Abstract

Lung cancer is the most common cause of cancer deaths worldwide. Early
detection is crucial for successful treatment and increasing patient survival
rates. Artificial intelligence techniques can play a significant role in the early
detection of lung cancer. Various methods based on machine learning and deep
learning approaches are used to detect lung cancer. This research works aims to
develop automated methods to accurately identify and classify lung cancer in
CT scans by using computational intelligence techniques. The process typically
involves lobe segmentation, extracting candidate nodules, and classifying
nodules as either cancer or non-cancer. The proposed lung cancer classification
uses modified U-Net based lobe segmentation and nodule detection model
consisting of three phases. The first phase segments lobe using CT slice and
predicted mask using modified U-Net architecture and the second phase extracts
candidate nodule using predicted mask and label employing modified U-Net
architecture. Finally, the third phase is based on modified AlexNet, and a
support vector machine is applied to classify candidate nodules into cancer and
non-cancer. The experimental results of the proposed methodology for lobe
segmentation, candidate nodule extraction, and classification of lung cancer
have shown promising results on the publicly available LUAN16 dataset. The
modified AlexNet-SVM classification model achieves 97.98% of accuracy,
98.84% of sensitivity, 97.47% of specificity, 97.53% of precision, and 97.70%
of F1 for the classification of lung cancer.
Introduction
Lung cancer has a high incidence and death among all other
cancers. Approximately 1,958,310 total new cancer cases and
609,820 cancer death are anticipated to occur in the United
State of America in 2023 including 350 deaths daily from lung
cancer . An early lung cancer diagnosis can significantly reduce
the mortality rate and approximately 54% increase in survival
rate up to 5 years .

Image processing methods have been utilized to examine

medical images for many years. The computer-aided diagnosis
(CAD) system can provide a rapid, accurate, and efficient
diagnosis of disease, which can help in the treatment of patients.
Early detection of diseases become a major reason for to decline
in death rate for various kinds of cancers such as breast cancer,
kidney stones, brain cancer, blood cancer, stomach cancer, and
lung cancer. In this regard, various research efforts have been
done to aid and improve the diagnosis process of diseases from
medical imagery .

Researchers have developed various segmentation frameworks

or models to detect lung cancerous tumors to provide help to
radiologists. Lung cancer segmentation methods are divided into
two types: The first type comprises traditional techniques while
the second type consists of deep learning (DL) techniques.
Traditional techniques mostly centered on intensity-based
methods such as region growing], adaptive threshold , the
morphological method, active-contour model and shape
analysis . However, these methods are not robust in the case of
variation of tumor sizes as well as not appropriate for lung
segmentation tumors. Moreover, in these methods, when the
tumors are attached to the other organs, the performance of
tumor segmentation methods affects the level of automation,
which is consequently low.

Therefore, some lung tumor segmentation methods might be

misguided. Therefore, traditional methods are being replaced by
deep convolutional neural network (DCNN) models.

Existing deep learning techniques have outperformed several

image recognition problems. DL techniques can efficiently
extract important features optimally without human
participation. These techniques can improve accuracy in the
detection of various diseases in the medical field. Numerous
imaging modalities such as X-Rays, magnetic resonance imaging
(MRI) positron emission tomography (PET), and computed
tomography (CT) have been applied to detect pulmonary nodules.
The researchers are applying deep learning techniques such as
CSE-GAN , MSU-Net , dual-branch residual network(DB-
ResNet) , 3D-UNet, MSDS-U-Net , DS-CMSF , dual-path lung
nodules segmentation based on boundary enhancement and
hybrid transformer (DPBET) , DAS-NET , Lung PAYNet, LungNet-
SVM to improve the segmentation task in medical images. The
mentioned networks apply benchmark U-Net architecture and
obtained different level of accuracy but still, there is a need to
improve the accuracy of the segmentation process.

Image segmentation divides an image into different image

objects and boundaries. Medical image segmentation plays a
decisive role in the detection of several diseases through deep
learning methods. Automated segmentation methods based on
CT and MRI have increased in demand. Deep learning networks
mostly used encoder-decoder architectures and deep generative
models for medical image segmentation. The U-Net-based model
crops the feature maps from the encoding component, copy them
to the decoding component, and for segmentation map
generation .
Pulmonary cancer nodules are detected by various researchers
using different segmentation methods. Deep learning-based CAD
solutions can decrease the burden of medical experts to detect
various diseases particularly segmentation, detection, and
classification of lung cancer nodules.
An automatic deep learning-based model that segments, detects,
and classifies lung nodules increases the accuracy rate, and
reduces false positives while detecting lung nodules. Eventually,
lung cancer detection at an early stage will reduce the mortality
rate.

SECTION II.
Literature Review
Pulmonary nodule detection is a crucial task and early detection
of lung cancer is needed to reduce the mortality rate and
appropriate treatment. Various computational techniques are
used to detect lung cancer and several research methods have
been reported in the literature. Therefore, we have analyzed the
techniques below including segmentation, classification, and
detection of lung cancerous nodules.

In medical imaging, deep convolutional neural networks (DCNN)

made fabulous achievements. Shelhamer et al. [32] presented an
end-to-end network based on a fully convolutional network which
is more accurate for image segmentation. Ronneberger et
al. [33] introduced U-Net architecture consisting of encoder-
decoder and skip connection used to retain important
information from the different sizes of feature maps and attained
remarkable performance in medical image segmentation tasks.
Singadkar et al. [34] used a deep deconvolutional residual
network (DDRN) in the 2D CT lung images for automatic lung
nodule segmentation and this model was end-to-end trained with
fully captured resolution features. Fu et al. [35] introduced a
multi-task learning model consisting of a convolutional neural
network (CNN) to segment 2D CT images. Their model used an
arbitrary depth technique on entire nodule volumes and a slice
attention module applied to drop irrelevant slices. Moreover,
attribute and cross-attribute modules represented meaningful
relationships between attributes. Bruntha et al. [28] suggested
an inverted residual block used by the encoder and decoder to
segment lung nodules. In their proposed Lung PAYNet
architecture, they applied a pyramid attention network to extract
dense features from the encoder and decoder. According to
Liang et al. [36] segmentation uncertainty at the pulmonary
nodule boundary is considered a challenge and to overcome this
challenge the authors presented Uncertainty Analysis Based
Attention UNet (UAA-UNet) model. The proposed network deals
with uncertainty in edge regions and it contains two stages. In
the first stage, initial segmentation maps of pulmonary nodules
were found and uncertainty regions are focused on in the second
step. A UAA UNet model has achieved a sensitivity of 85.11%
and Dice of 86.89% for nodule segmentation. Wang et
al. [37] have designed a selective kernel V-Net architecture for
the extraction of multi-scale feature information and improved
lung nodule segmentation performance with Dice of 0.796,
Jaccard of 0.665, and 0.789 sensitivity. He and Li [38] presented
an ISHAP (Improved SHapley Additive exPlanations)-based
model to classify lung nodules. Medical prior knowledge was
used to extract semantic and radiomics features. ISHAP
explanation and recursive feature elimination algorithm were
applied to guide important features and classifiers with
parameters. Then, the ISHAP-based model utilized to classify
pulmonary nodules into cancer and non-cancer on the LIDC
dataset obtained 0.873 of accuracy, 0.885 of specificity, and
0.862 of sensitivity. Huidrom et al. [39] focused neuro-
evolutional approach containing a feed-forward neural network
for the detection of pulmonary nodules. This method worked with
particle swarm optimization and cuckoo search algorithm and
yielded 95.5% of accuracy and 95.8% of sensitivity. Similarly,
another research presented by [40] to detect lung cancer
detection based on CNN and generative adversarial networks
(GANs). Li et al. [41] used handcrafted features followed by the
convolutional neural network. Nageswaran et al. [42] presented
a lung cancer classification technique using various machine
learning (ML) methods such as artificial neural network (ANN),
K-nearest neighbors (KNN), and random forest. Nodule
classification was performed by Zhao et al. [43] which consisted
of an attentive module that scratches spatial and global
information. Furthermore, multilevel contextual information
encoded by the adaptive conv-kernels method improved nodule
classification accuracy. Bhaskar and Ganashree [44] introduced
an effective method using multi-scale Laplacian of Gaussian
filters and deep convolutional neural network to detect
pulmonary nodules and achieved 71.2% recall, and 93.2%
accuracy.

Han et al. [45] detected and classified lung nodules by applying a

3D ResNet algorithm and a fully connected neural network to
reduce the medical expert’s workload on the LUNA16 dataset.
Similarly, Bruntha et al. [46] used ResNet50 and a handcrafted
histogram of oriented gradient (HOG) for deep feature extraction
and handcrafted feature respectively. A support vector machine
(SVM) was used to classify non-cancer and cancer nodules for
this proposed hybridized model on the LIDC dataset. Al-Shabi et
al. [47] introduced a model for lung nodule classification namely
Progressive Growing Channel Attentive Non-Local (ProCAN)
network reached an accuracy of 95.28%. Huang et
al. [48] introduced an effective model based on a deep feature
optimization framework (DFOF) for lung cancer classification.
The model yielded 92.13% accuracy and 87.16%. recall and
94.16% precision.
Mahmood and Ahmed [49] introduced an automatic CAD system
based on AlexNet architecture to classify lung nodules. The
proposed AlexNet architecture was tuned with several layers and
hyperparameters to achieve superior performance. The model
achieved results of the pulmonary cancer screening trial were
98.9% of specificity and 98.7% of accuracy. Another research
work by Dodia et al. [50], presented an elagha initialization-
based fuzzy c-means clustering (EFCM) and SVM presented for
segmentation and detection of nodules respectively. Lyu et
al [51] suggested a model based on a multi-level cross-residual
network (ML-xResNet) to classify the lung nodules and achieved
92.19% of accuracy.

The major limitation of previous studies is shown in Table 1.

TABLE 1 Limitation of Previous Work

Table 1 is showing some limitations of the previous studies

included hand crafted features [39], [41], [46], the need to
improve accuracy [35], [40], [44], [47], [48], [51] lack of
transparency [15].

The core contribution of this research is as following:

 The main objective of this research is to provide the lung

cancer classification method using modified U-Net based
lobe segmentation and nodule detection.

 To enhance the effectiveness of the segmentation model, we

have implemented modified U-Net architecture for lobe
segmentation and ensure that lobe-segmentation model
training, validation and testing are brought out efficiently.

 The performance of the suggested candidate nodule

extraction model has been used modified U-Net
architecture for the detection of nodule and it provides
better results and investigated using various performance
statistical indicators.

 Finally, lung cancer classification model using AlexNet and

Support vector machine (SVM) for the classification is
 proposed and it classifies the lung nodule into cancer and
non-cancer. The proposed model achieves better results for
accurate and effective lung cancer classification and
treatment.

The rest of the paper is structured as follows; Section II covers

the literature review, Section III illustrates the proposed
methodology, and the results and discussion wrap up in Section
IV. Section V concludes with a conclusion and Section
VI describes limitation and future work.

SECTION III.
Proposed Methodology
Lung cancer disease is referred to as the most lethal disease
among all other cancers. Detection of lung cancer at an early
stage plays a major role in the successful treatment and
increases the survival rate. Various methods are used to detect
lung cancer such as CT, biopsy, blood test, and X-ray. Pulmonary
nodule detection is a challenging task because of the various
size, shapes, locations of nodules, and densities. Computational
intelligence techniques have been utilized to detect lung cancer
timely.

In this section, we propose lung cancer classification using

modified U-Net based lobe segmentation and nodule detection
model as demonstrated in Figure 1.

FIGURE 1.
Proposed lung cancer classification using modified U-Net-based lobe
segmentation and nodule detection model.

Show All

The model consists of three phases: Lobe segmentation,

candidate nodule extraction, and lung cancer classification. In
the lobe segmentation phase, modified U-Net architecture is
used to segment the input CT scans, and lobes are derived as
output. Whereas the next candidate nodule extraction phase uses
predicted lobes as input and modified U-Net based model is
applied for the extraction of the candidate nodule. Furthermore,
modified AlexNet-SVM based model is applied on patches of
candidate nodules in the third phase and classifies candidate
nodules as non-cancer and cancer.

A. Lobe Segmentation Phase

Lobe segmentation is the first phase of the lung cancer
classification using modified U-Net based lobe segmentation and
nodule detection model as shown in Figure 2. In this phase,
modified U-Net architecture is applied to the segment lobe from
CT scan images. The segmentation phase consists of two steps:
seg-lobe training and validation step and lobe segmentation step.
In the seg-lobe step, modified U-Net based model trained and
validated on LUNA16 CT scans dataset.

FIGURE 2.
Proposed modified U-Net architecture for lobe segmentation.

Show All

In the second step, the modified U-Net based model for lobe
segmentation predicts the mask from the test CT scans dataset
and by using the predicted mask, the lobe is extracted. In this
phase, the LUNA16 dataset consists of CT scans with labels that
are used as input to the proposed method for segmentation. The
lung cancer dataset consists of 888 CT scans which is divided
into 589 for cancer and 299 indicates non-cancer. In this
research a total number of 30 cancer CT scans are separated for
testing of the proposed method by using random selection
technique. A total number of 858 cancer and non-cancer CT
scans is separated for training and validation, and it is further
divided into 80% (686 CT scans) for training and 20% (172 CT
scans) for validation of the model for the lobe segmentation. The
model is trained on the 686 CT scans training dataset. After the
training of the seg-lobe model, it is validated on 172 CT scans. A
total number of 30 CT scans are provided to the seg-lobe model
for testing for the segmentation of CT scans. The seg-lobe model
predicts the masks from the testing 30 CT scans. Finally, the lobe
from slices of 30 CT scans is segmented using the predicted
mask of slices. U-Net architecture was designed by Ronneberger
et al. [33] for medical image segmentation in 2015.
U-Net architecture consisted of three main blocks, encoder,
decoder, and skip connection as shown in Figure 3.

FIGURE 3.
Proposed modified U-Net architecture for lobe segmentation.

Show All

The Encoder block receives the image as input and then extracts
useful features from an image using multiple convolutional
layers. Decoder block U-Net architecture is a combination of
several convolutional layers and transposed convolutional layers.
The convolutional layer represents in Eq. (1) and transposed
convolutional layer represents in Eq. (2).

φ=f(ω∗α+♭)(1)
View Source where α denotes the input, ω shows the layer’s
weight and ♭ represents the bias parameter and expresses the

φ=f(ω′∗α+♭)(2)
activation function.

View Source where α denotes the input, ω′ shows the layer’s

transposed weight matrix and ♭ represents the bias parameter
and expresses the activation function.
U-Net architecture also comprises concatenation operations.
Where feature maps from contracting combine with feature maps
from the expanding paths. The mathematical representation of
concatenation operations is shown in Eq. (3).

φ=concatenate(α1+α2)(3)
View Source where α1 and α2 represent the feature maps.
In the first phase of the lung cancer classification using modified
U-Net based lobe segmentation and nodule detection method,
the image input dimension is 512×512×1 followed by two
convolutional layers. Convolutional operations are performed on
two convolutional layers with 8 filter size, ReLU activation
function is used 3×3 kernel size, and the same padding, the
output 512×512×8 is denoted by C1.
The convolutional layer is the primary component of CNN
architecture where important features are extracted from the
input data. For this, convolutional operations are performed and
denoted by ∗ , the output of the convolutional operations is called
a features map. The convolutional operations are represented
in Eq. (4).
(m∗n)(p)=∫m(u)n(p−u)du(4)
View Source where the input matrix (image) is denoted m, n is
the filter or kernel, and convolutional operation is represented
as ∗ . The output of the convolution is called a feature map and is
represented by (m∗n )(p) and forwarded to a nonlinear activation
function.
Various nonlinear activation functions such as ReLU, Softmax,
Hyperbolic tangent (Tanh), and Sigmoid are applied to remove
linearity values. In this model, the nonlinear ReLU activation
function is applied and mathematically represented in Eq. (5).

f(y)=max(0,y)(5)
View Source Kernel initializer such as ‘he_normal’ is applied to
initialize the weights of a layer and prevent the vanishing
gradients problem. The initializer ‘he_normal’ is mostly used
while the ReLU activation function is applied in this research.
Next, the pooling layer is applied to reduce the spatial
dimensionality of the feature maps however, it retains the
important information. Various types of pooling layers such as
min pooling, max pooling, sum pooling, and average pooling can
be applied to reduce the dimensionality of the feature map. In
the proposed model, max pooling is applied on the feature map,
and it is represented in the Eq. (6).

z[g][h]=max(y[g:g+j][h:h+j])(6)
View Source where y represents the input feature map, z denotes
the output feature map, k is pool size, and g,h are the indexes of
the output feature map. Max operation is performed a j × j
window of the input feature map and the maximum value is
assigned to the corresponding location in the output feature
map. Next, the sigmoid activation function applies
and 512×512×8 is forwarded to the sigmoid activation function
represented in Eq. (7).
sigmoid(C7)=1(1+ec7)(7)
View Source where C7 is the input to the sigmoid activation
function.
A deep learning-based model requires computational resources
and extensive time for training. Various optimizers such as
gradient descent, stochastic gradient descent (SGD), Adagrad,
root mean square propagation (RMSprop), and adaptive moment
estimation (Adam) can be utilized to optimize model performance
and reduce the error rate. Adam’s method uses an adaptive
learning rate to compute parameters at each iteration and
shows Eq. (8) to Eq. (11).

gtstΔjtjt+1=Υ1×gt−1−(1−β1)×ht=Υ2×st−1−(1−β2)×ht=−htgtm
t+ε−−−−−−√∗×ht=jt+η×Δjt(8)(9)(10)(11)

View Source where η represents the initial learning rate, denotes

exponential gradients average along with, shows gradient at time
t along, express the exponential average of squares of gradients
along, and shows hyperparameters. In this model, Adam
optimizer is applied to increase the efficiency and decrease the
error rate of the proposed model. Finally, 1×1 convolutional
operation is performed on C9 and 512×512×1 as the final
output is achieved for the segmentation of Lobes.
B. Candidate Nodule Extraction Phase
The second phase of the lung cancer classification using modified
u-net based lobe segmentation and nodule detection model is
called candidate nodule extraction is shown in Figure 4. In this
research, a total of 888 CT scans in which 589 CT scans for
cancerous and 299 CT scans for non-cancerous.

FIGURE 4.
Proposed modified U-Net architecture for candidate nodule extraction.

Show All

This phase uses 589 cancer CT scan for training, validation, and
testing of the modified U-Net architecture for the candidate
nodule extraction model. We use the slices from 30 cancer CT
scans for testing the model. Slices from 559 cancer CT scan is
further divided into 80% (slices from 447 CT scan) for training
and 20% (slices from 112 CT scan) for validation of the candidate
nodule extraction model. This model is trained on the slices
obtained from 447 cancer CT scans training dataset. When the
training of the candidate nodule extraction model, the model is
validated on slices obtained from 112 cancer CT scans.

Furthermore, in the testing step, slices from 30 CT scans are

used for testing the candidate nodule extraction model. The
modified U-Net architecture for the candidate nodule extraction
model predicts the candidate nodule mask from the lobes of
slices of testing 30 cancer CT scans. Finally, the model predicts
the candidate nodule by using the predicted mask and label.

C. Lung Cancer Classification Phase

Finally, the last phase of the lung cancer classification using
modified U-Net based lobe segmentation and nodule detection
model classifies cancer or non-cancer using patches from the
candidate nodules.

In this research, patch size 48×48 is selected that based on

nodule and pixel size to train, validate, and test the modified
AlexNet-SVM architecture for Lung Cancer Classification. A total
number of 17006 patches are obtained from slices of 858 cancer
and non-cancer CT scan. It is further divided into 80% (13605)
patches for training and 20% (3401) patches for validation of the
model. Patches obtained from slices of 30 CT scan is used to test
the model and it predicts lung cancer into non-cancer and
cancer. Modified AlexNet-SVM architecture for lung cancer
classification consists of lung cancer classification training and
validation phase and lung cancer classification phase is shown
in Figure 5.

FIGURE 5.
Proposed modified AlexNet architecture for classification of lung cancer
patches.

Show All

In the lung cancer classification training and validation phase,

modified AlexNet-SVM architecture model is trained and
validated on 48×48 patch size. Modified AlexNet architecture
extracted features from input patches to obtain important
information. Stochastic gradient descent (SGD) optimizer is used
with hyperparameters such as 200 epochs, 50 batch size, and
0.0001 learning rate. The modified AlexNet architecture
comprises eight convolutional and three max pooling layers. The
convolutional layer is responsible for extracting valuable features
from patches and pooling layer and which is used to reduce the
size of the feature map but keeps the important information. In
this research, max pooling layer is used on the feature map. After
the max pooling layer, the feature map matrix is transformed into
a single long vector, and it is called flattening.
The modified AlexNet architecture takes
input 48×48×1 grayscale patch size as demonstrated in Figure
6. The first three convolutional layers are used 32 filters along
with 3×3 filter size, the same padding, and ReLU AF is applied
to remove non-linearity from the feature map. Next, the max
pooling layer is used 2×2 filter size, stride 2 and resulting patch
size reduces and dimension of patches become 24×24×32 .

FIGURE 6.
Proposed modified AlexNet architecture for classification of lung cancer
patches.

Show All

The sigmoid activation function produces class score from the

output of a fully connected layer. Finally, SVM is utilized to
classify lung cancer into cancer and non-cancer. Afterward,
training and validation of the modified AlexNet-SVM model, the
patches from the 30 testing CT scans are forwarded to the model
to evaluate the performance of the modified AlexNet-SVM
architecture for lung cancer classification and classify patches
into cancer and non-cancer.

SECTION IV.
Simulation Results
Generally, digital imaging and communications in medicine
(DICOM) files are used to store CT scan images. The DICOM file
comprises the CT image along with image information in the
DICOM header and raw data related to the CT image. The slices
of CT scan have instance numbers which are also mentioned in
the DICOM header. The lung cancer classification uses modified
U-Net based lobe segmentation, and the nodule detection
method takes LUNA16 CT scans dataset as input for the
detection of lung cancer using computational intelligence
techniques. LUNA16 dataset is stored in the form of MetaImage
(mhd/raw) format. LUNA16 dataset contains 888 CT scans which
are further divided into 589 cancer CT scans and 299 non-cancer
CT scans.

A total number of slices from 30 CT scans are separated from

cancer CT scan for testing of the proposed methodology. In the
lobe-segmentation phase, the remaining slices from 858 CT
scans have been used for training and validation purposes while
slices from 30 CT scans have been used for testing the proposed
model. Next, slices from 589 cancer CT scans are utilized to
train, validate, and test the proposed methodology. Slices from
559 CT scans have been utilized for training and validation and
the remaining slices from 30 cancer CT scans are separated for
testing purposes. Finally, a total number of 17006 patches are
obtained from slices of 858 cancer and non-cancer CT scans have
been utilized to train and validate the proposed model the
trained model predicts cancer and non-cancer patches by using
slices from 30 cancer CT scans which are separated for testing
purposes.

The performance of the proposed methodology is measured using

various statistical metrics. Various performance metrics,
including Dice, IoU, sensitivity, and precision have been applied
to assess the performance of the proposed model. Mostly, Dice
and IoU statistical metrics are applied for segmentation
techniques. The Dice metric evaluates the connection between
the segmented output and ground truth.

The IoU calculates the area of overlap between the ground truth
and predicted segmentation based on the union of the outputs of
the ground truth and predicted segmentation. Sensitivity and
precision metrics are also applied to assess the robustness of the
proposed model. In the lobe segmentation step, the original
size 512×512 of the CT images are utilized in the proposed
model. Dice, IoU, sensitivity, and precision are used to evaluate
the performance of the Lobe segmentation model. Segmentation
lobe step, U-Net based modified architecture has been applied to
segment the lobe.
A. Results of Lobe Segmentation Phase
In the Lobe segmentation phase, modified U-Net architecture for
the lobe segmentation model segments the input CT scan images
into the lobe. The modified U-Net architecture for lobe
segmentation and Vanilla U-Net predicts mask from the input CT
scan dataset and results are shown in Figure 7. The lobe
segmentation phase consists of two steps, one is the Seg-Lob
training and validation step and the second is called Lobe
segmentation step. In the first step, Seg-Lobe based on modified
U-Net architecture consists of three encoders, three decoders,
and three skip connections. The seg-Lobe model is used to train
80% of the dataset and after the training step, the Seg-Lobe
model validates on 20% of the dataset of the CT scan images. In
the next step, CT scans test dataset is provided to Seg-Lobe
trained model to predict the mask of CT scan images. Then label
of the test dataset is provided to segment the lobe and the lobe is
extracted by using the predicted mask.

FIGURE 7.
Lobe mask prediction using modified U-Net architecture and vanilla U-Net
architecture.

Show All

The outcomes of performance indicators including Dice, IoU,

sensitivity, and precision are 90.32%, 82.35%, 87.5% and 93.33%
respectively obtains by modified U-Net architecture for lobe
mask prediction and Vanilla U-Net achieves 83.40% of Dice,
72.35% of IoU, 82.55% of sensitivity and 85.42% of precision.

B. Results of Candidate Nodule Extraction Phase

The next phase of lung cancer classification using modified U-
Net based lobe segmentation and nodule detection method is
candidate nodule extraction from the input segmented lobe. The
modified U-Net architecture for candidate nodule extraction and
Vanilla U-Net predicts candidate nodule mask from the
segmented lobe and results are shown in Figure 8.

FIGURE 8.
Candidate nodule extraction using modified U-Net architecture and vanilla U-
Net.

Show All
The candidate nodule extraction phase consists of two steps, one
is the candidate nodule extraction training and validation phase
where proposed model is trained and validated. In the second
step is called mask and candidate nodule extraction phase.
Candidate nodule extraction consists of three encoders, three
decoders and three skip connections. The candidate nodule
extraction model is used to train 80% of the dataset and after the
training step, the candidate nodule extraction model validates
from 20% of the dataset of the CT scan images. Candidate nodule
extraction model using modified U-Net architecture is trained
and validated on training and validating datasets.

In the next step, the segmented lobe test dataset is provided to

the candidate nodule extraction trained model to predict the
mask and then the label of the test dataset is provided to extract
the nodule, and the nodule is extracted by using the predicted
mask. Comparison analysis for candidate nodule extraction using
modified U-Net architecture, Vanilla U-Net and existing state-of-
the-art approaches illustrated in Table 2.

TABLE 2 Comparison Analysis of Candidate Nodule Extraction Using Modified

U-Net Architecture and Vanilla U-Net Model With Existing State-of-the-Art
Methods

C. Results of Lung Cancer Classification Phase

The last phase of lung cancer classification using modified U-Net
based lobe segmentation and nodule detection method consists
of lung nodule classification. The lung nodule classification phase
comprises two steps. The first step is called training and
validation of the lung cancer classification phase and the second
step is called the lung cancer classification phase. In the first
step, the patches from 858 CT scans are used and divided into
80% for training and 20% for validation purposes. Modified
AlexNet architecture consists of eight convolutional layers and
three max-pooling layers followed by two fully connected layers
and the SVM classifier has been applied to the classification of
lung cancer. When the classification model is trained on 80% of
patches and validated on 20% of patches. The trained model is
tested on patches obtained from 30 CT scans. A confusion matrix
has been employed to evaluate the performance of modified
AlexNet-SVM model to classify the lung nodules. A total number
of 13604 patches from 858 cancer and non-cancer CT scans are
obtained to train modified AlexNet-SVM model shown in Table 3.

TABLE 3 Confusion Matrix of Modified Alexnet-SVM Classification Model

(Training)

A total number of 13604 sample patches are divided into two

groups named cancer and non-cancer. In the non-cancer group, a
total number of 6802 non-cancer patches are used to train the
AlexNet-SVM model and it correctly predicts 6715 sample
patches as non-cancer and predicts wrongly 87 sample patches.
In the cancer group, a total number of 6802 sample patches are
used for the prediction of cancer, the modified AlexNet-SVM
model wrongly predicts 95 sample patches as non-cancer and
correctly predicts 6707 sample patches as cancer.

Confusion matrix of modified AlexNet-SVM classification model

for testing is shown in Table 4.

TABLE 4 Confusion Matrix of Modified AlexNet-SVM Classification Model

(Testing)

A total number of 1188 patches are obtained to test modified

AlexNet-SVM. Furthermore, a total number of 1188 sample
patches are divided into two groups named cancer and non-
cancer. In the non-cancer group, a total number of 594 patches
are used to test the performance of the modified AlexNet-SVM
model and it correctly predicts 579 sample patches as non-
cancer and wrongly predicts 15 sample patches. In the cancer
group, a total number of 594 sample patches are used for the
prediction of cancer, the modified AlexNet-SVM model wrongly
predicts 9 sample patches as non-cancer and correctly predicts
585 sample patches as cancer.

The various statistical metrics for evaluation such as accuracy,

miss rate, sensitivity, specificity, precision, and F1 are used to
test the performance of the modified AlexNet-SVM model as
presented in Table 5.
TABLE 5 Comparison Analysis of Lung Cancer Classification by Using
AlexNet-SVM Model

Other parameters have also been calculated for the performance

of the proposed model such as a Negative predictive value of
98.03%, false. negative rate of 2.14%, false positive rate of
2.27%, false discovery rate of 2.47%, and false omission rate of
1.96%.

SECTION V.
Conclusion
This research presents efficient and effective methods for lobe
segmentation, candidate nodule extraction, and lung cancer
classification which improved the accuracy. The model uses
modified U-Net architecture for lobe segmentation and candidate
nodule extraction. Furthermore, modified AlexNet-SVM applies
to the classification of pulmonary nodules. The modified AlexNet-
SVM based model comprises eight convolutional, three pooling,
two fully connected layers, and an SVM algorithm that is used to
classify lung cancer.

The experimental outcomes for the segmentation lobe using

modified U-Net model are 90.32% of Dice, 82.35% of IoU, 87.5%
of Sensitivity, and 93.33% of precision, whereas Vanilla U-Net
achieves 83.40% of Dice, 72.35% of IoU, 82.55% of sensitivity
and 85.42% of precision. Next, modified U-Net candidate nodule
extraction model shows results Dice 87.42%, 77.65% of IoU,
92.96% of sensitivity, and 82.5% of precision whereas Vanilla U-
Net achieves 78.57% of Dice, 64.71% of sensitivity, 74.83% of
precision and 82.71% of IoU for candidate nodule extraction.
Finally, the nodule classification phase of the proposed model
shows 97.98% of accuracy, 98.84% of sensitivity, 97.47% of
specificity, 97.53% of precision, and 97.70% of F1. The
experimental results of lung cancer classification using modified
U-Net based lobe segmentation and nodule detection model have
shown outstanding performance.

SECTION VI.
Limitation and Future Work
Lung cancer classification using modified U-Net based lobe
segmentation and nodule detection model segments candidate
nodules and classifies lung cancer into non-cancer and cancer.
The model based on modified U-Net architecture to segment lobe
and candidate nodule and modified AlexNet architecture with
SVM is applied to classify the lung nodules. The current research
has some limitations, for example, it used the LUNA16 dataset to
train, validate, and test purposes. The other publicly available
lung cancer datasets can be implemented to test the
performance of the Lung cancer classification using modified U-
Net based lobe segmentation and nodule detection model.




Authors
Figures
References
Citations
By Papers
By Patents
ExportReferences & Cited By
Cites in Papers - IEEE (1) | Other Publishers (4)

Select All

1.
Sabilla Halimatus Mahmud, Indah Soesanti, Rudy Hartanto, "Deep Learning Techniques for
Lung Cancer Detection: A Systematic Literature Review", 2023 6th International Conference
on Information and Communications Technology (ICOIACT), pp.200-205, 2023.
Show Article
Google Scholar
Cites in Papers - Other Publishers (4)
1.
Maheswari Sivakumar, Sundar Chinnasamy, Thanabal MS, "An efficient combined
intelligent system for segmentation and classification of lung cancer computed tomography
images", PeerJ Computer Science, vol.10, pp.e1802, 2024.
CrossRef Google Scholar
2.
V. Nisha Jenipher, S. Radhika, "Lung tumor cell classification with lightweight mobileNetV2
and attention-based SCAM enhanced faster R-CNN", Evolving Systems, 2024.
CrossRef Google Scholar
3.
Humberto de Jesús Ochoa Domínguez, Vianey Guadalupe Cruz Sánchez, Osslan Osiris
Vergara Villegas, "Demystifying Deep Learning Building Blocks", Mathematics, vol.12,
no.2, pp.296, 2024.
CrossRef Google Scholar
4.
P. Princy Magdaline, T.R. Ganesh Babu, "Detection of lung cancer using novel attention gate
residual U-Net model and KNN classifier from computer tomography images", Journal of
Intelligent & Fuzzy Systems, pp.1, 2023.
CrossRef Google Scholar