Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

Deep Learning-based Lung Cancer Classification of CT Images

using Augmented Convolutional Neural Networks
Bushara A. R.∗ and Vinod Kumar R. S.+

∗
Research Scholar, Department of ECE, Noorul Islam Center for Higher Education, Tamil Nadu, India
+
Professor, Department of ECE, Noorul Islam Center for Higher Education, Tamil Nadu, India.

Received 15th of August 2021; accepted 14th of June 2022

Abstract
Lung cancer is worldwide the second death cancer, both in prevalence and lethality, both for women and
men. The applicability of machine learning and pattern classification in lung cancer detection and classifi-
cation is proposed. Pattern classification algorithms classify the input data into different classes underlying
the characteristic features in the input. Early identification of lung cancer using pattern recognition can save
lives by analyzing the significant number of Computed Tomography (CT) images. Convolutional Neural
Networks (CNN) recently achieved remarkable results in various applications including Lung cancer detec-
tion in Deep Learning. The deployment of augmentation to improve the accuracy of a Convolutional Neural
Network has been proposed. Data augmentation is utilized to find suitable training samples from existing
training sets by employing various transformations such as scaling, rotation, and contrast modification. The
Lung Imaging Database Consortium-Image Database Resource Initiative (LIDC-IDRI) database is utilized
to assess the networks. The proposed work showed an overall accuracy of 95% Precision, recall, and F1
score for benign test data are 0.93, 0.96, and 0.95 respectively, and 0.96, 0.93, and 0.95 for malignant test
data. The proposed system has impressive results when compared to other state-of-the-art approaches.

Key Words: Lung Cancer Detection, Deep Learning, Convolutional Neural Networks, Computed Tomogra-
phy, Data Augmentation.

1 Introduction
Lung cancer is the second most lethal cancer in terms of prevalence and first in terms of mortality for women
and men all around the world. As per the report, new cases and sexual fatalities estimated in the United States,
2021 [1], 12% (119,100) of all incidents in men (116,660) and 13% (116,660) among women are lung cancer.
However, lung cancer is the major cause of mortality in both men (22%, 69410) and women (22%, 62470),
and patients with the early-stage disease have a survival rate of five years is 59%, whereas advanced cancer
people have a survival rate of 5 years of fewer than 6% [1]. There are two forms of lung cancer, Non-Small
Cell Lung Cancer (NSCLC) and Small Cell Lung Cancer (SCLC). Adenocarcinoma (ADC) and Squamous Cell
Carcinoma (SqCC) are sub types of NSCLC, which account for 80 to 85 percent of lung cancer cases [2]. The
Correspondence to: <[email protected]>
Recommended for acceptance by <Angel D. Sappa>
https://doi.org/10.5565/rev/elcvia.1490
ELCVIA ISSN:1577-5097
Published by Computer Vision Center / Universitat Autònoma de Barcelona, Barcelona, Spain
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 131

goal of early detection of lung cancer is required because symptoms only arise in advanced stages, and early
detection can cure or facilitate treatment, saving many lives [3]. Medical imaging is a critical tool for detecting
and diagnosing cancer early on. Medical imaging has been used for early cancer detection, evaluation, and
follow-up during procedures, as is well known [4]. Manual interpretation of a large number of medical images
can be tedious and time-consuming, as well as prone to human bias and error. Model classification algorithms
classify the input data into different classes as the basis for the characteristic features of the input. Early
identification of lung cancer using pattern recognition can save lives by analyzing large numbers of CT images.
As a result, starting in the early 1980s, Computer-Aided Diagnostic (CAD) systems were implemented to help
doctors interpret medical images more efficiently [5].
Image processing methods have been used in the past to diagnose lung cancer [6] & [7]. Machine learn-
ing techniques aid in the diagnosis and classification of lung nodules by evaluating CT images obtained using
artificial intelligence algorithms. Preprocessing, segmentation, object detection, feature extraction, and clas-
sification processes are used by such systems to investigate images. The main step in implementing machine
learning is feature extraction, which can be done in a CAD framework. [[8]-[13]] focused on various methods
for extracting features and classification for lung cancer. Here Haralick features, Gabor features, Local Bi-
nary Pattern [8], Scale Invariant Feature Transform, Histogram Oriented Gradient features [9], and minimum
Redundancy Maximum Relevance [10] have been adopted as feature selection, and Support Vector Machine,
k Nearest Neighbor [11], Rule-Based Classifier, Naive Bayes Classifier [12], Artificial Neural Network [10],
Feed-Forward Neural Network [13] are proposed for classification. A CAD system has been developed that
focuses on a new heuristic search technique for optimizing the Back Propagation Neural Network in distin-
guishing nodules from non-nodules [14]. The Fuzzy C-mean clustering [15] and other active contour models
[16] can be used to segment lung nodules in CT data. These approaches primarily concentrate on creating
good feature descriptors, many of which must be hand-crafted and machine learning techniques. These feature
extraction-based methods have several flaws, like a lower level of accuracy. This flaw prevents CAD systems
from improving their performance any further.
Several researchers have proposed the use of Artificial Intelligence (AI), especially Deep Learning to over-
come these limitations and improve results. Deep learning has the advantage of being able to perform end-to-
end detection in CAD systems by learning the most critical features during training and can learn features
from observed data without the need for hand-engineered features[17]. The Convolution Neural Network
(CNN), Deep Neural Network (DNN), and Stacked Auto Encoder (SAE) [18] are the three major deep learn-
ing structures for cancer detection. CNN yielded the best performance as compared to DNN and SAE [19].
Convolutional Neural Networks[20] (CNNs, or ConvNets) are a type of multi-layer neural network that can
distinguish visual patterns from pixel images with little or no preprocessing. AlexNet [21], VGG 16 [22], In-
ception (GoogleNet) [23], ResNet [24], and DenseNet [25] are the CNN architectures that have been employed
to investigate for lung nodule detection and classification.
The objective of this article is to develop a deep learning architecture for early lung cancer diagnosis and
classification using Augmented CNN. Presently overfitting is an incredibly common issue in the image classifi-
cation. Eventhough overfitting learned the features of the training set extremely well, it is unable to generalize
and accurately predict the output when given data that differs slightly from the exact data. The purpose of the
proposed research work is to alleviate the overfitting problem that occurs in CNN architecture by employing
data augmentation. Augmentation is a technique for expanding the size of data sets to improve accuracy. The
outcome of the research is a CNN-based lung cancer detection and classification method with better perfor-
mance matrices in terms of accuracy.
The following is a breakdown of the paper’s structure. The second section contains a relation between
CNN and the various state-of-the-art as CNN architectures. The third section discusses the architectures of the
developed augmented CNN along with model parameters and the performance matrices used. The experimental
results and discussion are elaborated in the fourth section. The paper ends with the conclusion in the last section.
132 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

2 Background

Early detection is crucial in the diagnosis of cancer and can improve long-term survival chances. As a result,
research into lung cancer detection and categorization has become a major area in recent years. Several studies
have used CNNs for lung cancer classification and detection to develop a reliable method. Some of them are
briefed in this section. S. Bhatia et al. [26] proposed an approach to lung cancer detection and feature extrac-
tion using deep residual networks. Features are extracted using UNet and ResNet models which are fed into
multiple classifiers. In addition, XGBoost and Random Forest, together with the individual forecasts, predict
the probability of malignancy in a CT scan. The accuracy obtained by the research work is 84% on LIDC-IRDI
datasets. M. Kriegsmann et al. [27] examined different lung cancer types associated with the University Clinic
Heidelberg with assistance from the Tissue Biobank of the National Center of Tumor Disorders has compiled,
scanned, annotated, and image patches taken from the Institute of Pathology Archive for 30 skeletal muscles
as control. They investigated different configurations of CNN architectures for classification and claimed that
the optimized InceptionV3 CNN architecture obtained the greatest classification accuracy. A new approach for
automatic pulmonary nodule detection from volumetric CT scans that use 3D CNN to reduce false positives
was proposed by Q. Dou et al. [28].
The LUNA16 Challenge, a dataset that was analyzed, achieved the highest Competition Performance Metric
0.827 score in the challenge. On histopathology slides, N. Coudray et al. [29] used convolutional neural
networks such as Google’s Inception v3 to diagnose lung cancer and classify lung cancer into Adenocarcinoma
and squamous cell carcinoma, which are the most common subtypes. The performance of this method is
comparable to that of pathologists, with an average area under the curve of 0.97. To address the noise in an
image and the morphology of nodules, W. J. Sori et al. [30] developed a “denoising first” two-path CNN and
tested it on the Kaggle Data Science Bowl 2017 challenge. This model combines denoising and detection in
an end-to-end approach, resulting in improved lung cancer detection. P. M. Shakeel et al. [31] proposed an
improved profuse clustering technique and the Deep Learning with Instantaneously Trained Neural Networks
method was used to evaluate lung CT images to predict lung cancer. The CT images of the lungs were taken
from the Cancer Imaging Archive dataset. The improvised 3D AlexNet with lightweight architecture was
used by E. S. Neal Joshua et al [32] to investigate lung nodule classification, and this network made use of
the multiview network technique in its entirety. They performed binary classification on CT images from the
LUNA 16 database.
Heuvelm et al. [33] suggested deep learning to identify malignant nodules and retrospectively validating
the Lung Cancer Prediction CNN on an isolated dataset of indeterminate nodules in a European multicentre,
with benign nodules excluded, but preserving a high sensitivity to lung cancer. Masud et al. [34] presented
a new supervised learning approach based on deep learning that identifies five different types of tissues found
in lung and colon lesions by observing their corresponding pathological data sets using LC25000 datasets to
train and validate the method. For image classification, two types of domain transformations were used to
extract four sets of features. The resulting features were combined to create a group of features containing
both kinds of information. T. L. Chaunzwa et al. [35] suggested a radiomic approach to the prediction of
NSCLC non-invasive CT tumor histology. They trained and validated Convolutional Neural Network data on
311 patients undergoing surgery at the Massachusetts General Hospital, prioritizing the two most common
forms of histology, named ADC and SqCC.
The complete system for detecting and diagnosing lung cancer has been introduced by O. Ozdemir et al.
[36] using low-dose CT scans, model uncertainty characterized by Monte Carlo dropout, and deep ensembles.
This has demonstrated that calculating model uncertainty allows the system to deliver calibrated classification
chances for diagnostic treatments based on utility or risk reliability. The background of this research paper
reveals that many methods for detecting and classifying lung cancer have been developed, with CNN being
the most popular. The authors discovered that a CNN system with data augmentation effectively mitigates
overfitting and improves accuracy.
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 133

3 Materials and Methods

The developed Augmented Convolutional Neural Network for classification consists of three main stages
namely i) CT Lung image Data Acquisition ii) Augmentation and iii) Classification using CNN. As shown
in Figure 1, the workflow begins with the gathering of a primary dataset containing two picture classes: one for
benign CT images that will not become cancer and the other for malignant which is cancerous CT images. The
data set was expanded using typical augmentation procedures to increase its size in the research’s following
step. In the next phase, the dataset was utilized to train the model. After training, the model was put to the test
for the performance of the proposed method in terms of cancer identification and classification. The chosen test
dataset of LIDC-IDRI was used to access the performance measures of augmented CNN.

Figure 1: Workflow of the Proposed System

3.1 CT Lung Image Data Acquisition

The images of the lungs were collected from the Lung Imaging Database Consortium-Image Database Resource
Initiative (LIDC-IDRI) [37]. DICOM format is used for all images, with slices ranging from 1.25 to 2.5 mm
in size, and pixels as small as 512 × 512 in size [38]. The thickness in LIDC-IDRI ranges from 0.48mm
to 0.72mm [39]. Several parameters, including lesion size, tumor type, and location, have been recorded by
four radiologists. All images in the LIDC datasets were classified as non-nodules, nodules less than 3 mm in
diameter, or nodules greater than 3 mm in diameter. For the proposed investigation, a benign nodule was defined
as one that was less than 3 cm in diameter, and a malignant nodule was defined as one that was greater than 3
cm in diameter. Figure 2 depicts the sample images of benign and malignant lungs on Computed Tomography
(CT).
The first step in building and training a convolutional neural network that can classify lung images as benign
or malignant is to obtain and prepare the datasets for training the model. The images are copied into the train,
test, and valid directories, which contain benign and malignant directories, after the datasets from the LIDC -
IDRI data sets are downloaded. This experiment used 2066 images, of which 80% (1653) of them were used
for training and 20% (413) for testing. Validation data were taken from 20% (330) of the training data.

3.2 Augmentation
In deep learning, data augmentation is used to find adequate training samples from existing training sets using
various transformations such as scaling, rotation, and contrast modification to improve the proposed method’s
accuracy[22]. Figure 3 depicts some of the augmented images used in the proposed work. Data Augmentation
allows for the creation of many more images with the same output, and as data size artificially extends, CNN
becomes more robust and can avoid overfitting problems. The rescale, shearing, zooming, and horizontal flip
134 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

(a) Benign nodule (b) Malignant nodule

Figure 2: Lung Images (Courtesy: LIDC- IDRI Dataset)

are performed for each image. This work employs an image data generator to accomplish the task. Here image
manipulation is accomplished through rescaling to normalize the image, and the scaling is accomplished by the
value 1/255. Shearing is another method that is used to produce excellent results with values ranging from 0 to
1, and in the proposed work, 0.2 is selected for shearing. The zoom parameter was set to 0.2, which resulted in
a random zoom range of 0% to 20%. The image will be randomly flipped if the Horizontal flip is set to true.

Figure 3: A subset of new images after the image augmentation technique

3.3 Convolutional Neural Networks (CNN)

The CNN is the central component of computer vision, learning feature maps of a given image and categorizing
it by detecting a variety of abstract features ranging from simple to complex. The network then uses these
discriminatory features to predict the correct image category [38]. The basic functionalities are normalization,
convolution, activation, pooling, fully connected layers, and classifier as shown in Figure 4.

Pre-Processing: The data must be pre-processed before being passed to the networks to make the CT scans
homogenous. The general pre-processing steps are mean subtraction [28], normalization [39], and local
contrast normalization. The pixel values of each image are converted to Hounsfield Units (HU), a ra-
diodensity calculation. Since tumors grow on lung tissue, segmentation can be used to mask the bone,
outside air, and other noise-causing substances, leaving only lung tissue details for the classifier [40].
The image data was subjected to isotropic rescaling using a linear interpolator and density normalization
using mean subtraction and linear transformation to reduce distortion [35]. Figure 5 depicts Hounsfield
Units (HU) of a Lung Image. Air, lipids, blood, and soft tissues are all found in the lungs. Figure 5
shows that HU units range from -1000 to 300, with a normal value of -700 to -600 for the lung. As a
preprocessing phase, the proposed approach used normalization and data augmentation.
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 135

Figure 4: Basic CNN architecture for Lung Cancer Detection

Figure 5: Hounsfield Units (HU) of a Lung Image

Convolution: The convolutional layer, which is at the heart of the network, performs either the convolution of
the input image if it is at the input layer or the convolution of the feature map available from previous
layers if it is at any other layers[41]. The convolution operation is performed using Equation (1)
A
X A
X
y(m, n) = w(p, q)x(m − p, n − q) (1)
p=−A q=−A

where y(m,n) represents the image sample at (m,n) in the output image, the convolution kernel coeffi-
cients that will be employed as weights are represented by w. All the convolutional layers in the proposed
system had a kernel size of w. In the proposed work a 3 × 3 kernel is chosen. The coefficients are ran-
domly initialized and they get adjusted automatically during iteration for obtaining maximum accuracy.

The pooling layer: The pooling layer uses a downsampling method to achieve spatial invariance following a
convolution layer. In Max pooling, the maximum values are used wherein in average pooling the average
values are used.

Activation Function: Every node in the convolutional layer has a non-linear activation function. Non-linear
mapping’s power is that it can map non-linear, non-separable data points into linearly separable data
points, which aids classification.

Fully Connected: One or more completely connected layers are included on the output side, and these fully
connected layers aim to classify the input data. The fully connected layer works with a flattened input
that connects each information to each neuron.
136 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

Softmax Classifier: Softmax is the last activation function of the neural network, converting a number vector
into a probability vector and assigning decimal probabilities to each class in a multiple-class problem
[42]. The proposed CNN model architecture included 14 layers, consisting of 5 Two dimensional Con-
volutional layers, 3 fully connected layers, ReLu and Sigmoid as Activation Function, 5 Max-pooling
layers, and 1 Flattening layer. Table 1 shows the parameters employed in the experimentation.

Parameter Value
Input Image Dimension (500, 500, 3)
Kernel Size 3×3
Filter 32, 64
Max Pooling 2×2
Activation Functions ReLU, Sigmoid
Optimizer Adam
Epochs 20
Loss Function Binary Cross-Entropy

Table 1: Parameters used in the proposed Augmented CNN

3.4 Performance Matrices Used

The trained models with different metrics assess the performance of the proposed approach. The quality of
learning algorithms is usually evaluated by evaluating the performance of testing data. Different evaluation
matrices such as Accuracy, Specificity, Sensitivity / Recall [43], Precision, F1-score, and Area under the ROC
Curve (AUC) [44] can be used to evaluate the performance of the proposed work and are defined as follows.
Accuracy is the ratio of correct predictions to total predictions and is expressed in Equation (2)

Accuracy = (T P + T N )/(T P + F P + T N + F N ) (2)

where TP = True Positive, FP = False Positive, TN = True Negative, and FN = False Negative
Specificity is the ratio of true negatives to total negatives in the data and is expressed in Equation (3)

Specif icity = T N/(T N + F P ) (3)

Sensitivity / Recall is defined as the ratio of true positives to total positives in the data and is expressed in
Equation (4)
Sensitivity/Recall = T P/(T P + F N ) (4)
Precision is the ratio of true positives to total predicted positives and is expressed in Equation (5)

P recision = T P/(T P + F P ) (5)

F1-score depends on the precision and recall value, which is expressed in Equation (6)

F 1 − score = 2 ∗ (P recision ∗ Recall)/(P recision + Recall) (6)

The range of area within the ROC curve (AUC) is 0.5 to 1. Higher AUC values mean that the device is
performing well. It is expressed in Equation (7).
Z 1
AU C = ROC(t)dt (7)
0
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 137

4 Experimental Results and Discussion

The experimentation was carried out using Python libraries such as Keras and TensorFlow on a GPU NVIDIA-
MX450, an i5-1135G7 processor, with Windows as the operating system. Lung CT images were collected
from the LIDC - IDRI dataset and partitioned into 80:20 for learning and testing during the analysis procedure.
The training dataset included 1653 Lung CT images, while the testing dataset comprised 413 CT images. The
proposed method allocated a training subset of 1653 CT images with a 20% validation size. Of the 1653 CT
images, 1323 images train the method and 330 images evaluate the method for each epoch. Table 2 shows that
utilizing these performance parameters with default values is preferred. In addition, the batch size was extended
to 16. As a result, the CNN model performed admirably, with an accuracy of 95%. In addition to accuracy, the
other critical metrics examined in this work to measure the overall performance include precision, recall, and
F1 score. They are summarized in Table 2.

Recall Precision F1-score Test Data

Benign 0.96 0.93 0.95 203
Malignant 0.93 0.96 0.95 210

Overall Accuracy 0.95 413

Table 2: Performance of the proposed Augmented CNN with LIDC – IDRI dataset

Figure 6 a) and 6 b) depict the accuracy and loss curves between training and validation respectively. Vali-
dation accuracy is maximum at epoch 8 and minimum loss at epoch 20. From this, it is known that accuracy
increases with an increase in the epoch, and the loss decreases with an increase in the epoch.

(a) Accuracy Curve (b) Loss Curve

Figure 6: Training Progress for Augmented CNN

Figure 7 depicts the confusion matrix of the proposed work. Of the 210 malignant data, 196 are expected to
be positive (TP) and 14 are anticipated to be negative (FN), and out of 203 benign data sets, 195 are expected
to be positive (TN) and 8 as negative (FP). Benign test data has a recall, precision, and F1 score of 0.96, 0.93,
and 0.95, while malignant test data has a recall, precision, and F1 score of 0.93, 0.96, and 0.95 respectively.
The performance of the proposed work is compared with the existing works and the values of the perfor-
mance are given in Table 3. In the classification of lung cancer, all the existing CAD systems in Table 3 have
reasonable accuracy values. While classifying a CAD system, it is pivotal to take into account the small nodule
size and the size of the datasets. The proposed work produced better results since a smaller data dimension of
138 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

Figure 7: Confusion Matrix of the proposed Augmented CNN

Methods Architecture Accuracy (%)

S. Bhatia et al.[26] UNet and ResNet 84
Q. Z. Song et al. [19] CNN 84.15
S. R. Jena et al. [45] KNG-CNN 87.3
Proposed Work Data Augmented CNN 95

Table 3: Performance of the proposed Augmented CNN with LIDC – IDRI dataset

500 × 500 pixels, a batch size of 16, and a kernel size of 3 × 3 are used in the experimentation. With these
modifications in the architecture and along with the use of augmentation techniques, thereby increasing the size
of the datasets, the proposed work provides relatively high accuracy when compared with the existing work.

5 Conclusion
This paper work was carried out to demonstrate the efficacy and accuracy of lung cancer diagnosis using CNN
trained on Lung CT image datasets. Because of the promising results of deep learning methods, the LIDC-
IDRI dataset was used in this work to detect lung cancer using CT scan images. In this paper, the most current
research on deep learning techniques to early detect and diagnose lung cancer using CT images was analyzed.
Data augmentation techniques were used to preprocess the data set. Following augmentation, the dataset was
trained and tested using a CNN method, and the performance parameters were identified. An accuracy of 95%
is obtained in the proposed work.
The Capsule Neural Network (CapsNet), a new deep learning architecture can be deployed to overcome
CNN challenges, and it can be trained with a smaller number of data. Through more research on the topic, the
computer vision community hopes to construct robust machine vision algorithms by using the triumphs and
flaws of CapsNets. So future research would concentrate on CapsNet which could improve the performance
measures using a limited number of datasets.

Acknowledgment
The authors acknowledge the National Cancer Institute and the National Institutes of Health for their contribu-
tions to the development of the LIDC/IDRI database, which is free and open to the public. No funding agency
in the governmental, commercial, or not-for-profit sectors provided a particular grant for this research.
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 139

References
[1] R. L. Siegel, K. D. Miller, H. E. Fuchs, and A. Jemal, “Cancer Statistics, 2021,” CA. Cancer J. Clin., vol.
71, no. 1, pp. 7–33, 2021, doi: 10.3322/caac.21654.

[2] W. D. Travis et al., “International Association for the Study of Lung Cancer/ American Thoracic So-
ciety/European Respiratory Society: International multidisciplinary classification of lung adenocarci-
noma - An executive summary,” Proc. Am. Thorac. Soc., vol. 8, no. 5, pp. 381–385, 2011, doi:
10.1513/pats.201107-042ST.

[3] D. E. Midthun, “Early detection of lung cancer,” F1000Research, vol. 5, pp. 3–12, 2016, doi:
10.12688/f1000research.7313.1.

[4] L. Fass, “Imaging and cancer: A review,” Mol. Oncol., vol. 2, no. 2, pp. 115–152, 2008, doi:
10.1016/j.molonc.2008.04.001.

[5] K. Doi, “Computer-aided diagnosis in medical imaging: Historical review, current status and
future potential,” Comput. Med. Imaging Graph., vol. 31, no. 4–5, pp. 198–211, 2007, doi:
10.1016/j.compmedimag.2007.02.002.

[6] A. El-Baz et al., “Computer-aided diagnosis systems for lung cancer: Challenges and methodologies,” Int.
J. Biomed. Imaging, vol. 2013, 2013, doi: 10.1155/2013/942353.

[7] R. Afroze, M. Atikur, and M. Karam, “Detection of Lung Nodules using Image Processing Techniques,”
Int. J. Comput. Appl., vol. 177, no. 19, pp. 31–37, 2019, doi: 10.5120/ijca2019919638.

[8] F. Han et al., “Texture Feature Analysis for Computer-Aided Diagnosis on Pulmonary Nodules,” J. Digit.
Imaging, vol. 28, no. 1, pp. 99–115, 2015, doi: 10.1007/s10278-014-9718-8.

[9] W. J. Choi and T. S. Choi, “Automated pulmonary nodule detection based on three-dimensional shape-
based feature descriptor,” Comput. Methods Programs Biomed., vol. 113, no. 1, pp. 37–54, 2014, doi:
10.1016/j.cmpb.2013.08.015.

[10] A. Tartar, N. Kilic, and A. Akan, “Classification of pulmonary nodules by using hybrid features,” Comput.
Math. Methods Med., vol. 2013, 2013, doi: 10.1155/2013/148363.

[11] S. R. A. Ahmed, I. Al-Barazanchi, A. Mhana, and H. R. Abdulshaheed, “Lung cancer classification using
data mining and supervised learning algorithms on multi-dimensional data set,” Period. Eng. Nat. Sci., vol.
7, no. 2, pp. 438–447, 2019, doi: 10.21533/pen.v7i2.483.

[12] S. H. Hawkins et al., “Predicting outcomes of nonsmall cell lung cancer using CT image features,” IEEE
Access, vol. 2, no. March 2015, pp. 1418–1426, 2014, doi: 10.1109/ACCESS.2014.2373335.

[13] N. Camarlinghi et al., “Combination of computer-aided detection algorithms for automatic lung nodule
identification,” Int. J. Comput. Assist. Radiol. Surg., vol. 7, no. 3, pp. 455–464, 2012, doi: 10.1007/s11548-
011-0637-6.

[14] E. E. Nithila and S. S. Kumar, “Automatic detection of solitary pulmonary nodules using swarm intelli-
gence optimized neural networks on CT images,” Eng. Sci. Technol. an Int. J., vol. 20, no. 3, pp. 1192–1202,
2017, doi: 10.1016/j.jestch.2016.12.006.

[15] E. E. Nithila and S. S. Kumar, “Segmentation of lung nodule in CT data using active contour
model and Fuzzy C-mean clustering,” Alexandria Eng. J., vol. 55, no. 3, pp. 2583–2588, 2016, doi:
10.1016/j.aej.2016.06.002.
140 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022

[16] E. E. Nithila and S. S. Kumar, “Segmentation of lung from CT using various active contour models,”
Biomed. Signal Process. Control, vol. 47, pp. 57–62, 2019, doi: 10.1016/j.bspc.2018.08.008.

[17] S. Kamal, S. K. Mohammed, P. R. S. Pillai, and M. H. Supriya, “Deep learning architectures for underwa-
ter target recognition,” International Symposium on Ocean Electronics, SYMPOL, 2013, pp. 48–54, doi:
10.1109/sympol.2013.6701911.

[18] S. S. S. R. Jayadevan and S. S. Sannakki, “Recognition of Devanagari Scene Text Using Autoencoder
CNN,” Electron. Lett. Comput. Vis. Image Anal., vol. 20, no. 1, pp. 55–69, 2021.

[19] Q. Z. Song, L. Zhao, X. K. Luo, and X. C. Dou, “Using Deep Learning for Classification of Lung Nodules
on Computed Tomography Images,” J. Healthc. Eng., vol. 2017, 2017, doi: 10.1155/2017/8314740.

[20] M. K. Mahto, K. Bhatia, and R. K. Sharma, “Deep Learning Based Models for Offline Gurmukhi Hand-
written Character and Numeral Recognition,” Electron. Lett. Comput. Vis. Image Anal., vol. 20, no. 2, pp.
69–82, 2021, doi: 10.5565/rev/elcvia.1282.

[21] G. E. Krizhevsky, A.; Sutskever, I.; Hinton, “ImageNet Classification with Deep Convolutional Neural
Networks.,” Adv. Neural Inf. Process. Syst. 25 (NIPS 2012),Lake Tahoe, NV, USA, pp. 3–6, 2012, doi:
10.1201/9781420010749.

[22] M. Toğaçar, B. Ergen, and Z. Cömert, “Detection of lung cancer on chest CT images using minimum
redundancy maximum relevance feature selection method with convolutional neural networks,” Biocybern.
Biomed. Eng., vol. 40, no. 1, pp. 23–39, 2020, doi: 10.1016/j.bbe.2019.11.004.

[23] C. Szegedy et al., “Going deeper with convolutions,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recognit., vol. 07-12-June, pp. 1–9, 2015, doi: 10.1109/CVPR.2015.7298594.

[24] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” Lect. Notes Comput.
Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 9908 LNCS, pp. 630–645,
2016, doi: 10.1007/978-3-319-46493-0-38.

[25] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional net-
works,” Proc. - 30th IEEE Conf. Comput. Vis. Pattern Recognition, CVPR 2017, vol. 2017-Janua, pp.
2261–2269, 2017, doi: 10.1109/CVPR.2017.243.

[26] S. Bhatia, Y. Sinha, and L. Goel, “Lung cancer detection: A deep learning approach,” Adv. Intell. Syst.
Comput., vol. 817, pp. 699–705, 2019, doi: 10.1007/978-981-13-1595-4-55.

[27] M. Kriegsmann et al., “Deep learning for the classification of small-cell and non-small-cell lung cancer,”
Cancers (Basel)., vol. 12, no. 6, pp. 1–15, 2020, doi: 10.3390/cancers12061604.

[28] Q. Dou, H. Chen, L. Yu, J. Qin, and P. A. Heng, “Multilevel Contextual 3-D CNNs for False Positive
Reduction in Pulmonary Nodule Detection,” IEEE Trans. Biomed. Eng., vol. 64, no. 7, pp. 1558–1567,
2017, doi: 10.1109/TBME.2016.2613502.

[29] N. Coudray et al., “Classification and mutation prediction from non–small cell lung cancer histopathology
images using deep learning,” Nat. Med., vol. 24, no. 10, pp. 1559–1567, 2018, doi: 10.1038/s41591-018-
0177-5.

[30] W. J. Sori, J. Feng, A. W. Godana, S. Liu, and D. J. Gelmecha, “DFD-Net: lung cancer detection from de-
noised CT scan image using deep learning,” Front. Comput. Sci., vol. 15, no. 2, 2021, doi: 10.1007/s11704-
020-9050-z.
 Bushara A. R. et al. / Electronic Letters on Computer Vision and Image Analysis 21(1):130-142, 2022 141

[31] P. M. Shakeel, M. A. Burhanuddin, and M. I. Desa, “Lung cancer detection from CT image using improved
profuse clustering and deep learning instantaneously trained neural networks,” Meas. J. Int. Meas. Confed.,
vol. 145, pp. 702–712, 2019, doi: 10.1016/j.measurement.2019.05.027.

[32] E. S. Neal Joshua, D. Bhattacharyya, M. Chakkravarthy, and Y. C. Byun, “3D CNN with Visual Insights
for Early Detection of Lung Cancer Using Gradient-Weighted Class Activation,” J. Healthc. Eng., vol.
2021, 2021, doi: 10.1155/2021/6695518.

[33] M. A. Heuvelmans et al., “Lung cancer prediction by Deep Learning to identify benign lung nodules,”
Lung Cancer, vol. 154, no. January, pp. 1–4, 2021, doi: 10.1016/j.lungcan.2021.01.027.

[34] M. Masud, N. Sikder, A. Al Nahid, A. K. Bairagi, and M. A. Alzain, “A machine learning approach to
diagnosing lung and colon cancer using a deep learning-based classification framework,” Sensors (Switzer-
land), vol. 21, no. 3, pp. 1–21, 2021, doi: 10.3390/s21030748.

[35] T. L. Chaunzwa et al., “Deep learning classification of lung cancer histology using CT images,” Sci. Rep.,
vol. 11, no. 1, pp. 1–12, 2021, doi: 10.1038/s41598-021-84630-x.

[36] O. Ozdemir, R. L. Russell, and A. A. Berlin, “A 3D Probabilistic Deep Learning System for Detection
and Diagnosis of Lung Cancer Using Low-Dose CT Scans,” IEEE Trans. Med. Imaging, vol. 39, no. 5, pp.
1419–1429, 2020, doi: 10.1109/TMI.2019.2947595.

[37] S. G. Armato et al., “The Lung Image Database Consortium (LIDC) and Image Database Resource Initia-
tive (IDRI): A completed reference database of lung nodules on CT scans,” Med. Phys., vol. 38, no. 2, pp.
915–931, 2011, doi: 10.1118/1.3528204.

[38] S. Khan, H. Rahmani, S. A. A. Shah, and M. Bennamoun, “A Guide to Convolutional Neural

Networks for Computer Vision,” Synth. Lect. Comput. Vis., vol. 8, no. 1, pp. 1–207, 2018, doi:
10.2200/s00822ed1v01y201712cov015.

[39] W. Alakwaa, M. Nassef, and A. Badr, “Lung cancer detection and classification with 3D convolu-
tional neural network (3D-CNN),” Int. J. Adv. Comput. Sci. Appl., vol. 8, no. 8, pp. 409–417, 2017, doi:
10.14569/ijacsa.2017.080853.

[40] H. Eedi*, “A Deep Convolutional Neural Network Framework for Lung Cancer Detection,” Int. J. Recent
Technol. Eng., vol. 8, no. 4, pp. 10105–10109, 2019, doi: 10.35940/ijrte.d9764.118419.

[41] A. Patil and M. Rane, “Convolutional Neural Networks: An Overview and Its Applications in Pattern
Recognition,” Smart Innov. Syst. Technol., vol. 195, pp. 21–30, 2021, doi: 10.1007/978-981-15-7078-0-3.

[42] X. Kang, B. Song, and F. Sun, “A deep similarity metric method based on incomplete data for traffic
anomaly detection in IoT,” Appl. Sci., vol. 9, no. 1, 2019, doi: 10.3390/app9010135.

[43] A. Baratloo, M. Hosseini, A. Negida, and G. El Ashal, “Part 1: Simple Definition and Calculation of Accu-
racy, Sensitivity and Specificity.,” Emerg. (Tehran, Iran), vol. 3, no. 2, pp. 48–9, 2015, doi: 10.22037/emer-
gency.v3i2.8154.

[44] W. Zhu, N. Zeng, and N. Wang, “Sensitivity, specificity, accuracy, associated confidence interval and ROC
analysis with practical SAS® implementations.,” Northeast SAS Users Gr. 2010 Heal. Care Life Sci., pp.
1–9, 2010.

[45] S. R. Jena and S. T. George, “Morphological feature extraction and KNG-CNN classification of CT images
for early lung cancer detection,” Int. J. Imaging Syst. Technol., vol. 30, no. 4, pp. 1324–1336, 2020, doi:
10.1002/ima.22445.