NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.

NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

LUNG CANCER DETECTION BY USING CNN

ARCHITECTURE MODELS
1
Dr. Dattatray G. Takale, 2 Dr. Parishit N. Mahalle, 3Dr. Sachin R. Sakhare, 4Prof. Piyush P.
Gawali, 5Prof. Gopal Deshmukh, 6Dr. Vajid Khan, 7Dr. Chitrakant B. Banchhor
1
Assistant Professor, Department of Computer Engineering, Vishwakarma Institute of information Technology,
SPPU Pune
2
Professor and Head, Department of AI & DS, Vishwakarma Institute of information Technology, SPPU Pune
3
Professor and Head, Department of Computer Engineering, Vishwakarma Institute of information Technology,
SPPU Pune
4
Assistant Professor, Department of Computer Engineering, Vishwakarma Institute of information Technology,
SPPU Pune
5
Assistant Professor, Department of Computer Engineering, Vishwakarma Institute of information Technology,
SPPU Pune
6
Associate Professor, Department of Computer Engineering, KJ College of Engineering and management
Research, SPPU Pune
7
Assistant Professor, Department of Computer Engineering, Vishwakarma Institute of information Technology,
SPPU Pune
1
[email protected], [email protected], [email protected], [email protected],
5
[email protected], [email protected], [email protected]

Abstract:
Cancer disorders are caused when certain body cells enlarge uncontrollably and spread throughout the body.
Lung cancerous lesion tumors can be detected by radiologists with the help of medical imaging techniques like
Computed Tomography (CT) scan images, by using chest X-rays, or by using MRI. We examine various
medical image-based datasets, their availability, and the number of cases they contain. Due to the abundance of
pictures, a CT scan-based strategy is employed. The vastness of the collection made it difficult for radiologists
to diagnose quickly and correctly. Among the data that was supplied, there is a great deal of computer
tomography (CT) scan pictures. Chest radiographs are the source of diagnostic mistakes in around ninety
percent of instances involving lung cancer. In chest radiographs, it might be difficult for radiologists to
differentiate a lung lesion from bones, pulmonary arteries, mediastinal structures, and other complicated
anatomical features. This can be a challenge for diagnostic purposes. In order to tackle this issue, distinct
predictive analytics and intelligent retrieval techniques are employed, like traditional Computer Aided Diagnosis
(CAD), Convolutional Neural Network (CNN), Dens-Net, Agile CNN, and LeNet. These deep structured
learning-based methods mostly used complex neural networks, which caused vanishing gradient problems. To
solve such a problem, we are using the ResNet50 method approach to get better result accuracy. CT scan images
from (LIDC/IDRI) are being used for the dataset
Keywords: AlexNet, GoogleNet, ResNet50, Convolutional neural network (CNN), lung cancer, transfer
learning, Lung Imaging Database Consortium (LIDC)

DOINumber:10.14704/nq.2022.20.10.NQ551250 NeuroQuantology2022;20(10):11-19

I. INTRODUCTION encourage the growth of tumors. The vast

Cancer is a disease that may occur in a specific majority of lung cancer cells originate in these
area of the body when certain tissues have a tubes. Lung cancers that are malignant are
tendency to proliferate excessively and normally classified into the fundamental

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

categories of adenocarcinoma and squamous some distance from the lungs. There are many
cell carcinoma. Shouldering nicotine is different kinds of image scans, such as X-ray
connected with around 75%-90% of all cases pictures, images based on CT (ccomputer
of fatal lung cancer in the United States. Tomography) scans, and images based on MRI
According to the Global Cancer Statistics (Magnetic Resonance Imaging) scans. The
2020, globally, there have been around 19.4 next updated version of X-Ray pictures is
million people who have recently succumbed called CT, which stands for computer
to cancer. Of them, 11.4% are lung malignant tomography. Traditional thoracic x-rays are
tumors. There will be over 10.5 million deaths not as effective in detecting lung malignancies
caused by cancer in the year 2020, and there as CT (computer Tomography) scans[5]. The
will be 19.4 million new cases of cancer CT scans offer a greater probability of
diagnosed. According to the data, there were detecting lung tumors. In addition to this, it
2,206,711 newly identified cases in the year may be helpful in detecting any enlarged
2020. Lethargy, a lump, persistent, lymph nodes that may contain cancerous cells
inexplicable muscle or joint pain, inexplicable that have migrated to other parts of the body,
pyrexia or night sweats, constant tussis, in addition to the size, form, and location of
sputum flecked with blood [1], pectoralgia, these lymph nodes. Magnetic resonance
shortness of breath, and a few other symptoms imaging, sometimes known as an MRI scan, is
may be considered a syndrome in the early a diagnostic procedure that produces very
stages of lymphoma. However, these detailed images of the body's soft tissues. The
symptoms can also be signs of cancer in later CAD (Computer-Aided Diagnosis Systems)
stages of the disease. While even non-smokers approach is one of the more time-honoured
are capable of developing malignant lung methods for identifying potentially cancerous
lesions, the risk of developing them is higher lesions in the lung. The following are the four
for smokers [2, 3]. primary processing phases that are included in
There are two types of lung cancerous lesions: a modern CAD system approach: Image Pre-
Lung lesions caused by cancerous oat cells processing [7,8], Image Clustering,
Chain smokers have a much higher risk of Characteristic Extraction, and the diagnosis of
developing oat cell lung malignant lesion than the nodules as either benign or malignant are
non-small cell lung cancerous lesion [3]. the steps that need to be taken. Other methods,
A non-small cell lung malignant tumor. such as Predictive Analytics and hierarchical
Cancer and carcinomas of the epidermis are learning techniques, were also utilized. “These
both kinds of non-small cell lung cancer methods included (CNN)[1, 2D, and 3D
(NSCL). In this stage, the cancerous lesions (CNN)[2, respectively. Dense-Net, Agile
have not spread beyond the original location of (CNN), SVM (Support Vector Machine)
the lung, but they have been found there. In classifier, and XGBoost are all examples of
the second stage, a malignant disease may neural network architectures. X-rays, magnetic
develop in the lungs as well as the lymph resonance imaging (MRI) scans, computed
nodes that are located nearby. In the third tomography (CT) scans, chest radiographs,
phase, a lung tumor and lymph node growths and positron emission tomography (PET)
appear in the middle of the thoracic cavity[4]. scans are all examples of datasets that are
As is the case with malignant lesions of any available to the public”. In this survey, a
kind, cancerous lymph nodes are often only variety of deep learning approaches to the
present on the same side of the thoracic cavity diagnosis of lung cancer are investigated. This
as the primary tumor, but only on that side. assists in the improvement of technique for
Infection has spread to the lungs, the tissues enhanced performance.
around the lungs, or organs that are located

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

The findings of the global cancer Satanists degrees of accuracy possessed by a number of
2021 study show how harmful and widespread machine learning algorithms [13].
the lung cancer illness is. A large part of the
reason why lung cancer has such a high II. LITERATURE SURVEY
mortality rate is because it is often not Predictive analysis and hierarchical learning-
diagnosed in its early stages. It led to the based approaches are applied by many
patient's treatment being delayed, which researchers to improve the accuracy of results.
increased the likelihood of the patient passing Medical images like CT scans, MRIs, and X-
away. The radiologist, with the use of X-rays, Ray images are used in many more formats.
“CT (Computed Tomography)”, and MRI As a result of applying this codification or
scans, detects lung cancer (Magnetic algorithms to various datasets, lung cancerous
resonance imaging). The diagnostic procedure lesions were detected with high accuracy and
takes a significant amount of time. with the effectiveness
purpose of assisting persons working in the Yang et al. [3] throughout, the LIDC data
medical area by speeding up the diagnostic or collection will be used as a point of reference.
analytical process. The use of machine The photos were investigated using
learning methods was offered here for the convolution neural networks to get a better
diagnosis of lung cancer. The “CAD comprehension of the data after the original
(Computed Assist Diagnosis), CNN images were divided into smaller patches
(Convolutional Neural Network), 2D CNN, based on the location of the centroids of the
and 3D CNN [3,5] are all examples of neural cancerous nodules. The researchers were able
network-based algorithms” that suffer from the to get a better comprehension of the data as a
vanishing gradient issue, which may result in result of this.
the loss of data. As a consequence, ResNet50 Fan et al. presented in their study [4] a method
implements the skip connection, which for locating nodules in chest CT scans that
ultimately leads to improved result accuracy makes use of 3D convolutional neural
[15]. networks in conjunction with other, more
Detection of lung cancer using a CT scan that conventional image processing techniques.
is based on an image data collection. In This was done so that the nodules could be
machine learning, neural networks proved to identified. The purpose of doing this was to
be the most effective method for image make their system more accurate, and it was a
processing. Due of the heavy traffic on this successful endeavour. Beginning with a
network, a vanishing gradient issue has arisen. version in black and white, the artwork
It is beneficial to use a model that is based on progresses through many iterations before
ResNet50 in order to avoid losing data. The arriving at its completed, colourful state.
skip connection mechanism that ResNet50 During the journey, it goes through a process
introduces helps to avoid features from being of metamorphosis (RGB). Following then, a
used. We hope to get better results by significant portion of the morphological
increasing the accuracy of the results with the operations will have reached their conclusion.
assistance of the ResNet50 and ResNet152 Last but not least, the area of the picture that
models [7]. was connected will act as a mask for the CT
The goal of this study is to establish whether image. According to the researchers, they
or not MR imaging of lung cancer may be obtained their data from CNN, and the
categorized as non-specific, multifocal, findings showed that the network had an
multicentric, or gliomatosis. An investigation accuracy rate of 67.7%.
was carried out to determine the relative Victor et al. [5], the accuracy of 88.41% was
achieved by the application of deep transfer

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

learning. Jan et al. proposed a method for lung coefficient characteristics for each sub-band. A
segmentation that was based on morphological probabilistic Neural Network (PNN) contains
analysis and circular filters. In the end, CNN an Input layer performing each neuron layer
was used, and the network had an accuracy containing N-1 neurons to N categories. Input
percentage of 84.6% throughout the whole neurons feed value to each hidden neuron
process. layer. Based on the contributions from each
Lyu et al. [7] an accuracy of 84.81% was class of inputs, the pattern and the summation
achieved with the Multi-Level CNN when it layer produces a probabilistic vector. Finally,
was applied to the LIDC dataset. Kumar et al. the output layer chooses the option with the
[8] proposed the use of deep features derived highest yield and 80% accuracy.
from an auto encoder in conjunction with a Krishna, S.T et. al. [15] given proposed a
binary decision tree as a classifier for the system in which they are using MRI-based
purpose of developing a CAD system that lung images. A Region Proposal Network
could identify lung cancer. The data provided (RPN)-based approach generated proposed
by a number of practitioners is utilized in regions for each image. Through optimization
conjunction with lung scans in order to extract of parameters, building spatial inputs in three
nodules. After that, the retrieved nodules are channels, and transmitting knowledge. “Aspect
put into the auto encoder. Further features are ratio = width of image/height of the image,
extracted using the fourth layer of the auto scale is the size of the image”. Then,
encoder, which has a total of five layers. The Anatomical features are used to design a false
categorization that used these criteria had an positive reduction scheme to avoid false
accuracy of 75.01% overall. Researchers from positives rate. Using 3.47 scan frames per
a variety of fields have each proposed their scan, the proposed method has 85.2%
own unique algorithm for screening for lung sensitivity.
cancer. It is possible for radiologists to make Diego Riquelme et. al.[17] Work proposed by
use of machine learning algorithms in order to given system used (CNN).The LIDC/IDRI
recognize and classify malignant tumors in dataset was used by this system. Arrays of
medical images. This can assist radiologists in pixels in an image are segmented to allow
making decisions, particularly in them to share certain characteristics based on
circumstances that are difficult to diagnose, their labels. It's typically used to locate objects
and it can also increase both speed and and boundaries like curves, and lines. (CNN)
accuracy. The analysis of the relevant research input layer containing a nonlinear regression
[2–8] reveals the importance of developing a among sets of arithmetical equations ions
method for improving accuracy via related to a vector matrix with the following
classification as a strategy. shape: (number of inputs) (h) (w) (channel).
Kumar, D., et. al. [14] is based on Thoracic X- For local bundling, which combines small
Ray images. They employed a two-stage clusters, a typical tile size of 4 x 4 is prevalent.
image rendition and categorization procedure During the output layer of the neural network,
(PNN). The pre-processing is necessary to the desired predictions are obtained. An
enhance the discrepancy, translucency, and individual set of biases and weights are
separation of background noise. To remove the applied before the final output is generated.
noise from the images, median filtering is The activation function was applied to define
used. As it runs through the signal, it replaces the final class. The proposed CNN architecture
each entry with its neighbouring median. Haar obtained 87.76% accuracy.
wavelet transforms for feature extraction. Pouria Moradi et.al.[18] They used the
During the classification process, the Haar LUNA16 dataset to train and evaluate the
wavelet transform generates power and system. In the proposed 3D CNN-based

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

model, false positives can be reduced while are determined by a brief grid search
sensitivity is high in detecting lung cancerous performed on the initial tiny model. A model
lesions. They use multilevel con-textual CNN, may be improved such that it makes the most
due to different sizes of nodules. Designed accurate predictions possible based on its
different size 3D CNN so that fusion of these hyperparameters via the use of grid search.
four networks can provide a good coverage Not only does the approach enhance the
error of classification. The input of the first accuracy of the calculation, but it also
convolution layer was convolved with two accelerates it, which can be measured in terms
kernels of size 3x3x16. ReLU (Rectified of the number of floating-point operations that
Linear Input) was used for activation. For the can be carried out in one second (FLOPS).
pre-processing stage, first, convert input They were able to accomplish a level of
images into Hounsfield Unit (HU) which is a accuracy that was 93.33%, sensitivity that was
standard unit for normalizing the intensity of 92.75%, and precision that was 93.75%.
CT voxels. The HU for water and air is 0 to
1000, respectively. “Stochastic Gradient III. CNN ALGORITHMS
Descent” (SGD) is used to train network Deep learning [9] has emerged as one of the
weights. 91.23% accuracy was obtained. most promising new areas for pattern
Sekar Sari et. al.[19] provided system is based identification and categorization in recent
on Resnet50. The Resnet50 dataset had its years. One of the most often used examples of
final two layers changed, and the public a deep neural network is the CNN. “There are
dataset that was available on Kaggle was several components that make up the network,
utilized. There were two levels of pooling for including the input layer, the hidden layers,
the global average: a layer that flattened, and a and the output layer. Within of the hidden
layer that was dense. They also introduced a layer are several other layers, including the
zero-padding layer as a component of the convolutional layer, the ReLU (Rectified
procedure of feature extraction that they used. Linear Unit), the pooling layer, the fully
They compared resnet50 with two models connected layers, and many more layers. In
based on Efficient Net, which is architecture order to build the Convolutional Network,
and a technique of scaling that was created these layers served as the basis for its
based on convolutional neural networks that construction”. CNN [10] provides users with
function evenly across all dimensions. This access to a variety of pre-trained architectural
comparison was made as part of their research. models, some of which are LeNet, AlexNet,
The depth of the network is denoted by, the GoogleNet, VGGNet, ResNet50, and others.
width by, and the size of the picture is denoted Figure 1 illustrates the CNN architecture that
by, where,, and are constant coefficients that is used for the identification of lung cancer.

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

Fig 1: Lung cancer detection Using CNN architectures

During the training phase of deep learning, the might bring up a few issues on its own. The
model is exposed to a significant amount of additional depth required a modification in the
data and is taught how to adjust its own model weights, which in turn elevated the end of the
weight and bias. These weights are then network. As a result, the prediction became
implemented into several different network less accurate in the initial layers as a result of
models so they may be tested. The new the higher end of the network. Another issue is
network model can get going using weights the massive amount of available parameter
that have been pre-trained [11]. A pre-trained space that was essential. In order to find a
model is one that has previously undergone solution to this problem, residual modules
training in the given field. AlexNet is a neural have been included into the overall image.
network that is far more advanced than LeNet. Two examples of ResNet networks are known
Rectified Linear Units, also known as ReLUs, as ResNet50 and ResNet152.
are employed in this network to introduce non- The Convolutional Neural Network can be
linearity, which in turn speeds up the network. built using a variety of architectures, including
“This network comprises of 62.3 million “AlexNet [12], GoogleNet [13], and
parameters, five convolutional layers, three ResNet50” [14]. Classify the images obtained
fully connected layers, and one output layer. In from the CT scan as either benign or malignant
addition, it has one output layer”. While by using these networks. In addition to an
GoogleNet achieved a decent level of unknown quantity of hidden layers and
accuracy, it required a significant amount of something known as an output layer, every
computer resources because to the very high network also contains something known as an
order of computations. After the final input layer. [15]. The information contained
convolutional layer, an average pooling within these networks is put to use in order to
method was used in lieu of fully-connected effectively classify the images. Learning that
layers; this change will result in a lower total is acquired in this manner is referred to as
number of parameters. GoogleNet was transfer learning [16].
replaced with this method. Up to this point, In this architecture, the first 5 sets of layers
increasing the network depth has likewise consist of the convolutional layer, the ReLU
automatically resulted in an improvement in layer, the normalization layer, and the max-
accuracy. Yet, the ResNet network depth pooling layer. After those levels come the fully

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

connected layers and the dropout layers [17], information from the pictures it uses as input.
and then finally the soft-max layer. The first This article enhanced ResNet50 such that it
Convolutional Network ever built was called was able to classify the binary class as either
AlexNet, and it included eight different layers. benign or malignant. Formerly, ResNet50 had
This network assigns categories to the the ability to distinguish between 1000
photographs that are fed into it by different classes. This upgraded version of
mechanically locating distinct elements within ResNet50 is able to identify the photographs
the images and then extracting those more accurately than previous systems that are
characteristics. AlexNet was able to identify a presently being used.
thousand distinct classes; the purpose of this
study is to modify this network so that it can IV. PROPOSED WORK
categorize the binary class, such as benign or Both are included in this design. The structure
malignant. This improved version of AlexNet of this model is comprised of one maxpool
performs picture classification more layer, one average pool layer, and a total of 48
effectively than other approaches currently in convolutional layers. “The ResNet-50 model is
use. comprised of five stages, and each stage has a
There are a total of 22 covert layers in convolution as well as an Identity block. Each
GoogleNet. The neural network has a depth convolution block has three convolution
that is much greater than that of AlexNet. The layers, and each identity block also has three
network is able to accurately classify the data convolution layers [9]. There are also three
in a more time efficient manner as a result of convolution layers in each identity block. Each
its enhanced depth. This network also convolution block has three layers of the
automatically identifies the photos it receives convolution operation applied to it”. The
and extracts information from the images it formation of a ResNet is accomplished by
receives as input. GoogleNet was able to constructing several Residual blocks in a
categorize a thousand distinct classes; This stacked fashion, one above the other. By using
study uses the same network, but updates it so x, we are able to sidestep the identity
that it can also categorize binary data, such as connections, so we can go ahead and connect
whether or not something is benign or directly to the layer. We are working on the
malignant. This enhanced version of assumption that identity links are what
GoogleNet identifies the photos more contribute to the value F. (x). The sum of X
effectively than other systems currently in use. and F is the answer (X). Using the ResNet-50,
There are 50 covert layers concealed inside you may train over 23 million distinct
ResNet50. The network has a depth that is parameters at the same time. With ResNet
greater than that of the GoogleNet. The assistance, getting around the connection is
network is able to accurately classify the data much easier. This makes it much simpler to
in a more time efficient manner as a result of get around the problem with the disappearing
its enhanced depth. This network is also gradient.
capable of automatically identifying the
photographs it obtains and deriving

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

Fig 2: System Architecture

V. RESULT ANALYSIS The statistical measures are computed after

The performance of the suggested approach is being based on the properties of the confusion
analysed via the use of the statistical metrics. matrix in order to arrive at their conclusions. A
Accuracy, sensitivity, specificity, and number of different statistical metrics,
precision are the statistical metrics. Accuracy including accuracy, sensitivity, specificity, and
is the most important of these. Taking into precision, are being tested during the course of
account the statistical measures, the suggested this research. This research analyzes many
technique need to provide a high degree of different statistical metrics, and there is a large
emphasis on the positive findings. The selection of statistical measurements
statistical measures are generated with the accessible to choose from. The mathematical
assistance of a confusion matrix. This expressions needed to compute the statistical
confusion matrix uses four criteria, including parameters are shown in the following list,
“false positives (FP), true positives (TP), false (TP + TN)
Accuracy =
negatives (FN), and true negatives (TN)”, (TP + FN + FP + TN)
among others. The following is an explanation × 100 (1)
of the terminology used in the confusion (TP)
Sensitivity = (2)
matrix: (TP + FN)
TP: The accurate forecast of the fault count in (TN)
specificity = (3)
the defence Medical whenever there is a (FP + TN)
positive instance. (TP)
Precision = (4)
FP: The inaccurate projection of the fault (TP + FP)
count in the Medical industry when the case is
positive. Using the formulae that have been described
TN: The accurate forecast of the number of before, an evaluation of each statistical
errors in the Medical industry's count when the measurement and a measurement of how well
case is negative. the suggested approach performs are carried
FN: The erroneous defect prediction count in out. A graphical depiction of the statistical
the Medical industry, which occurs when the measures that were obtained using the
situation is bad.

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

approach that is being suggested as well as the than or equal to 3 millimetres and a presumed
one that is presently being utilized is presented histology; nodules with a diameter of less than
in this section. 3 millimetres and an unknown origin; non-
The LIDC was established by the National nodules with a diameter of less than 3
Cancer Institute with the goal of enhancing millimetres but which are benign. There are a
research and development endeavours (NCI). total of 1008 patient records in the database.
The LIDC database was built using three types Figure 3 displays several examples of the CT
of items that needed to be tagged by four scan pictures that were taken from the LIDC
radiologists, and these were the categories that dataset.
were used: Nodules with a diameter of larger

Fig 3: Dataset images of a) Malignant b) benign from LIDC dataset

Table 1: CNN networks Accuracy
Method Accuracy Precision
AlexNet 89 80
GoogleNet 95.42 90
ResNet50 97.42 95

Dataset Size: 100 Accuracy

97.42
95.42

89

AlexNet GoogleNet ResNet50

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

Fig 4: Accuracy of Proposed System

Dataset Size: 100 Precision

95
90

80

AlexNet GoogleNet ResNet50

Fig 5: Precision of Proposed System

VI. CONCLUSION GLOBOCAN estimates of incidence and

The condition that causes lung malignant mortality worldwide for 36 cancers in 185
tumors is fatal, and the mortality rate countries. CA: A Cancer Journal for The
associated with it is extremely significant. The Clinician, 68(6): 394-424.
key to successfully treating lung malignant https://doi.org/10.3322/caac.21492
tumors illness is an early diagnosis. In this 2. Jiang, H., Qian, W., Gao, M., Li, Y.
research study, we presented a model for the (2018). An automatic detection system of
diagnosis of lung malignant lesions that was lung nodule based on multigroup patch-
built on ResNet50 and ResNet152. There were based deep learning network. IEEE
a great number of approaches and algorithms Journal of Biomedical and Health
to choose from; however, due to the utilization Informatics, 22(4): 1227-1237.
of deep neural networking, input values were https://doi.org/10.1109/JBHI.2017.272590
discarded and reset to zero. This results in the 3
omission of some characteristics, which in turn 3. Yang, H., Yu, H., Wang, G. (2016). Deep
reduces the accuracy of the outcome. As may learning for the classification of lung
be seen from the above observation, nodules. arXiv preprint,
Resnet152 achieved an accuracy of 0.93, arXiv:1611.06651.
whereas Resnet50 achieved 0.9123. The 4. Fan, L., Xia, Z., Zhang, X., Feng, X.
majority of the time, issues with vanishing (2017). Lung nodule detection based on
gradients occur in deep neural networks. The 3D convolutional neural networks. In the
ResNet50-based model proved useful in Frontiers and Advances in Data Science
assisting with the resolution of issues of this (FADS). 2017 International Conference on
kind. This helps to increase the accuracy of the IEEE, Xi'an, China.
results. https://doi.org/10.1109/FADS.2017.82531
. 84
References 5. Victor, R., Peixoto, S., Pires, S., Silva, P.,
1. Bray, F., Ferlay, J., Soerjomataram, I., Pedrosa, P., Filho, R. (2018). Lung nodule
Rebecca, L.S., Torre L.A., Jemal, A. classification via deep transfer learning in
(2018). Global cancer statistics 2018: CT lung images. In 2018 IEEE 31st

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

International Symposium on Computer- Vision, Halifax, 343 NS, Canada.

Based Medical Systems (CBMS), https://doi.org/10.1109/CRV.2015.25
Karlstad, Sweden. 15. Krishna, S.T., Kalluri, H.K. (2019). Deep
https://doi.org/10.1109/CBMS.2018.00050 learning and transfer learning approaches
6. Jin, X., Zhang, Y., Jin, Q. (2016). for image classification. International
Pulmonary nodule detection based on CT Journal of Recent Technology and
images using Convolution neural network. Engineering (IJRTE), 7(5S4): 427-432.
2016 9th International Symposium on 16. Riquelme, Diego, and Moulay A.
Computational Intelligence and Design Akhloufi. 2020. "Deep Learning for Lung
(ISCID), Hangzhou, China. Cancer Nodules Detection and
https://doi.org/10.1109/ISCID.2016.1053 Classification in CT Scans" AI 1, no. 1:
7. Lyu, J., Ling, S.H. (2018). Using multi- 28-67. https://doi.org/10.3390/ai1010003
level convolutional neural network for 17. Diego Riquelme, T., Yakopcic, C.,
classification of lung nodules on CT Westberg, S., Sidike, P., Nasrin, M.,
images. In 2018 40th Annual International Essen, B., Awwal, A., Asari, V. (2018).
Conference of the IEEE Engineering in The history began from AlexNet: A
Medicine and Biology Society (EMBC), comprehensive survey on deep learning
Honolulu, HI, USA, pp. 686-689. approaches. arXiv preprint
https://doi.org/10.1109/EMBC.2018. arXiv:1803.01164.
8512376 18. Sekar Sari , A survey on transfer learning.
8. J. Clerk Maxwell, A Treatise on IEEE Transactions on Knowledge and data
Electricity and Magnetism, 3rd ed., vol. 2. Engineering, 22(10): 1345-1359.
Oxford: Clarendon, 1892, pp.68–73. https://doi.org/10.1109/TKDE.2009.191
9. I. S. Jacobs and C. P. Bean, “Fine 19. Krizhevsky, A., Sutskever, I., Hinton, G.E.
particles, thin films and exchange (2017). ImageNet classification with deep
anisotropy,” in Magnetism, vol. III, G. T. convolutional neural networks.
Rado and H. Suhl, Eds. New York: Communications of the ACM, 60(6): 84-
Academic, 1963, pp. 271–350. 90. https://doi.org/10.1145/3065386
10. K. Elissa, “Title of paper if known,” 20. S, M. T., & Dr. Dattatray G. Takale, P. S.
unpublished. (May 2019). Load Balancing Energy
11. R. Nicole, “Title of paper with only first Efficient Protocol for Wireless Sensor
word capitalized,” J. Name Stand. Network. International Journal of
Abbrev., in press. Research and Analytical Reviews
12. Y. Yorozu, M. Hirano, K. Oka, and Y. (IJRAR), 153-158.
Tagawa, “Electron spectroscopy studies on 21. Szegedy, C., Liu, W., Jia, Y.Q., Sermanet,
magneto-optical media and plastic P., Reed, S., Anguelov, D., Erhan, D.,
substrate interface,” IEEE Transl. J. Magn. Vanhoucke, V., Rabinovich, A. (2015).
Japan, vol. 2, pp. 740–741, August 1987 Going deeper with convolutions. In
[Digests 9th Annual Conf. Magnetics Proceedings of the IEEE Conference on
Japan, p. 301, 1982]. Computer Vision and Pattern Recognition,
13. M. Young, The Technical Writer’s Boston, MA, USA.
Handbook. Mill Valley, CA: University https://doi.org/10.1109/CVPR.2015.72985
Science, 1989. 94
14. Kumar, D., Wong, A., Clausi, D.A. 22. He, T., Zhang, Z., Zhang, H., Zhang, Z.,
(2015). Lung nodule classification using Xie, J., Li, M. (2019). Bag of tricks for
deep features in CT images. 2015 12th image classification with convolutional
Conference on Computer and Robot neural networks. In Proceedings of the

ISSN 1303-5150 www.neuroquantology.com

NeuroQuantology | August 2022 | Volume 20 | Issue 10 | Page 12903-12914 | doi: 10.14704/nq.2022.20.10.NQ551250
Dr. Dattatray G. Takale/LUNG CANCER DETECTION BY USING CNN ARCHITECTURE MODELS

IEEE Conference on Computer Vision and

Pattern Recognition, pp. 558-567.
23. Liu, Y.H. (2018). Feature extraction and
image recognition with convolutional
neural networks. Journal of Physics:
Conference Series, 1087(6): 062032.
https://doi.org/10.1088/1742-
6596/1087/6/062032
24. Hussain, M., Bird, J.J., Faria, D.R. (2019).
A study on CNN transfer learning for
image classification. In UK Workshop on
Computational Intelligence, Springer,
Cham, pp. 191-202.
25. Yang, J., Yang, G. (2018). Modified
convolutional neural network based on
dropout and the stochastic gradient descent
optimizer. Algorithms, 11(3): 28.
https://doi.org/10.3390/a11030028
26. Fedorov, A., Hancock, M., Clunie, D.,
Brochhausen, M., Bona, J., Kirby, J.,
Freymann, J., Pieper, S., Aerts, S., Kikinis,
R., Prior, F. (2018). Standardized
representation of the LIDC annotations
using DICOM., PeerJ Preprint, e27378v1.
https://doi.org/10.7287/peerj.preprints.273
78v2
27. LIDC-IDRI. database available
https://wiki.cancerimagingarchive.net/disp
lay/Public/LI DC-IDRI, accessed on 18
March, 2019.
28. Hossin, M., Sulaiman, M.N. (2015). A
review on evaluation metrics for data
classification evaluations. International
Journal of Data Mining & Knowledge
Management Process, 5(2): 1-11.
https://doi.org/10.5121/ijdkp.2015.5201
29. Cui, X., Zhang, W., Tüske, Z., Picheny,
M. (2018). Evolutionary stochastic
gradient descent for optimization of deep
neural networks. In Advances in Neural
Information Processing Systems, pp.
6048-6058.
https://arxiv.org/abs/1810.06773

ISSN 1303-5150 www.neuroquantology.com

Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.