016/j.rineng.2023.101111
Received 25 August 2022; Received in revised form 16 March 2023; Accepted 16 April 2023
Available online 18 April 2023
2590-1230/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Fig. 1. Proposed CNN architecture.

quantitative determination of slides, reducing the amount of time disease prediction with CT scan and histopathological images was
needed for manual testing and trying to improve pathologist precision, employed. The prediction analysis is made with different algorithms like
repeatability, and workflow efficiency. The use of deep learning ap­ CNN with quasi-convex gradient descent, VGG16, InceptionV3, VGG 19
proaches to aid diagnostics has recently sparked a great deal of interest and Resnet50. A novel strategy for CNN optimization with quasi-
in histopathology. gradient descent is employed which results in higher accuracy of lung
Convolutional neural networks (CNNs) in general have shown cancer detection and reduced false positives. The main contributions of
tremendous promise in medical disease identification. CNNs have been the proposed approaches are listed as follows:
employed in pathology with the promising outcomes from tumour cell
identification in metastatic breast cancer to grade glioma and colon • Designing a deep learning methodology for lung cancer detection
cancer. When it comes to classifying cancer and making the necessary based on CT scan images and Histopathological images
therapeutic options, a thorough examination of lymph glands is essen­ • Models are designed with CNN, CNN GD, Inception V3, Resner-50,
tial. Multiple lymph node levels are considered when determining VGG-16, and VGG-19
prognosis and appropriate staging necessitates meticulous examination • Analyze the performance with various metrics like recognition ac­
of lymph node health [5]. Histopathological pictures have recently been curacy, precision, Specificity, Sensitivity, F-Score.
discovered to be an accurate predictor of various therapeutic bio­
markers. Manually screening several slides, on the other hand, can be 2. Related works
tiresome and time-consuming for pathologists, and people are capable of
making mistakes since they must keep track of which sections they The development of malignant cells in the lungs is known as lung
already studied. Several kind of available solutions in these competitions cancer. Overall men and women’s mortality rates have increased as a
were capable of demonstrating greater efficiency in terms of accuracy result of growing cancer incidence. Lung cancer is a disease wherein the
micro metastases than a pathologist operating in a busy practise. In lung cells in the lungs quickly multiply. Lung cancer cannot be eradicated,
cancer, primary tumour metastases have an identical role in predicting but it can be reduced [7]. The number of people affected with lung
cancer stage, therapy options, and outcomes as they do in breast cancer cancer is precisely equal to the number of people who smoke continu­
[6]. ously. Lung cancer treatment was evaluated using classification ap­
In this proposed methodology, a deep learning-based lung cancer proaches such as Naive Bayes, SVM, Decision Tree, and Logistic

Fig. 2. Different layers in CNN.

• Input Layer:

2
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

ranked, and a subset of meaningful features was chosen based on Gini

coefficient scores for histopathological class relationships [12]. In the
estimation of survival, a deep learning network with a tumour cell and
metastatic staging system was used to examine the dependability of
individual therapy suggestions supplied by the deep learning preserva­
tion neural network. The C statistics were employed to evaluate the
performance of the model. The computational intelligence survival
neural network model’s longevity forecasts and treatment strategies
Fig. 3. Pooling layer.
were made easier with the use of a customer interface [13].
A lung cancer detection model that utilizes image analysis and ma­
chine intelligence to identify the occurrence of lung cancer in CT scans
Regression. Pradhan et al. [8] conduct an empirical evaluation of mul­
and blood tests has been developed. Despite the fact that CT scan find­
tiple machine learning methods that can be used to identify lung cancer
ings are more efficient than mammograms, patient CT scan pictures are
using IoT devices. A survey of roughly 65 papers employing machine
divided into normal and abnormal categories [14,15]. Even in the same
learning techniques to forecast various diseases was conducted in this
tumour stage, non-small-cell cancer patients have a wide range of clin­
study. The study focuses on a variety of machine learning methods for
ical performance and results. This research investigates deep learning
detecting a variety of diseases in order to identify a gap in prospective
applications such as medical imaging, which could help with patient
lung cancer detection in medical IoT. Deep residual learning is used by
stratification by automating the measurement of radiographic proper­
Bhatia et al. [9] to identify lung cancer from CT scans. With the UNet
ties. Ardila et al. [16] suggest a novel system for predicting lung cancer
and ResNet algorithms, we propose a series of pre-processing ap­
risk based on a patient’s historical and current computerized tomogra­
proaches for emphasising cancer-prone lung regions and retrieving
phy dimensions. Our model obtains a 94.4% state-of-the-art efficiency.
characteristics. The extracted features are fed through several classifiers,
When prior computed tomography imagery was unavailable, their
namely Adaboost and Random Forest, and the individual predictions are
model excelled all six radiologists, reducing false positives by 11% and
ensembled to calculate the likelihood of a CT scan becoming cancerous.
false negatives by 5%. For CT scans of the lungs, Lakshmanaprabhu et al.
Without learning inadequate human information, Shin et al. [10,11]
[17] introduce a new automatic diagnosis categorization system. The CT
use deep learning to investigate the characteristics of cell exosomes and
scan of lung pictures was processed using Optimal Deep Neural Network
determine the similarities in human plasma extracellular vesicles. The
(ODNN) and Linear Discriminate Analysis in their research (LDA). The
deep learning classifier was tested with exosome SERS data from regular
remainder of the paper is organized as section 3 describes the deep
and lung cancer cell lines and was able to categorise them with 95%
learning techniques used in the proposed methodology. Section 4 de­
efficiency. The deep learning algorithm projected that 90.7% of pa­
scribes the result comparison of various proposed techniques. Section 5
tients’ plasma exosomes were more similar to lung cancer cell extra­
describes the conclusion and future extension of the proposed work.
cellular vesicles than the mean of healthy controls in 43 patients,
encompassing stage I and II cancer patients. In the ability to forecast
lung ADC subtypes, researchers looked at four clinical factors: age, sex, 3. Proposed methodologies
tumour size, and smoking status, as well as 40 radiomic markers. The
LIFEx software was used to extract radiomic characteristics from PET 3.1. CNN
scans of segmented cancers. The clinical and radio mic variables were
Multilayer perceptrons are regularised variations of CNNs. Multi­
layer perceptrons are a dynamic network that is linked together where
each neuron is interconnected to all neurons in the next layer. Because of
their total connection, these systems are susceptible to an information
explosion. CNNs need very little pre-processing compared to certain
other image categorization algorithms. This means that the system uses
automated retraining to enhance the filters whereas traditional algo­
rithms require hand-engineering. The lack of prior information and
human interaction in extracting features is a significant benefit.
Normalization or preventing overfitting can be accomplished in a vari­
ety of methods, including compensating parameters throughout training
or reducing interconnectivity. CNNs consider a new method to
normalization: they reap the benefits of another patient’s patterns in
Fig. 4. VGG-16 architecture.
data and generate models of increasing the complexity using smaller and
• The first two layers are 3 * 3 filter convolution layers with a total of 64 filters. simpler patterns imprinted in their filters. As a result, CNNs are at the
Because the same convolution is utilised, the size will be 224 * 224 * 64. The bottom end of the connectedness and complexity range.
incremental is always 1 and the filter has always been 3 * 3.
• Subsequently, the volume height and breadth were reduced from 224 * 224 *
3.1.1. Layers in CNN
64 to 112 * 112 * 64 using a pooling layer with such a maxpool size of 2 * 2
The various five layers in CNN are a) Input layer b) Convo layer c)
and an incremental of 2.
• Two convolution layers containing 128 filters are then applied. This will Pooling layer d) Fully connected layer e) Softmax/logistic layer and f)
result in the creation of a different dimension of 112 * 112 * 128. Output layer.
• The volume decreases to 56 * 56 * 128 after utilising the pooling layer. This layer in CNN should have to include the imaging information. A
• Two 256-filter fully connected layers were added, accompanied by a down- three-dimensional matrix is used to represent image data. This layer
sampling layer that reduced the size to 28 * 28 * 256. requires reshaping into a single column. Before feeding an image over
• The max pool layer separates two further stacks, each one with three the dimensions of 28 × 28 = 784, it will be converted to 784 × 1.
convolution layers.
• The 7 * 7 * 512 level is reduced to a fully connected (FC) layer with 4096 • Convo layer (Convo þ ReLU)
inputs and classes 1000 softmax output after the last max pooling.
This layer is also called the features extraction layer since it retrieves
components of the image. A piece of the information is originally

3
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

attached to the Convo layer to perform convolution and filtration for locating a function’s local minimum or maximum. To reduce a cost or
[18–20]. The result size is represented by a single integer returned by loss function, these techniques are frequently used in machine learning
the procedure. Then slide the filter in each of the next regions of interest (ML) and deep learning (DL). The two specific requirements of gradient
with the same input images using a Stride and continue the operation. descent are.
The same technique is repeated till the whole process is controlled. The
result will be used as the feed for the following layer. The Convo layer • Differentiable
also has ReLU activation, which lowers all negative numbers to zero. • Convex
To reduce the temporal volume of the source images, max pooling is
used after convolution. It is utilised between two convolution layers. Spatially separated functions have a derivative at each point in their
Applying FC after the Convo layer without employing pooling or domain, although not all functions meet those standards. Convex implies
maximum pooling will be operationally demanding. As a result, using that the line segment joins different function points lies on or above the
maximal pooling seems to be the only technique to minimise the input curve of the univariate variable. For two points x₁, x₂ laying on the
image’s geographic volume. In the above instance, max pooling was function’s curve this condition is expressed in equation 4.
applied to a single spatial dimension with a Span of 2.
f (λx1 + (1 − λ)x2) ≤ λf (x1) + (1 − λ)f (x2) (4)
The four-by-four-dimensional input is reduced to two-by-two di­
mensions. The pooling layer has no characteristics, but it does have two where λ signifies a point’s position on a section line and its value should
hyperparameters, Filter(F) and Stride (S). be between 0 which is left most point and 1 which is right most point.
In general, if the input dimension is W1 x H1 x D1, then Similarly, 0.5 denotes an intermediate location. Compute the second
W2 = (W1 − F) / S + 1 (1) derivative to see if its value has always been greater than zero to see if a
univariate function is convex.
H2 = (H1 − F) / S + 1 (2)
d2 f (x)
(5)
D2 = D1 (3) dx2

With a gradient descent approach, quasi-convex functions can also be

where W2, H2 and D2 are the width, height and depth of output.
used. However, they frequently feature “saddle points,” also known as
“minimax points,” where the method might become stuck.
• Fully Connected Layer (FC)

The completely linked layer includes weights, biases, and neurons. It 3.2. VGG16
links neurons in one layer to those in another. It is used to teach in­
dividuals how to categorise images into different groups. VGG16 is a CNN architecture that took first place in the 2014 ILSVR
ImageNet contest. One of the greatest Vision model architectures re­
• Softmax Layer (FC) mains. VGG16 is distinct in that, rather than many hyperparameters,
authors centred on the 3 × 3 filter’s convolution layer in step 1 and
The Softmax or Stochastic layer is the final layer of CNN. It was at the always used the same padding and max pool layer in step 2. All through
very bottom of the Convolution layers. Softmax is suitable for multi- the architecture, the sequence of convolutional levels and maximal pool
classification, whereas stochastic is suitable for binary classification. tiers is maintained [21]. The VGG architecture serves as the foundation
for cutting-edge object classification models. The VGGNet, which was
• Output Layer (FC) created as a deep neural network, outperforms baselines on a variety of
tasks and datasets other than ImageNet. Furthermore, it remains one of
The output label in the form of one hot encoding is available in the the best most widely used image recognition architectures today.
output layer.
3.3. Inception V3
3.1.2. CNN with quai gradient descent
Gradient descent (GD) is a recurrent first-order optimization process
Convolutional Neural Networks are used in the Inception V3 learning

Fig. 5. VGG-19 architecture.

4
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Fig. 6. CT scan images of Lung dataset.

5
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Fig. 7. CT scan images of Lung dataset.

algorithm for picture categorization. The Inception V3 is a much more Convolution layers. Consider the basic module of the conceptualization
enhanced form of the basic model Inception V1, which was first V1 module [22–24]. It has 55 convolutional layers, which are compu­
launched as Google Net in 2014. When several deep levels of convolu­ tationally very expensive, as previously stated. To reduce the compu­
tions were used in a model, overfitting of the data happened. To get tational cost, the 55% convolutional layer was replaced by 33%
around this, the V1 model uses the concept of several filters of different convolutional layers. Max pooling and average pooling were typically
diameters on the same level. Therefore, instead of having in-depth layers lower than the grid size of feature maps. In order to efficiently reduce the
in our inception models, we now have parallel layers, which makes the grid size, the activation dimensions of the network filters are increased
system larger instead of deeper. One of the great benefits of the Incep­ in the inception V3 design.
tion V1 model was the considerable dimension reduction. To improve
the model, the larger Convolutions were factored into smaller

6
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Fig. 8. Recognition accuracy comparison on a) CNN b) CNN GD c) Resnet-50 d) VGG-16 e) Inception V3 f) VGG-19.

3.4. VGG-19 calculated for the whole training set. Kernels are used with a size of (3 *
3) and a stride size of 1 pixel to span the entire image. To maintain the
The VGG19 model is slight variation of the VGG model with 19 layers resolution of the image, spatial padding was applied. With sride 2,
(16 convolution layers, 3 Fully connected layers, 5 MaxPool layers and 1 maximum pooling was achieved over a 2 * 2 pixel window. This has
SoftMax layer). There are 19.6 billion FLOPs in VGG19. This structure is been accompanied by the Rectified linear unit (ReLu) to bring non-
assigned with a fixed size (224 * 224) RGB picture as input, indicating linearity into the method to enhance classification and reduce process­
that the matrix was of size (224, 224, 3). The required pre-processing ing time, while earlier models employed tanh or sigmoid components.
was to subtract the average RGB value from each pixel, which was Three fully connected layers were constructed, the first two layers have

7
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Table 1 TN
Classification results of various models. Specificity = (9)
TN + FP
Parameters CNN CNN VGG Inception VGG Resnet
GD 16 V3 19 50 FP
FPR = (10)
Accuracy 93.64 97.86 96.52 93.54 92.17 93.47 TN + FP
(%)
Precision (%) 89.76 96.39 92.14 90.57 91.46 93.31 ∫1
TP FP
Sensitivity 91.70 96.79 93.71 91.73 93.20 91.14 AUC = d (11)
(%) TP + FN FP + TN
0
Specificity 88.64 97.40 92.91 92.29 92.00 92.73
(%)
2 X TP
FPR (%) 91.42 98.63 93.48 93.75 93.39 95.01 F − Score = (12)
F-Score (%) 92.58 97.96 94.24 95.68 94.78 91.63 (2 X TP) + FP + FN
In addition to the CT scan images, the histopathological images are
been of size 4096, followed by a layer with 1000 channels for 1000-way also considered for the identification of lung cancer and are shown in
ILSVRC categorization, and finally a softmax function. Fig. 7. The training and testing dataset are split from the main dataset in
the ratio if 3:1 respectively. The various deep learning models are
3.5. Resnet-50 executed against the two datasets and the values of the result are shown
in Table 1 (see Fig. 8).
One of the most major benefits of the ResNet is that it avoids bad The result analysis of recognition accuracy is denoted with six
consequences while growing network depth. The fundamental goal of different models CNN, CNN GD, Resnet-50, VGG-16, Inception V3 and
the recurrent neural network is to build a deeper neural network. It is VGG-19. The CNN GD shows a higher accuracy of 97.86%. The higher
essential for keeping precision and margin of error as one proceeds precision is obtained in the CNN GD model with 96.39% whereas the
deeper into the execution of a large number of hidden layers. This is lower precision is 89.76% in CNN. The higher sensitivity rate is 96.79%
where identification mappings can assist. Continue to learn the residuals whereas the lower sensitivity rate is 91.14% in Resnet-50. The speci­
to align the anticipated with the actual. Deep learning researchers desire ficity is obtained maximum in CNN GD with 97.40% whereas the lower
to add so many layers to extract significant information from complex rate is obtained at 88.64%. The higher False Positive Rate is obtained
images. As a result, the initial layers may detect edges, while the latter maximum at 98.63% in CNN GD and the lower FPR is 91.42% in CNN.
layers may identify recognised objects, such as automobile tyres [25]. Similarly, the CNN-GD shows the maximum F-Score of 97.96% where as
However, if the network has more than 30 layers, the efficiency falls and 91.63% is obtained for Resnet-50 as the lowest F-Score.
the accuracy drops. This contradicts the belief that adding layers to a The methods used for the detection of lung cancer in the proposed
neural network will improve its performance. This is not due to approach are CNN, CNN GD, Inception V3, Resnet-50, VGG-16 and VGG-
over-fitting, because in that instance, dropout and regularisation tech­ 19. CNN are effective models for image classification and recognition.
niques can be used to eliminate the problem entirely. It’s mostly there CNN GD is the basic algorithm for CNN model optimization. The ad­
because of the well-known vanishing gradient issue (see Fig. 5). vantages of using CNN GD are the training data helps these models learn
over time, and the cost function within gradient descent specifically acts
4. Results and discussions as a barometer, gauging its accuracy with each iteration of parameter
updates. Resnet uses a larger number of layers to train the model and
The Cancer Imaging Archive provided the dataset for this investi­ that can be trained easily without increasing the training error per­
gation, which included normal and tumorous CT images of 15,500 pa­ centage. VGG 16 and VGG 19 have faster training speeds, fewer training
tients. The size of the dataset for training includes the real-time data samples per time, and higher accuracy compared with existing methods.
collected from 180 distinct patients. In terms of scanning modes and The inception family makes several improvements including using Label
manufacturers, the dataset is likewise quite diverse. The photos were smoothing, factorized 7 × 7 convolutions, and the use of an auxiliary
most likely taken as part of regular treatment rather than as part of a classifier to propagate label information lower down the network. Thus,
supervised literature review or clinical trial. Experienced professionals the usage of all the models in the proposed architecture provides
analyzed and labelled each item in the dataset. The dataset is freely enhanced accuracy with a lower recognition rate.
available. The various images of the lung and its disease categories are
shown in Fig. 6. From the total dataset size 93,000 images are used for 5. Conclusion
training the model and 1080 images are used for testing the model. A
confusion matrix was built to examine the tumour identification out­ In the proposed approach six deep learning models were built for
comes at the slide stage in detail, providing reliability in terms of per­ efficient diagnosis of lung cancer. The methods used for the detection of
centage for tumour prediction in various-sized metastases. The lung cancer are CNN, CNN GD, Inception V3, Resnet-50, VGG-16 and
confusion matrix is computed with four values TP (True Positive), TN VGG-19. The experimental analyses were done based on the CT scan
(True Negative), FP (False Positive) and FN (False Negative). images and Histopathological images. The performance of the proposed
The performance of the proposed approach is analyzed in terms of approach was done based on the recognition accuracy, F-Score, Preci­
accuracy, precision, sensitivity, specificity, False Positive Rate (FPR) sion, Sensitivity, Specificity etc. Because of the inherent advantage of
and F-Score. The values of these parameters are computed as follows. the proposed methodology, this will be an efficient method for lung
The metrics are computed as follows. cancer detection and will be beneficial to needed people. The enhanced
performance is seen in the CNN GD model with an accuracy of 97.86%.
Accuracy =
TP + TN
(6) The higher precision is 96.39% and the sensitivity rate is 96.79%. The
TP + TN + FP + FN specificity is obtained maximum in CNN GD with 97.40%. Similarly, the
CNN-GD shows the maximum F-Score of 97.96%. In future, the proposed
TP
Pr ecision = (7) methodology can be enhanced by including fuzzy genetic optimization
TP + FP
techniques with the deep learning approaches.
TP
Sensitivity = (8)
TP + FN

8
V. Rajasekar et al. Results in Engineering 18 (2023) 101111

Credit author statement [9] S. Bhatia, Y. Sinha, L. Goel, Lung cancer detection: a deep learning approach, in:
Soft Computing for Problem Solving, Springer, Singapore, 2019, pp. 699–705.
[10] H. Shin, S. Oh, S. Hong, M. Kang, D. Kang, Y.G. Ji, Y. Choi, Early-stage lung cancer
Vani Rajasekar: Conceptualization, Methodology, Software. diagnosis by deep learning-based spectroscopic analysis of circulating exosomes,
Vaishnnave M: Data curation, Writing- Original draft preparation. ACS Nano 14 (5) (2020) 5435–5444.
Prem Kumar S: Visualization, Investigation. Rangaraaj V: Software, [11] V. Rajasekar, B. Predić, M. Saracevic, M. Elhoseny, D. Karabasevic, D. Stanujkic,
P. Jayapaul, Enhanced multimodal biometric recognition approach for smart cities
Validation.: Velliangiri Sarveshwaran: Writing- Reviewing and based on an optimized fuzzy genetic algorithm, Sci. Rep. 12 (1) (2022) 1–11.
Editing. [12] S.H. Hyun, M.S. Ahn, Y.W. Koh, S.J. Lee, A machine-learning approach using PET-
based radiomics to predict the histological subtypes of lung cancer, Clin. Nucl.
Med. 44 (12) (2019) 956–960.
Declaration of competing interest [13] Y. She, Z. Jin, J. Wu, J. Deng, L. Zhang, H. Su, C. Chen, Development and
validation of a deep learning model for non–small cell lung cancer survival, JAMA
The authors declare the following financial interests/personal re­ Netw. Open 3 (6) (2020) e205842, e205842.
[14] W. Rahane, H. Dalvi, Y. Magar, A. Kalane, S. Jondhale, Lung cancer detection using
lationships which may be considered as potential competing interests: image processing and machine learning healthcare, 2018, March, in: International
Velliangiri Sarveshwaran reports administrative support was provided Conference on Current Trends towards Converging Technologies (ICCTCT), IEEE,
by SRM Institute of Science and Technology. Velliangiri Sarveshwaran 2018, pp. 1–5.
[15] V. Rajasekar, S. Krishnamoorthi, M. Saračević, D. Pepic, M. Zajmovic, H. Zogic,
reports a relationship with SRM Institute of Science and Technology that Ensemble machine learning methods to predict the balancing of ayurvedic
includes: employment. Velliangiri Sarveshwaran has patent pending to constituents in the human body : ensemble machine learning methods to predict,
Assignee. nil. Comput. Sci. 23 (1) (2022), https://doi.org/10.7494/csci.2022.23.1.4315.
[16] A. Hosny, C. Parmar, T.P. Coroller, P. Grossmann, R. Zeleznik, A. Kumar, H.
J. Aerts, Deep learning for lung cancer prognostication: a retrospective multi-
Data availability cohort radiomics study, PLoS Med. 15 (11) (2018), e1002711.
[17] D. Ardila, A.P. Kiraly, S. Bharadwaj, B. Choi, J.J. Reicher, L. Peng, S. Shetty, End-
Data will be made available on request. to-end lung cancer screening with three-dimensional deep learning on low-dose
chest computed tomography, Nat. Med. 25 (6) (2019) 954–961.
[18] S.K. Lakshmanaprabu, S.N. Mohanty, K. Shankar, N. Arunkumar, G. Ramirez,
References Optimal deep learning model for classification of lung cancer on CT images, Future
Generat. Comput. Syst. 92 (2019) 374–382.
[1] Y. Xu, A. Hosny, R. Zeleznik, C. Parmar, T. Coroller, I. Franco, H.J. Aerts, Deep [19] H. Chao, H. Shan, F. Homayounieh, R. Singh, R.D. Khera, H. Guo, P. Yan, Deep
learning predicts lung cancer treatment response from serial medical imaging, Clin. learning predicts cardiovascular disease risks from lung cancer screening low dose
Cancer Res. 25 (11) (2019) 3266–3275. computed tomography, Nat. Commun. 12 (1) (2021) 1–10.
[2] M.I. Faisal, S. Bashir, Z.S. Khan, F.H. Khan, An evaluation of machine learning [20] S. Krishnamoorthi, P. Jayapaul, V. Rajasekar, R.K. Dhanaraj, C. Iwendi, A futuristic
classifiers and ensembles for early stage prediction of lung cancer, December, in: approach to generate random bit sequence using dynamic perturbed chaotic
2018 3rd International Conference on Emerging Trends in Engineering, Sciences system, Turk. J. Electr. Eng. Comput. Sci. 30 (1) (2022) 35–49.
and Technology (ICEEST), IEEE, 2018, pp. 1–4. [21] B. Ricciuti, G. Jones, M. Severgnini, J.V. Alessi, G. Recondo, M. Lawrence,
[3] N. Coudray, P.S. Ocampo, T. Sakellaropoulos, N. Narula, M. Snuderl, D. Fenyö, M. Awad, Early plasma circulating tumor DNA (ctDNA) changes predict response
A. Tsirigos, Classification and mutation prediction from non–small cell lung cancer to first-line pembrolizumab-based therapy in non-small cell lung cancer (NSCLC),
histopathology images using deep learning, Nat. Med. 24 (10) (2018) 1559–1567. J. Immunother. Cancer 9 (3) (2021).
[4] D.M. Ibrahim, N.M. Elshennawy, A.M. Sarhan, Deep-chest: multi-classification [22] Y. Chen, E. Zitello, R. Guo, Y. Deng, The function of LncRNAs and their role in the
deep learning model for diagnosing COVID-19, pneumonia, and lung cancer chest prediction, diagnosis, and prognosis of lung cancer, Clin. Transl. Med. 11 (4)
diseases, Comput. Biol. Med. 132 (2021), 104348. (2021) e367.
[5] M. Avanzo, J. Stancanello, G. Pirrone, G. Sartor, Radiomics and deep learning in [23] J. Hanaoka, M. Yoden, K. Hayashi, T. Shiratori, K. Okamoto, R. Kaku, A. Sonoda,
lung cancer, Strahlenther. Onkol. 196 (10) (2020) 879–887. Dynamic perfusion digital radiography for predicting pulmonary function after
[6] S.H. Hyun, M.S. Ahn, Y.W. Koh, S.J. Lee, A machine-learning approach using PET- lung cancer resection, World J. Surg. Oncol. 19 (1) (2021) 1–10.
based radiomics to predict the histological subtypes of lung cancer, Clin. Nucl. [24] V. Rajasekar, J. Premalatha, K. Sathya, Cancelable Iris template for secure
Med. 44 (12) (2019) 956–960. authentication based on random projection and double random phase encoding,
[7] P.R. Radhika, R.A. Nair, G. Veena, A comparative study of lung cancer detection Peer-to-Peer Network. Appl. 14 (2) (2021) 747–762.
using machine learning algorithms, 2019, February, in: IEEE International [25] M. Gaga, J. Chorostowska-Wynimko, I. Horváth, M.C. Tammemagi, D. Shitrit, V.
Conference on Electrical, Computer and Communication Technologies (ICECCT), H. Eisenberg, Q. Zhou, Validation of Lung EpiCheck, a novel methylation-based
IEEE, 2019, pp. 1–4. blood assay, for the detection of lung cancer in European and Chinese high-risk
[8] K. Pradhan, P. Chawla, Medical Internet of things using machine learning individuals, Eur. Respir. J. 57 (1) (2021).
algorithms for lung cancer detection, J. Manag. Anal. 7 (4) (2020) 591–623.