IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 3, September 2023, pp. 1044~1061

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i3.pp1044-1061  1044

A review of convolutional neural network-based computer-

aided lung nodule detection system

Sekar Sari1, Tole Sutikno2, Indah Soesanti1, Noor Akhmad Setiawan1

1
Department of Electrical and Information Engineering, Faculty of Engineering, Universitas Gadjah Mada, Yogyakarta, Indonesia
2
Department of Electrical Engineering, Faculty of Industrial Technology, Universitas Ahmad Dahlan, Yogyakarta, Indonesia

Article Info ABSTRACT

Article history: Worldwide, lung cancer is the major cause of death and rapidly spreads. Lung
tissue that is benign does not grow significantly, but lung tissue that is
Received Nov 14, 2022 malignant grows rapidly and attacks the body, posing a grave threat to one's
Revised Jan 10, 2023 health. This paper provides a literature review of computer-aided detection
Accepted Jan 30, 2023 (CAD) systems for lung cancer diagnosis. Preprocessing, segmentation,
detection, and classification are the stages of the CAD system. This review
divides the preprocessing into three stages: image smoothing, edge
Keywords: sharpening, and noise removal. Additionally, lung segmentation is divided
into three stages: histogram-based thresholding, linked component analysis,
Computer-aided detection and lung extraction. The detecting phase aids in decreasing the workload.
Convolutional neural network Several techniques are briefly described, including random forest, naive
False-positive reduction bayes, k-nearest neighbor (k-NN), support vector machine (SVM), and
Feature extraction convolutional neural network (CNN). Classification is the final stage; the
Lung cancer detection image is then identified as containing or not possessing nodules. The prospect
Lung nodule detection of incorporating CNN-based deep learning techniques into the CAD system
is discussed. This paper is superior to other review studies on this topic due to
its comprehensive examination of pertinent literature and structured
presentation. We hope that our research may help professional researchers and
radiologists design more effective CAD systems for lung cancer detection.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Tole Sutikno
Department of Electrical Engineering, Faculty of Industrial Technology, Universitas Ahmad Dahlan
South Ringroad st., Tamanan, Banguntapan, Bantul, Special Region of Yogyakarta, Indonesia 55191
Email: [email protected]

1. INTRODUCTION
Lung cancer is characterized by uncontrolled cell division in the lungs. Cells divide and replicate
themselves. When damaged cells divide uncontrollably, they become tumors, which eventually hinder organs
from functioning correctly. When compared to other kinds of cancer, lung cancer is by far the deadliest and
most common form of the disease, making it the top cause of death globally [1]–[6]. Lung cancer is the world's
second most common cancer, behind breast cancer. According to World Health Organization (WHO) figures
issued in 2020, lung cancer mortality in Indonesia reached 28,633 in 2019. In 2022, the United States is
expected to have 1,918,030 new cancer cases and 609,360 cancer deaths, with lung cancer, the leading cause
of cancer death, accounting for approximately 350 deaths per day [7]. However, if nodules are found at an
early stage, there is a chance that the human survival rate can be raised.
Lung cancer detection methods using computer-aided detection (CAD) have been developed in recent
years [8]–[16]. Early identification of lung cancer has been shown to both lower mortality rates and increase
the likelihood of survival. Nodules in the lungs that are still relatively tiny can be benign or malignant,
depending on the circumstances, even if the nodules themselves are not cancerous. Lung tissue that is benign

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  1045

does not experience much growth, however lung tissue that is malignant will develop swiftly and assault the
body, making it extremely hazardous to one's health [17]. The most important goals of computer-aided design
(CAD) are for reliably recognizing images and extracting regions of interest (ROI) from images obtained from
a variety of imaging modalities. These imaging modalities include computed tomography (CT) scans, X-rays,
position emission tomography (PET), and magnetic resonance imaging (MRI) [18]–[27]. CAD systems are
further subdivided into computer-aided detection (commonly abbreviated as CADe) and computer-aided
diagnosis (CADd) (CADx). The CADe system's capabilities are limited to identifying abnormal tissue regions
and pictures, but the CADx system can be used to diagnose a disease by determining the sort of abnormality
present and whether it is malignant [28].
Previously, a substantial amount of study was conducted on the examination of lung nodule
identification using the CADe system. Because of limited resources and enormous volumes of data, this system
is often based on traditional machine learning, as well as linear discriminant analysis, multiple gray-level
thresholding, distance transformation, and support vector machines (SVM) [29]–[33]. However, in recent
years, a number of researchers have created deep learning-based lung detection techniques. One of these ways
is the convolutional neural network (CNN) approach, It has exceptional computer vision performance values,
and increases the CADe system's accuracy and sensitivity [34]. Trends for the years 2019-2022 include deep
learning in the form of CNN as well as the performance of each individual approach. The objective of this
study is to give researchers with an overview of the CNN-based CAD system for the identification of lung
nodules. The review will cover the phases of the CAD system for lung nodule identification in general, the
preprocessing, the segmentation and detection techniques that have been extensively utilized from the
beginning to the present, as well as the approach of lung nodule identification that is now being utilized.

2. LUNG NODULES
Lung nodules are abnormal growths that develop within the lungs. Lung nodules are extremely
frequent. The lung may have one or many nodules. Up to fifty percent of persons who have chest X-rays or CT
scans possess them. Nodules can form in either lung. About 95% of lung nodules are benign (not cancerous).
Infrequently, lung nodules indicate lung cancer. Because small lung nodules rarely cause symptoms, more
testing are required to identify whether or not it is lung cancer.
Analysis of lung nodules is one of the stages that must be taken in order to successfully prevent lung
cancer, which is accomplished through identification and categorization. Dark-level lung nodules are typically
between 3 and 30 millimeters in diameter [34]. In general, these nodules have a diameter of around 3
millimeters. Figure 1 depicts samples of various types of lung nodule categories. The circumscribed, juxta
vascular, juxta pleural, and pleural tails of nodules are represented in Figures 1(a) to 1(d), respectively. In
contrast, juxta vascular nodules are firmly linked to blood vessels, while juxta pleural nodules are located in
the region around the pleura [28]. Circumscribed nodules are not associated with any other tissue structures
and are seen in a dispersed manner throughout the tissue.

The nodule is centrally While placed centrally in the A substantial amount of the The nodule is located at the
positioned in the lung and has lung, the nodule has nodule is connected to the pleural surface and is
no link to the vascular. extensive linkages to pleural surface. connected by a thin surface.
adjacent arteries.
(a) (b) (c) (d)

Figure 1. Categories of lung nodule samples [28], (a) well-circumscribed (b) juxta vascular, (c) juxta pleural,
and (d) juxta pleural tail

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1046  ISSN: 2252-8938

3. COMPUTER-AIDED DESIGN FOR LUNG NODULE DETECTION

There are five stages to the computer-aided design (CAD) for the lung nodule detecting system, which
are as follows: i) image acquisition, ii) preprocessing, iii) lung segmentation, iv) nodule detection, and
v) classification. A comprehensive schematic representation of the lung CAD process is shown in Figure 2.
Medical images are acquired by a variety of imaging modalities, including CT scans, X-rays, and MRI images
[35]. Medical images can be gathered from publically available image databases to give researchers with a
source of data for CADe system research, development, testing, evaluation, and benchmarking. Table 1 shows
several databases of lung scans that are open to the public and can be used by anyone. These include Japanese
society of radiological technology (JSRT), early lung cancer action program (ELCAP),

Figure 2. Acquisition, preprocessing, lung segmentation, nodule detection, and classification are the steps in
the CAD schematic for lung nodule detection

Table 1. List of public lung image databases

Public Databases Released Year Scans Modalities
PLCO Trial 1993 52,320 CXR
JSRT [36] 1998 247 CXR
ELCAP [37] 2003 50 CT
NELSON [38] 2003 15,822 CT
NELSON [39] 2003 15,822 CT
RIDER [40] 2004 48,698 CT
ANODE09 [41] 2009 55 CT
NLST Trial [42] 2009 75,000+ CT
Lung TIME [43] 2009 157 CT
LIDC–IDRI [1] 2011 1,018 CT, CR, DX
ILD [44] 2012 905 CT
ACRIN-NSCLC-FDG-PET [45] 2013 3,377 PT, CT, MR, CR, DX, SC, NM
NSCLC-Radiomics [46] 2014 1,265 CT, RTSTRUCT, SEG
LungCT-Diagnosis [47] 2015 61 CT
QIN Lung CT [48] 2015 47 CT
LISS [49] 2015 271 CT
LUNA 16/Ali Tianchi [50] 2016 888 CT
Italung-CT [51] 2016 122 CT
Kaggle Data Science Bowl [17] 2017 285,380 CT
UniToChest [52] 2021 306,440 CT
Note: JSRT: Japanese Society of Radiological Cancer Screening Trial
CXR: Chest X-Ray Technology PLCO: Prostate, Lung,
CT: Computed Tomography ELCAP: Early Lung Cancer Action Project Colorectal and
CR: Computed Radiography NELSON: Nederlands Leuvens Ovarian
DX: Digital X-ray Longkanker Screeningsonderzoek NLST: National Lung
PT: Positron emission tomography (PET) ANODE: Automatic Nodule Detection Screening Trial
MR: Magnetic Resonance LIDC: Lung Image Database Consortium NELSON:
SC: Secondary Capture IDRI: Image Database Resource Initiative Nederlands-Leuvens
NM: Nuclear Medicine ILD: Intersitial Lung Disease Longkanker
RTSTRUCT: Radiotherapy Structure Set LUNA: Lung Nodule Analysis Screenings Onderzoe
SEG: Segmentation Italung-CT: Italian Lung

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1047

Dutch-Belgian randomized lung cancer screening trial or in Dutch Nederlands–Leuvens Longkanker

Screenings Onderzoek (NELSON), automatic nodule detection 2009 (ANODE09), lung image database
consortium and image database resource initiative (LIDC-IDRI), lung nodule analysis 2016 (LUNA16),
Kaggle, and others. The lung ROI is segmented and noise is reduced during the preprocessing phase of the
lung nodule CAD system. As a result, the prospecting range for lung nodules has been narrowed and the data
have been normalized [10]. These include the weiner filter, the grayscale filter, the histogram equalization
filter, the linear mapping filter, and the Gaussian lowpass filter [53], [54].
Lung segmentation is the process of separating a nodule from surrounding areas of a CT scan and then
enhancing the resulting image to obtain more information about the nodule [55]. Accurate lung segmentation
and a range of methods for obtaining lung volume from CT images are two indicators that reflect the efficiency
of a system designed to identify lung nodules. Among them are quality thresholding, morphological
procedures, the watershed approach, the active shape model (ASM), the active appearance model (AAM), and
the faster region CNN (faster R-CNN) [50], [56].
The process of finding nodules in the lung that may progress to lung cancer is known as nodule
detection [28]. Naive bayes, random forests, SVM’s, multi-layer perception (MLP), and CNN’s are some of
the approaches used to identify lung nodules [1], [41], [57]. Because high sensitivity and poor accuracy values
are frequently obtained during the nodule identification stage, a subsequent phase, termed false-positive
reduction, is required [34]. The feature extraction method and nodule categorization with feature-based
classifiers were engaged at this stage [41]. The form and texture are used to extract features. Shape qualities
are measured using the geometric value of each structure (such as form proportion, density, roundness,
elongation, weighted radial distance, and Boyce-Clark radial shape index); nodules are rounder than other
tissue structures, thus look for the most spherical shape [58]. After the feature extraction step, use many
supervised or unsupervised classifiers to detect the nodule and lower the FPs value [28]. In order to reduce the
CADe system's false-positive rate, this FPs reduction stage focuses on recognizing true lung nodules from all
suspected nodules and removing phony nodules. There will be four possible outcomes throughout the
classifying procedure. A lung nodule that is correctly identified is referred to as a true positive (TP) or false
negative (FN). If a lung nodule is identified wrongly, it is referred as as a "true negative" (TN) or "false
positive" (FP) [59].

4. PREPROCESSING
Preprocessing is an essential first step in the CT images used for lung detection because the raw CT
images contain a great deal of noise and extraneous data that will hinder the CAD system's ability to detect
lung nodules [34]. This review divides preprocessing into three stages: image smoothing, edge sharpening, and
noise removal. The preprocessing stages will be explained briefly in this section, and the section will conclude
with the possibility of using CNN-based methods to remove noise from lung images.

4.1. Image smoothing

Images are impacted by a number of factors, not all of which are related to the viewer's ability to
visually perceive the image. In addition to this, they make it more difficult to recognize and differentiate
between image features that are important for various applications, such as pattern recognition and image
segmentation. Noise is one of these factors that occurs quite frequently, and it has the potential to have a
significant impact on both the visual quality of images and the performance of the majority of image processing
tasks. It is due to inaccuracies that occurred during the image acquisition process [60].
Images are frequently captured despite the fact that the conditions are not optimal; for example, there
may be insufficient light, excessive clarity, or unfavorable weather. The inability to acquire an image due to
transmission errors, problems with networked cables, signal disturbances, problems with sensors, and so on
can be caused by equipment of inferior quality. Therefore, the pixel intensity values do not accurately reflect
the true colors that were captured during the real acquisition process. Because of these factors, a wide variety
of techniques have been developed in order to retrieve lost image information and improve the details of
images. Image smoothing is a technique that is included in preprocessing techniques and is used to remove
possible image perturbations without losing any of the image's information [60]. Image smoothing is used to
define smooth, consistent borders for segmented lungs near the mediastinum. As a result, smoothing operations
must be restricted to the region around the mediastinum so that the contour of the lung in areas other than the
mediastinum is not affected [61].
Some smoothing methods include restricted cubic splines (RCS), penalized splines (P-splines), natural
splines (NS), and fractional polynomials (Fracpoly). The RCS is a cubic regression spline with continuous first
and second derivatives at the knots for visual smoothness. RCS are further limited to be linear above the last
knot and below the first. The linearity in the tails allows for a more compact model. P-splines were fitted using
R's standard software implementation. By modeling the smooth function, P-splines provide an approach to
A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1048  ISSN: 2252-8938

determining optimal smoothing via degrees of freedom that is relatively robust to the choice of location and a
relatively large number of knots. The natural spline is simply a constrained cubic spline that employs B-spline
basis functions instead of piecewise polynomials. Fractional polynomials, like P-splines and RCS, can be
employed with any generalized linear model to analyze survival data. Although a global (rather than local)
approach, the FP model has the advantage of being a simpler form than the other two possibilities and
incorporating a wider range of functional forms than the normal polynomial family allows. A broader range of
possible dose-response interactions could be accommodated [62], [63]. The mean filter, the Gaussian filter,
anisotropic diffusion, the median filter, the adaptive median filter, conservative smoothing, and the alpha trim
mean filter are other image smoothing methods that can be used to improve low-level distortion in lung images
[37], [53]–[55], [58], [64].

4.2. Edges sharpening

Depending on the image domain, the sharpening methods can be divided into two groups: spatial-
based and frequency-based methods. In the first scenario, we work directly with the pixel, whereas in the
second, we work with the image's transform coefficients (Fourier or wavelet). Only when we recover the image
using the inverse transform can we notice the effect of the alteration. The alteration of pixel values is the
foundation of spatial domain techniques for sharpening images. Enhancing the contrast between various
elements of the image is one method to make it better. There are various techniques for sharpening images in
the spatial domain. Histogram equalization is one of the most well-known techniques (HE). Contrast stretching
(CS), which is based on altering the dynamic range, or the range between the minimum and maximum intensity
values of the image's gray levels, is another well-known method within the field of spatial domain sharpening.
The simplest contrast stretch algorithm, linear contrast stretch (LCS), stretches the pixel values of a low- or
high-contrast image by stretching the dynamic range across the entire image spectrum. This method's loss of
certain details due to saturation and clipping is one of its drawbacks [1], [53], [55], [60], [65].
Frequency domain approaches rely on transformations such as the discrete Fourier (or cosine)
transform or wavelet transforms. Each of these strategies is not unique; in fact, they are part of a family of
methods that are fundamentally the same yet varies slightly from one another. They operate as follows: First,
we apply one of these transformation methods; then, we process the changed image using one of these methods;
and finally, the inverse transformation of the processed image yields the output. This method has a significant
advantage: the ability to discern various parts in an image. Higher frequencies correspond to image edges or
features, while lower frequencies belong to image smoothness. This simple split enables the image to be
processed correctly depending on the aim. However, this also implies that we are processing details from
multiple locations indistinguishably at the same time. This also occurs in smooth areas. In recent years, wavelet
theory has emerged as a powerful image processing tool. This approach offers us with image spatial and
frequency information. Adding high-pass or eliminating low-pass filtered versions from the image might
improve the image. One of the early works on contrast sharpening in the wavelet domain was applying a
parametrized hyperbolic function to the gradient of the wavelet coefficients. Since then, a great deal of work
has been done in the wavelet domain. Loza et al., for example, suggested a non-linear augmentation strategy
based on the local dispersion of wavelet coefficients. This technique adaptively improves image contrast
depending on local characteristics of the image's wavelet coefficients. A contrast enhancement technique based
on scaling the internal noise of a dark image in the discrete cosine transform (DCT) domain. It is based on a
physics concept known as "dynamic stochastic resonance" (DSR), which employs noise to increase the
performance of a system [60].
The traits of smoothing and sharpening are diametrically opposite. As a result, there are few solutions
to meet both goals concurrently, either jointly or independently. The first strategy is to process the image in
two steps: first by performing one operation, and then by applying the second procedure on the produced image.
The order in which we do the operations has a considerable influence on the result in this situation. Sharpening
before smoothing may increase the significance of image noise, complicating the smoothing process.
Smoothing before sharpening, on the other hand, risks losing information that the sharpening technique cannot
recover. Although the second technique yields better outcomes in general, it is not the best option. As a result,
in recent years, techniques capable of combining smoothing and sharpness have been presented [60]. The
methods of smoothing and sharpening known as contrast limited adaptive histogram equalization (CLAHE)
[66] and anisotropic filtering introduced by Perona and Malik (PM) [67] have been combined simultaneously
by means of a synchronization algorithm, and the improvement in comparison to the corresponding two-step
methods that are based on them can be seen here. The technique makes use of the benefits offered by these
distinct models and combines them with the intention of developing a powerful instrument for the creation of
medical images, more specifically for lung images.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1049

4.3. Noise removing

Image denoising processes eliminate noise from images and restore them to their original state. How
to discern between noise, edge, and texture is a fundamental issue in image denoising. Image denoising
techniques are utilized in medical imaging, remote sensing, military surveillance, biometrics and forensics,
industrial and agricultural automation, and individual recognition. Denoising algorithms are important pre-
processing processes in medical and biomedical imaging that are used to remove medical noise such as speckle,
Rician, quantum, and others [68].
The mean filter (MF), the adaptive mean filter (AMF), and the bilateral filter are three examples of
the many algorithms that can be used to remove or reduce noise (BF). The mean filter (MF) is the simplest of
the three, the AMF is a refinement of the mean filter, and the background subtraction filter (BF) is considered
a state-of-the-art noise reduction technique. The BF compresses the filter range as well as the filter domain into
a particular window size. The BF is a non-iterative adaptive smoothing filter that reduces noise while
maintaining the edges of the objects contained within an image. The BF is comprised of three variables, which
are as follows: the window size, the filter range, and the filter domain. Because of this, the camera is able to
preserve a significant amount of the image's fine detail as well as its texture. However, the BF technique
requires a considerable amount of computation and is mathematically complex; consequently, the development
of a simple and quick algorithm for noise reduction that maintains excellent image quality would be very
significant [69].
The majority of the filters listed above have generated results that are considered to be of a satisfactory
quality; nonetheless, each of these filters has a few deficiencies. These disadvantages include insufficient test
phase optimization, the need for manual parameter settings, and particular denoising models. The adaptability
of CNN’s has, fortunately, demonstrated that it is possible to overcome these shortcomings [70]–[73]. An
overview of CNN's image denoising procedures is presented in Figure 3. It is intended that the explanations
contained within this study will make it possible to comprehend the CNN architectures that are utilized in
image denoising [68].

Figure 3. CNN-based methods for removing noise for lung images [68]

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1050  ISSN: 2252-8938

5. LUNG SEGMENTATION
Lung segmentation is a procedure that is used in lung detection. This process involves isolating lung
nodules from other portions of the lung CT scan image and then further improving the image that is produced
so that detail can be seen [55]. Lung segmentation techniques are broadly classified into four categories:
i) deformable boundary-based techniques; ii) edge-based techniques; and iii) threshold-based techniques, and
iv) registration-based method [28], [74]–[77]. Each of these categories has its own subcategories. Several of
these methods of segmentation each have their own set of benefits and drawbacks. This review divides lung
segmentation into three stages: histogram-based thresholding, connected component analysis, and lung
extraction. The stages will be explained briefly in this section, and the section will conclude with the possibility
of using CNN-based lung segmentation methods.

5.1. Histogram based thresholding

Because of its ease of use, the global histogram of a digital image is a popular tool for real-time image
processing. It provides an important foundation for statistical techniques in image processing by giving a global
description of the image's information. A pixel's color in a color image using RGB representation is a blend of
the three primitive hues red, green, and blue. Each image pixel may be seen as a three-dimensional vector with
three components representing the three colors of each image pixel. As a result, the global histograms
representing the three primitive components may yield global information about the entire image. Histogram-
based thresholding is a common segmentation approach that looks for peaks and troughs in the histogram. A
standard segmentation strategy based on histogram analysis can only be carried out if the dominant peaks in
the histogram can be accurately identified. Several frequently used peak-finding methods assessed the
sharpness or area of the peak to determine the dominant peaks in the histogram. Although these peak-finding
techniques are beneficial in histogram analysis, they may not always function effectively, especially when the
image contains noise or radical change [78], [79].
In the research carried out by Filho et al. [58], the researchers decided to utilize a threshold value of
90 since they found that it was effective in the identification of lung nodules. The use of threshold approaches
in lung segmentation helps make it simpler to locate nodules and reduces the amount of lung tissue that does
not contain a nodule [1]. In the process of lung segmentation, in addition to thresholding techniques, there are
also watershed techniques [54], morphological operations [65], active contour modeling (ACM) [56], active
appearance modeling (AAM) [53], active shape modeling (ASM) [36], and fissure region segmentation [37].
Widodo et al. [53] came up with the idea for the AAM methodology, which is a statistical learning method that
models parameters to characterize form differences across classes and variation in texture. In order to carry out
the covariance matrix eigen analysis of the training vector in a way that is consistent with the shape and texture
of the training image, the principal component analysis (PCA) is utilized. The AAM model as a whole is
composed of three subcomponents, which are as follows: i) alignment of the shape data; ii) creation of a
parameter model based on statistical and real data; and iii) template matching. The findings that were acquired
from this segmentation were able to distinguish the lungs from the other tissues that were adjacent [53]. An
active contour model was integrated with the field formulation of the locally biased image by using the ACM
approach that was suggested by Kasinathan et al. [56] for the segmentation of lung tumors (ACM). In order to
reconcile properly homogenous CT images and quickly split the tumor zone with an inhomogeneous intensity,
the mean square error was utilized. The LIDC-IDRI database was utilized to test the suggested ACM approach,
which is comprised of 850 images of the lesion and is able to reliably locate lung tumors in CT scans. The
results of the testing were positive.

5.2. Connected component analysis (CCA)

Connected component analysis (CCA) also known as connected component labeling, blob extraction,
or region labeling is a graph theory-based approach for determining the connectedness of "blob"-like regions
in a binary image. Connected component analysis is frequently utilized in the same contexts as contours are
used; however, connected component labeling can often provide more granular filtering of blobs in a binary
image. The outline hierarchy is commonly a constraint when employing contour analysis (i.e., one contour
contained within another). We can more readily segment and examine these structures with connected
component analysis. The primary purpose of CCA is to extract related components in a binary image and
synthetic data, such as area, bounding boxes, center of gravity, and so on, which will then be further processed
based on the application. CCA is traditionally implemented as a mixture of two following computations:
connected component labeling (CCL) and feature computation (FC). The CCL procedure distinguishes various
connected components in the input binary image by assigning a unique label to all pixels that belong to the
same connected component. The FC algorithm then creates these labeled images in order to obtain one or more
of the above-mentioned synthetic data parameters required by the subsequent phases [80].

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1051

The labeled connected component in the binarized image is analyzed using connected component
analysis. The pixels in the binary image are divided into multiple linked components based on their pixel
connection. The lung lobes are segmented using the first-level linked components. The designated components'
area and boundary information are utilized to determine the choice. The second level of linked components is
identified, and geometric characteristics are retrieved from each component. The geometric characteristics used
include area, bounding box, eccentricity, equivalent diameter, main axis length, minor axis length, and
perimeter. The information gathered from the components is evaluated to identify whether or not they are
tumors. An overall evaluation of lung nodules is provided in [81]. A lung nodule greater than 5 mm in size has
a significant likelihood of becoming malignant cancer. According to their findings, an eccentric component
with an overall size of 5 mm or greater is likely to be malignant. The findings of these tests are utilized to label
the section as malignant and remove the remainder [82].

5.3. Lung extraction

The purpose of lung extraction is to identify the thoracic wall and mediastinum voxels, allowing the
subsequent stages of work to focus entirely on the region that forms the lung parenchyma. This is accomplished
once more through the use of the region-growing algorithm. Occasionally, the lung extraction stage incorrectly
removes certain voxels from the pulmonary parenchyma. These errors can result in the exclusion of all potential
nodules, resulting in a detection error. As a result, the rebuilding stage is critical for the preservation of
peripheral nodules. In order to perform the reconstruction of the incorrectly eliminated lung outline, a previous
knowledge about the object which is being segmented is used. The lung is a well-known organ with a gentle
contour and no re-entrances. As a result, any hole or abrupt discontinuity observed on its outline is a clear
indication of perimeter collapse and must be repaired [83].
Kuruvilla and Gunavathi [65] proposed a method for lung segmentation that makes use of
morphological procedures on CT images. These operations are carried out by transforming grayscale images
to binary images. The speed and user-friendliness of the morphological operation approach are two of its
defining characteristics. Although Jayaraj and Sathiamoorthy [54] were the ones who initially introduced the
watershed segmentation approaches, its most distinguishing feature is the capacity to isolate and identify items
that are in close proximity to the image. This paradigm of mathematical morphology is built on the concept of
regions. It is a kind of apparent image decomposition that assigns each pixel to a region or watershed. Deep
learning, namely the faster convolutional neural network, was proposed by Huang et al. [50] as a method for
performing the segmentation procedure for identifying lung CT images. It employs five layers of convolution
at a rate that is lower than usual because to the improved resolution and increased segmentation precision it
possesses. The suggested fully convolutional network (FCN) method was tested on the LIDC-IDRI database,
and the results showed that it had an accuracy value of 94.6%. When separating CT images from the LIDC-
IDRI database in 2018, Chunran and Yuanyuan [84] employed a FCN.
The regression neural network (RNN) segmentation approach was proposed by Messay et al. [85].
They developed a system that is completely automated (FA), a system that is semi-automated (SA), and a
system that is hybrid. The FA and SA systems are what give rise to the hybrid system. These systems then
yield many parameters, which are subsequently decided in an adaptive manner for each nodule by use of RNN.
Additionally, the RNN method was presented by Sankar and George [86] in the year 2020. They are employing
RNNs in an effort to enhance the identification of juxtapleural and juxtavascular lesions. Their suggested RNN
technique performs better than the skeleton graph cut method and the level set method, which are both used to
recognize lesions with the same intensity level. UNet and CNN are two approaches that were suggested by
Shaziya and Shyamala [87] for segmenting lung CT images. According to the results of the dice similarity
coefficient (DSC), the UNet approach is 1.27% more effective than CNN when it comes to image segmentation.
In addition, Arora et al. [88] completed segmentation using a total of 662 chest X-ray (CXR) images using the
UNet approach. When they segmented the lungs of TB patients, they found that the DSC value was 0.9680.
In point of fact, performance segmentation based on a rules-based approach is exactly the same as
performance segmentation based on a data-based approach. Nevertheless, in order to train the learning model,
a data-based method takes a significant amount of time, and the associated computational costs will be higher
than those associated with a rule-based approach to CAD system optimization. In order to make researchers
feel more at ease when using a rule-based approach to the processing of lung CT images, a rule-based approach
can be accomplished by altering the manual settings of a data-based method [34]. As a consequence of this,
the techniques of thresholding, FCN, RNN, and UNet are the most effective ways of segmentation for
performing image detection tasks. For more details, the lung nodule preprocessing and segmentation techniques
are shown in Table 2.

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1052  ISSN: 2252-8938

Table 2. Techniques for lung preprocessing and segmentation

Authors Year Databases Scans Nodules size Preprocessing Methods Segmentation methods
Kuruvilla and 2014 LIDC-IDRI 155 260-400 mm Gray scale to binary image Morphological
Gunavathi [65] operations
De Carvalho Filho 2014 LIDC-IDRI 640 3 mm Contrast, gaussian filter Quality threshold
et al. [58] and median filter clustering and region
growing
Messay et al. [85] 2015 LIDC-IDRI 456 4.21-31.62 mm A semi-automated (SA) RNN
system, a fully-automated
(FA) system, and a hybrid
system
Chen et al. [89] 2015 LIDC-IDRI 1010 8 mm Resizing and Z-score CNN and deep belief
normalization using mean networks (DBN)
and standard deviation
Widodo et al. [53] 2017 Private 120 0.5-10 mm Histogram equalization and AAM
gaussian lowpass filter
Jiang et al. [1] 2018 LIDC-IDRI 1006 3 mm Brighnest, position, shape, Local threshold
and repair of the lung segmentation
contour for juxta-pleural
nodules.
Chunran and 2018 LIDC-IDRI 1010 3-30 mm Original images FCN
Yuanyuan [84]
Gong et al. [41] 2018 LUNA16 and 1079 0.5–2 mm OTSU thresholding Segmentation of 3D
ANODE09 levels and local image
characteristics

Ausawalaithong et 2018 JSRT, NIH 247 2048×2048 and Increasing contrast, noise CNN
al. [90] and 1024×1024 removal, image resizing
100 pixels and image normalizing
Anitha and Babu 2019 LIDC and 50 and 0.625 mm Morphological Fissure regions
[37] ELCAP 30 transformation and weiner segmentation
filter
Kasinathan et al. 2019 LIDC-IDRI 850 0.45-0.75 mm Remove the mediastinum ACM
[56] region and thoracic wall
Huang et al. [50] 2019 LIDC-IDRI 888 0.6-5.0 mm Linear mapping FCN
and LUNA16
Ardila et al. [91] 2019 LIDC, LUNA, 1139 8–15 mm CNN CNN
and NLST

Li et al. [36] 2020 JSRT 247 17.3 mm Rib suppression CNN

Sankar and 2020 LIDC-IDRI 1018 3 mm Gaussian filtering RNN
George [86]
Shaziya and 2020 private 267 128×128 pixels Data augmentasi are crop, CNN and UNet
Shyamala [87] zoom, rotate and flip
Arora et al. [88] 2021 NLM-China 662 15 mm CLAHE method UNet++
CXR
Nazir et al. [92] 2021 LIDC-IDRI 4682 3–30 mm Laplacian pyramid (LP) CNN
sparse vector fusion
Osadebey et al. 2021 LIDC-IDRI 1100 3–30 mm CNN CNN
[93]
Chavan et al [94] 2022 Shenzhen and 800 256×256 CNN CNN
Montgomery
Tandon et al [95] 2022 LIDC-IDRI 1018 3 mm Data augmentation CNN
methods (rescaling,
rotation, horizontal and
vertical flip)

6. LUNG NODULE DETECTION

The process of detecting objects in the lung tissue that are assumed to be nodules is referred to as
"candidate nodule detection," which is also the name of the phrase. This detection stage is carried out after the
lung segmentation stage. Lung segmentation is useful for lowering the burden of detecting CT input pictures
because the background and undesired areas have been removed prior to this stage. It is possible to recognize
lung nodules by the application of a number of distinct methods, such as random forest, SVM, naive bayes, k-
nearest neighbor (k-NN), and CNN. The CNN system is comprised of several distinct designs, the most
prominent of which are the CNN, a quicker R-CNN, a 3D CNN, and an R-CNN.
Using random forests and the 10-fold cross method, Gong et al. [41] were able to identify lung nodules
for CAD systems. The created CAD system was validated using two datasets, specifically LUNA16 and
ANODE09, respectively. Detection of lung nodules was carried out by De Carvalho Filho and colleagues using
the SVM method. The SVM is a cutting-edge algorithm that is derived from the Vapnik-Chervonenkis theory.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1053

The level of accuracy achieved by the SVM relies on the selection of kernel parameters such as C and the radial
basis function (RBF). For the testing, the LIDC-IDRI database was utilized, and in total, 140 new exams were
developed [58]. The Naive Bayes and KNN methods were proposed for the purpose of lung nodule detection
by Nóbrega et al. [57]. The builder for these methods was based on a Gaussian distribution of the probability
density function.
When attempting to detect lung nodules using machine learning techniques, the difficulty of defining
and choosing the attributes of a certain image arises. As the number of photographs in each category rises, the
process of extracting features from images gets increasingly time-consuming and labor-intensive [34] In
addition, over the past few years, methods based on deep learning have been developed for the detection of
lung nodules. The CNN method was developed by a group of researchers specifically for the purpose of
detecting lung nodules. In its most basic form, a CNN is made up of three layers: a convolutional layer (CONV),
a pooling layer, and a fully connected layer (FC). Jiang et al. [1] conducted research on the detection of lung
nodules using the CNN method, modifying it to use the function of rectified linear units in place of the
convolutional layer Rectified Linear Unit (ReLU). When compared to other activation functions on the CNN
method, the process of error training is sped up by using ReLU, which is one of the advantages obtained from
using it. At the pooling layer, operations known as max-pooling and average-pooling are carried out, whereas
the FC layer is made up of four separate channels. assessment of the suggested CNN approach using the LIDC
IDRI database. The CNN method, which consists of several layers, is utilized by Wang et al. [96] in their lung
nodule detection process. The activation function of the convolution layer is Leaky ReLU, the pooling layer
uses means and averages, and the final layer, the FC layer, uses global average pooling. The FC layer
implements a feature mapping strategy that makes use of a 4 by 4 matrix kernel in an effort to cut down on the
number of connected parameters. In addition, the batch normalization layer is implemented to lessen the impact
of overfitting and speed up the convergence of the network [96]. Kasinathan et al. [56] and Li et al. [36] carried
out yet another study for the purpose of detecting lung nodules using the CNN method. The CNN method is in
the process of being developed, and it currently incorporates architectural models such as region-based fully
convolutional networks (RFCN) [97], regional-CNN (R-CNN) [98], faster regional CNN (Faster R-CNN) [50],
ResNet50 [57], and 3D-CNN [99].
Deep learning is crucial for enabling the CNN approach to be employed in the analysis of medical
pictures as technology develops, the detection process gets computationally faster, and the amount of data
accessible increases. This was made possible by the use of the CNN method. The CNN method has a number
of benefits, the most notable of which are an enhanced image detection performance, high flexibility and
adaptability to a wide variety of datasets, and the capacity to be designed automatically and effectively by
making use of black-box operations [34]. According to the findings of previous research, the following are
some advantages of deep learning:
− The performance of the CAD system in detecting nodules in lung cancer may be improved through the
use of techniques from deep learning. Not only does the CAD system detect the presence of lung nodules,
but it also provides information on the location of those nodules and has the ability to categorize detected
nodules as either benign or malignant [97].
− Deep CNN has the potential to increase the sensitivity of the detection of lung nodules by reducing the
value of FPs/scan, thereby lowering the error rate in detection and, of course, improving the quality of
detection [99].
− Deep CNN can be used to detect lung nodules in a variety of dataset sources. For example, it can be used
to detect data from hospital A, and then it can be used to detect data from hospital B. Deep CNN is able
to discover various CT scans of the lungs and categorize them into distinct groups [37], [41], [50], [100].

7. CLASSIFICATION: FALSE POSITIVE REDUCTION

Following the step of candidate nodule detection, the image is then further classified as either
containing nodules or not containing nodules. This part of the process is referred to as the False Positive (FPs)
reduction stage. The FPs reduction process can be broken down into two distinct categories: the first is the
traditional feature-based classification, and the second is the CNN-based classification. In conventional
feature-based classification, feature extraction and the detection of nodule candidate nodules are both utilized.
There have been a few different approaches to feature extraction and candidate nodule detection that have been
suggested. The following is a review of some of the publications that pertain to the aforementioned two
classification stages for CAD systems in lung images.
The SVM method was proposed by Filho et al. [58] for the purpose of classifying lung images. This
method utilizes data obtained from the LIDC-IDRI database as well as a feature extraction process that makes
use of shapes and textures. With a FPs/scan value of 0.008 and a free-response operating characteristic (FROC)
value of 0.8062, the findings indicated that the test had an accuracy of 97.55%, a sensitivity of 85.91%, and a

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1054  ISSN: 2252-8938

specificity of 97.70%. Additionally, the test had a sensitivity of 85.91%. Han et al. [101] and Boroczky et al.
[102] carried out yet another investigation in which the SVM method was utilized. In this particular
investigation, the classification performed by feature-based SVM was dependent on categorical rules. These
categorical rules took the form of geometric or shape features, intensity features, gradient features, and
eigenvalue-based features. The proposed system was tested on 205 patient cases taken from the openly
accessible online LIDC database. The experimental results obtained a sensitivity of 89.2% at 4.14 FPs/scan,
and the system was found to be successful [101]. The results of the study showed that the genetic algorithm
method developed by Boroczky et al. for extracting features from private data obtained from lung CT scans
(including 52 real nodules and 443 fake ones) was 100% accurate and 56.4% accurate [102].
Jia et al. [103] proposed several three-dimensional methods for lung detection, some of which are
surface shadow display (SSD), volume rendering, maximum intensity rendering (MIP), and minimum intensity
rendering (MIP) (VR). The process of feature extraction uses ROI to more accurately identify suspicious
regions, paying particular attention to their shape, gray value, position, circularity, mean value of gray level,
and smoothness. Also taken into consideration is the mean value of gray level. With a sensitivity of 95% on
FPs/scan of 0.91, the proposed method was able to successfully detect lung image nodules. Tariq and Akram
[55] investigated the use of the neurofuzzy method in the detection of lung nodules. The neurofuzzy
classification can be broken down into two distinct sub-networks: the fuzzy self, which is responsible for
managing the network, and the multilevel, multilayer perception (MLP). In order to generate a pre-
classification vector, the feature vector is used as an input to the fuzzy layer. This pre-classification vector is
then assigned to the MLP as a sample test classification. The fuzzy self-layer network is in charge of locating
nodule pixels and organizing them into groups according to the similarity of the nodules (regardless of whether
or not there are nodules), but with varying membership values. In addition to this, the MLP network will
classify the input vectors that have been applied in order to select candidates from the appropriate category.
The testing was done with a total of one hundred datasets of lung CT images taken from various patients. The
accuracy reached using the strategy that was suggested is 95%. The back propagation neural network method,
which is composed of three layers, was utilized by Talebpour et al. in the process of classifying nodules as
distinct from other objects that are not nodules. There are 22 input neurons in the first layer, five hidden neurons
in the second layer, and one neuron in the output layer of the third layer. The first layer is considered the input
layer. Tan-sigmoid is an internal function that is present in every neuron. The proposed method was put to the
test using the LIDC-IDRI database, and the results showed that it had a sensitivity of 90% on an FP/scan of 10
[104]. Kuruvilla and Gunavathi [65] carried out yet another study, this time making use of the back propagation
neural network. The findings of the study indicated that a sensitivity of 91.4% could be achieved with an
FP/scan value of 30.
A random forest method was proposed by Gong et al. [41] for the purpose of classifying lung nodules
using the LUNA 16 and ANODE 09 databases. A sensitivity value of 79.3% was obtained from testing in both
databases with an FP/scan value of 4, and a sensitivity value of 84.62% was obtained from testing with an
FP/scan value of 2.8. Detecting lung cancer on CT images using the LIDC dataset was the focus of another
study that was carried out by Jayaraj and Sathiamoorthy [54], which utilized the random forest method. The
random forest method relies on a classification that takes into account both the index and the entropy to arrive
at its conclusions. The results obtained using the proposed method had an accuracy of 89.90%, a sensitivity of
90.85%, and a specificity of 88.32%, respectively. Nobrega et al. [57] conducted research on the effectiveness
of deep transfer learning when applied to the classification of lung nodule malignancy tasks. Their goal was to
improve such systems and put them to the test using the LIDC database. The proposed method is a contrast
between deep transfer learning and deep feature learning, and it yields the following results: an area under the
curve (AUC) of 93.1%, a true positive rate (TPR) of 85.38%, an evaluation metrics accuracy (ACC) of 88.41%,
precision of 73.48%, and an F1-score of 78.83%. They found that the deep transfer learning method is an
effective way to take CT images of lung nodules and extract the most important features from those images.
This was the finding of their research.
The CNN classification method is utilized next in the FPs reduction stage. Shin et al. [44] developed
the CNN approach for lung nodule identification, an architecture created with GoogLeNet that comprises of a
convolution layer, three pooling layers, and nine inception layers. GoogLeNet's inception layers are made up
of six convolutional layers and one pooling layer. The system was tested using the ILD dataset, and the findings
indicated a reduced accuracy (79%). To identify lung CT images, Golan et al. [105] suggested a deep CNN
technique utilizing a back-propagation algorithm. CNN is divided into two segments. The first section
comprises of numerous volumetric convolutions, rectified linear units (ReLU), and max-pooling layers. The
second component is a classifier composed of many fully linked, threshold, and softmax layers. The system
was tested using the LIDC dataset and yielded a poor sensitivity result of 78.9% at 20 frames per second.
Anthimopoulos et al. devised and tested CNN for the categorization of ILD patterns. The suggested technique
comprises of five convolutional layers, leaky ReLU activations, a pooling layer, and three dense layers. The

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1055

classification performance of lung patterns using the CNN approach is 85.5% [100]. Dou et al. proposed
employing three-dimensional (3-D) CNN to improve automated lung nodule identification from volumetric
CT data. The suggested approach was thoroughly tested in the LUNA16 challenge, yielding a sensitivity result
of 90% with 8 FPs per scan [99]. Tekade and Rajeswari [17] presented lung cancer detection and classification
using deep learning, and the technique is a 3D multipath VGG-like network that is tested on the LIDC,
LUNA16, and Kaggle datasets. The result for lung nodule identification and classification is 95.60% accuracy
and 0.387732 log loss. Kido et al. [98] developed a CAD method for lung anomalies based on CNN and areas
with CNN properties (R-CNN). R-CNN is an object detection system that use a CNN to categorize picture
areas inside an image. R-CNN was trained using marked aberrant lesions, and it indicated abnormal lesion
bounding boxes on the test picture. Their proposed approach has an accuracy of 84.7%. Jiang et al. [1]
suggested an automated detection approach for lung nodules based on a multigroup patch and a deep learning
network. With 15.1 FPs per scan, the CAD system achieved a result sensitivity of 94%. Huang et al. [50]
proposed utilizing deep convolutional neural networks to detect and separate lung nodules in thoracic CT
images in a quick and fully automated manner. The accuracy gained at the false-positive (FP) reduction step,
which is conducted using CNN, is 94.6% with 4 FPs per scan. The average dice coefficient of nodule
segmentation compared to the ground truth is 0.793.
Lung nodule identification using multi-resolution convolutional networks was proposed by Li et al.
[36] for the purpose of chest X-ray radiography. In order to extract the feature, they used patch-based multi-
resolution convolutional networks, and for classification, they used four distinct fusion algorithms. They
employed the JSRT database in order to evaluate their suggested technique. Within this database, they exhibited
an accuracy of 99% while only using 0.2 FPs each scan. Kasinathan et al. [56] proposed utilizing CNN to
automate the process of detecting and classifying three-dimensional lung tumors. The LIDC-IDRI dataset,
which included 850 lung nodule-lesion pictures, was utilized in the evaluation of the suggested model. The
outcome was that the model was accurate 97% of the time. In their study, Shi et al. [106] suggested a CNN
multi-scale feature fusion approach for the identification of lung nodules. The framework for detection is made
up of two parts: the production of region proposals and the minimization of false-positive results. The
architecture of the CNN model is VGG16, and trials performed on the LUNA16 dataset demonstrate that it has
an average sensitivity of 82.62%. Masood et al. suggested an automated technique for the identification of lung
cancer by employing a method known as the improved multidimensional region-based fully convolutional
network (mRFCN). The LIDC dataset was used to train and test their system, and the experiment findings
demonstrate that their system has a sensitivity of 98.1% and an accuracy of 97.91% [97]. Wang et al. [96]
proposed utilizing a raw patch-based CNN for the identification of lung nodules in CT images. On CT images
taken from the LIDC-IDRI dataset, they evaluated the performance of ResNet in comparison to that of many
alternative CNN architectures. As a result, they achieved a high detection sensitivity of 92.8% with 8 FPs per
scan.

8. CNN-BASED COMPUTER-AIDED LUNG NODULE DETECTION SYSTEM

It is abundantly evident that the CAD system is undergoing continuous development year after year
based on the investigations and analyses that we have outlined above for the automatic CAD detection system.
This takes place so that a higher-quality nodule detection may be achieved, which ultimately results in
increased efficiency. When it comes to identifying and categorizing lung nodules, the most effective CAD
system is one that is capable of achieving high levels of both accuracy and sensitivity. We conducted a literature
review on lung nodule detection and summarized many articles that were published between 2006 and 2022 in
the databases Science Direct, Springer Link, IEEE Xplore, and Web of Science. Our goal was to identify
potential research areas and future problems. It is crucial to analyze the framework and assessment process of
the technique that has been proposed, so it is not the primary focus while reviewing some of the most recent
research, even if it is important to directly compare the outcomes of the studies. On the other hand, we assessed
the technique of lung nodule identification in a CAD system based on the data that was utilized, the number of
nodules, the image size, and the best performance, which included sensitivity and the reduction of false-positive
results. In addition, we analyzed and evaluated a number of different techniques for detecting lung nodules at
each stage, including preprocessing, segmentation, nodule identification, and classification between nodules
and non-nodules using feature extraction and FPs reduction.
According to the findings of our review of the pertinent research literature, a number of studies
examining the identification of lung nodules made use of a substantial number of datasets. CT scans are the
most popular sort of dataset currently being used. Jiang et al. utilized the LIDC-IDRI dataset of 1006 images
[1], Huang et al. used a total of 888 images [26], Kasinathan et al. used 850 images [56], Wang et al. [96],
Masood et al. [97], and other researchers also used the same dataset. In addition to that, the LUNA16 dataset,
which is utilized by a large number of people, is available. Gong et al. made use of the 1186-image LUNA16
dataset [41], which was previously utilized by Tekade et al. [17], Dou et al. [99], Shi et al. [106], and a number
A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1056  ISSN: 2252-8938

of other researchers. Table 1 provides an overview of a variety of datasets taken from several different public
databases, in addition to providing information on these datasets.
In addition to the dataset that was utilized, we also performed an analysis of the CAD system by
employing the methods of preprocessing, segmentation, detection, and FPs reduction. The Table 2 provides a
summary of some of the preprocessing and segmentation techniques that we use. The Gaussian filter and the
Median Filter are the two types of preprocessing procedures that are utilized the most frequently. In order to
analyze all 1018 photos that were taken from the LIDC-IDRI database, Jayaraj et al. utilized Gaussian and
median filters. A gaussian filter is applied in order to reduce noise in lung cancer diagnosis, and a median filter
is employed in order to remove salt and pepper noise, which is a type of very tiny noise that can be found in
CT scans. In the CT picture of the lung, this will result in the smoothing of the image and a reduction in the
speckle noise [54]. Widodo et al. [53], Filho et al. [58], and Sankar et al. [86] utilized the same method in their
works.
The subsequent step in the process carried out by a CAD system is image segmentation. The process
of distinguishing lung nodules from other sections of a CT scan picture of the lung and then enhancing the
image that is produced as a consequence in order to acquire additional detail is referred to as lung segmentation.
The thresholding method is one of the most used approaches to segmentation [1], [55], [58]. Deep learning as
a segmentation approach has developed over time and can take on a variety of forms, depending on the lung
imaging nodule being analyzed. Using a faster convolutional neural network, Huang et al. suggested a
segmentation strategy for recognizing lung CT images [50], and they attained an accuracy of 94.6 f. RNN is a
method that was proposed by Sankar and George [86] to enhance the identification of juxtapleural and
juxtavascular lesions. UNet and CNN are two approaches that were proposed by Shaziya and colleagues for
segmenting lung CT images. When it comes to picture segmentation, the findings of the Thedicee similarity
coefficient (DSC) demonstrate that the UNet approach is 1.27 percent more effective than CNN [87]. In the
TB category, Arora et al. performed segmentation with the help of the UNet, and as a result, they achieved a
DSC value of 0.9680 [88]. As a direct consequence of this, deep learning segmentation strategies such as FCN,
RNN, CNN, and UNet offer a great deal of promise.
The procedure of segmentation is followed by the identification of nodules and the extraction of
features. The purpose of this step is to ascertain whether or not a certain picture is recognized as a nodule.
Using SVM’s and feature extraction with form and texture, Filho et al. [58] devised a method for the
identification of nodules that was both highly accurate and sensitive. Tariq et al. [55] employ neurofuzzy to
identify nodules, and they extract characteristics using vector and intensity. Back propagation neural network
(BPNN) was utilized by Talebpour et al. for the purpose of detecting nodules, in conjunction with geometric
and textural feature extraction [104]. However, machine learning algorithms for detecting nodules and
extracting features have difficulty with the work of defining and selecting the characteristics of a particular
picture, which also causes the task to become more time-consuming. In the process of nodule detection, deep
learning CNNs are utilized extensively. Several studies, such as the ones by Jiang et al. [1] using CNN, Tekade
et al. [17] using CNN with 3D multipath VGG, Wang et al. using CNN (ResNet) [96], and Masood et al. [97]
using RFCN, Dou et al. [99] using 3D-CNN, have achieved an accuracy and sensitivity that is more than 90%.
Performing methods like as feature extraction and false-positive reduction are essential for achieving the best
results possible in lung nodule detection.
Moving forward, there is a pressing need for more study into the development of CAD systems for
the identification of lung nodules. The target is to acquire a detection result that is more accurate as well as a
high sensitivity value that has the potential to lower the value of FP reduction. The following is a list of
significant concepts for potential CAD systems that can be utilized in the future to locate lung nodules:
− Developing new methods of deep learning, such as the CNN method, with the primary goal of enhancing
the performance of lung nodule identification. In addition, the batch normalization layer can be included
in order to lessen the effects of overfitting and speed up the convergence of the network [96].
− Developing a CAD system that is capable of detecting all types of lung nodules with high accuracy and
sensitivity, as well as a low percentage of false-positive results while maintaining these characteristics.
− If the suggested technique is trained and evaluated using a large number of datasets, such as the LIDC-
IDRI and LUNA16 public databases, then it will be able to provide a more comprehensive assessment of
the general and clinical performance of the detection system.

9. CONCLUSION
In this study, we have presented a critical analysis of some of the research that has been done on CAD
systems for lung nodule identification by doing literature investigations and analyzing the results. The research
that has been done on CAD systems for lung nodule identification focuses on the identification of nodules in
the lungs. There have been a few distinct lines of inquiry that have utilized CT scan photos to examine the

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1057

efficacy of the suggested method. After providing an overview of the computer-aided diagnosis (CAD) system
for detecting lung nodules, it has been established that the system consists of various steps. These phases
involve preparing or obtaining image data, preprocessing, segmenting, lung nodule detection, and FPs
reduction, which incorporates feature extraction. Other steps include detecting and segmenting lung nodules.
The steps are broken down into their component parts farther down in this article. We discovered that some of
the works on lung nodule detection had better results than others based on parameters such as sensitivity,
specificity, accuracy, and the number of FPs per scan, as well as other parameters, after reviewing a few of the
most well-known works on the subject and evaluating the proposed method with a dataset taken from the
LIDC-IDRI database. This was the conclusion we reached after looking at a few of the most well-known works
on lung nodule detection and evaluating the method. In addition, we have provided an overview of the various
methods that can be used for each of the steps of the process of identifying lung nodules. In recent years, deep
learning algorithms such as CNN have been an increasingly popular means of locating prospective nodules and
extracting attributes. Over the course of the last few years, this strategy has become increasingly common.
Despite the fact that we discovered that some CAD systems achieved outstanding sensitivity with low false-
positive rates, there are still a great many barriers to overcome in order to optimize CAD systems for the
detection of lung cancer. It is expected that a capable CAD system will be able to assist radiologists in the
detection of lung cancer. This is the single most critical thing that it ought to be able to accomplish.

ACKNOWLEDGMENT
The authors gratefully acknowledge Universitas Gadjah Mada and Universitas Ahmad Dahlan for
their financial support of this collaborative work, as well as the Embedded Systems and Power Electronics
Research Group (ESPERG) for their publication assistance.

REFERENCES
[1] H. Jiang, H. Ma, W. Qian, M. Gao, and Y. Li, “An automatic detection system of lung nodule based on multigroup patch-based
deep learning network,” IEEE Journal of Biomedical and Health Informatics, vol. 22, no. 4, pp. 1227–1237, 2018, doi:
10.1109/JBHI.2017.2725903.
[2] T. Uemura and T. Hida, “Durvalumab showed long and durable effects after chemoradiotherapy in stage III non-small cell lung
cancer: Results of the PACIFIC study,” Journal of Thoracic Disease, vol. 10, pp. S1108–S1112, 2018, doi:
10.21037/jtd.2018.03.180.
[3] T. Hou et al., “Alpha thalassemia/intellectual disability X-linked deficiency sensitizes non-small cell lung cancer to immune
checkpoint inhibitors,” Frontiers in Oncology, vol. 10, 2020, doi: 10.3389/fonc.2020.608300.
[4] C. Olingy et al., “CD33 expression on peripheral blood monocytes predicts efficacy of anti-PD-1 immunotherapy against non-small
cell lung cancer,” Frontiers in Immunology, vol. 13, 2022, doi: 10.3389/fimmu.2022.842653.
[5] M. Casás-Selves et al., “Tankyrase and the canonical Wnt pathway protect lung cancer cells from EGFR inhibition,” Cancer
Research, vol. 72, no. 16, pp. 4154–4164, 2012, doi: 10.1158/0008-5472.CAN-11-2848.
[6] Z. Huang et al., “Correlation of cancer stem cell markers and immune cell markers in resected non-small cell lung cancer,” Journal
of Cancer, vol. 8, no. 16, pp. 3190–3197, 2017, doi: 10.7150/jca.20172.
[7] R. L. Siegel, K. D. Miller, H. E. Fuchs, and A. Jemal, “Cancer statistics, 2022,” CA: A Cancer Journal for Clinicians, vol. 72, no.
1, pp. 7–33, 2022, doi: 10.3322/caac.21708.
[8] M. S. AL-Huseiny and A. S. Sajit, “Transfer learning with GoogLeNet for detection of lung cancer,” Indonesian Journal of
Electrical Engineering and Computer Science, vol. 22, no. 2, pp. 470–478, 2020, doi: 10.11591/ijeecs.v22.i2.pp470-478.
[9] R. Pandian, D. N. S. Ravi Kumar, and R. R. Kumar, “Development of algorithm for identification of maligant growth in cancer
using artificial neural network,” International Journal of Electrical and Computer Engineering, vol. 10, no. 6, pp. 5709–5713, 2020,
doi: 10.11591/ijece.v10i6.pp5709-5713.
[10] H. F. Kareem, M. S. AL-Husieny, F. Y. Mohsen, E. A. Khalil, and Z. S. Hassan, “Evaluation of SVM performance in the detection
of lung cancer in marked CT scan dataset,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 21, no. 3, pp.
1731–1738, 2021, doi: 10.11591/ijeecs.v21.i3.pp1731-1738.
[11] R. D. Abdu-Aljabar and O. A. Awad, “Parallel extreme gradient boosting classifier for lung cancer detection,” Indonesian Journal
of Electrical Engineering and Computer Science, vol. 24, no. 3, pp. 1610–1617, 2021, doi: 10.11591/ijeecs.v24.i3.pp1610-1617.
[12] S. Kothari, S. Chiwhane, S. Jain, and M. Baghel, “Cancerous brain tumor detection using hybrid deep learning framework,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 26, no. 3, pp. 1651–1661, 2022, doi:
10.11591/ijeecs.v26.i3.pp1651-1661.
[13] O. R. Kadhim, H. J. Motlak, and K. K. Abdalla, “Computer-aided diagnostic system kinds and pulmonary nodule detection
efficacy,” International Journal of Electrical and Computer Engineering, vol. 12, no. 5, pp. 4734–4745, 2022, doi:
10.11591/ijece.v12i5.pp4734-4745.
[14] M. F. Abdullah, S. N. Sulaiman, M. K. Osman, N. K. A. Karim, S. Setumin, and I. S. Isa, “A new procedure for lung region
segmentation from computed tomography images,” International Journal of Electrical and Computer Engineering, vol. 12, no. 5,
pp. 4978–4987, 2022, doi: 10.11591/ijece.v12i5.pp4978-4987.
[15] A. A. Abd Al-Ameer, G. A. Hussien, and H. A. Al Ameri, “Lung cancer detection using image processing and deep learning,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 28, no. 2, pp. 987–993, 2022, doi:
10.11591/ijeecs.v28.i2.pp987-993.
[16] R. S. Shankar, R. S. Chigurupati, P. Voosala, and N. Pilli, “An extensible framework for recurrent breast cancer prognosis using
deep learning techniques,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 29, no. 2, pp. 931–941, 2023,
doi: 10.11591/ijeecs.v29.i2.pp931-941.
[17] R. Tekade and K. Rajeswari, “Lung cancer detection and classification using deep learning,” Proceedings - 2018 4th International

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1058  ISSN: 2252-8938

Conference on Computing, Communication Control and Automation, ICCUBEA 2018, 2018, doi:
10.1109/ICCUBEA.2018.8697352.
[18] N. E. A. Khalid, M. F. Ismail, M. A. A. B. Manaf, A. F. A. Fadzil, and S. Ibrahim, “MRI brain tumor segmentation: A forthright
image processing approach,” Bulletin of Electrical Engineering and Informatics, vol. 9, no. 3, pp. 1024–1031, 2020, doi:
10.11591/eei.v9i3.2063.
[19] I. Assini, A. Badri, A. Sahel, and A. Baghdad, “Hybrid multiple watermarking technique for securing medical images of modalities
MRI, CT scan, and X-ray,” International Journal of Electrical and Computer Engineering, vol. 10, no. 3, pp. 2349–2356, 2020,
doi: 10.11591/ijece.v10i3.pp2349-2356.
[20] N. Raju, H. B. Anita, and P. Augustine, “Identification of interstitial lung diseases using deep learning,” International Journal of
Electrical and Computer Engineering, vol. 10, no. 6, pp. 6283–6291, 2020, doi: 10.11591/IJECE.V10I6.PP6283-6291.
[21] N. Saha and M. A. Rahaman, “Bone cancer detection using electrical impedance tomography,” Indonesian Journal of Electrical
Engineering and Computer Science, vol. 24, no. 1, pp. 245–252, 2021, doi: 10.11591/ijeecs.v24.i1.pp245-252.
[22] R. Supriyanti, M. Alqaaf, Y. Ramadhani, and H. B. Widodo, “Morphological characteristics of X-ray thorax images of COVID-19
patients using the Bradley thresholding segmentation,” Indonesian Journal of Electrical Engineering and Computer Science, vol.
24, no. 2, pp. 1074–1083, 2021, doi: 10.11591/ijeecs.v24.i2.pp1074-1083.
[23] W. S. Alazawee, Z. H. Naji, and W. T. Ali, “Analyzing and detecting hemorrhagic and ischemic stroke-based on bit plane slicing
and edge detection algorithms,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 25, no. 2, pp. 1003–
1010, 2022, doi: 10.11591/ijeecs.v25.i2.pp1003-1010.
[24] A. Elaraby and A. Taha, “An approach for cross-modality guided quality enhancement of liver image,” International Journal of
Electrical and Computer Engineering, vol. 12, no. 2, pp. 1449–1455, 2022, doi: 10.11591/ijece.v12i2.pp1449-1455.
[25] H. A. Owida, H. S. Migdadi, O. S. M. Hemied, N. F. F. Alshdaifat, S. F. A. Abuowaida, and R. S. Alkhawaldeh, “Deep learning
algorithms to improve COVID-19 classification based on CT images,” Bulletin of Electrical Engineering and Informatics, vol. 11,
no. 5, pp. 2876–2885, 2022, doi: 10.11591/eei.v11i5.3802.
[26] V. Bhavana and H. K. Krishnappa, “Multi-modal image fusion using contourlet and wavelet transforms: a multi-resolution
approach,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 28, no. 2, pp. 762–768, 2022, doi:
10.11591/ijeecs.v28.i2.pp762-768.
[27] T. Asmaria, D. A. Mayasari, A. D. G. Febrananda, N. Nurul, A. J. Rahyussalim, and I. Kartika, “Computed tomography image
analysis for Indonesian total hip arthroplasty designs,” International Journal of Electrical and Computer Engineering, vol. 12, no.
6, pp. 6123–6131, 2022, doi: 10.11591/ijece.v12i6.pp6123-6131.
[28] F. Shaukat, G. Raja, and A. F. Frangi, “Computer-aided detection of lung nodules: A review,” Journal of Medical Imaging, vol. 6,
no. 02, p. 1, 2019, doi: 10.1117/1.jmi.6.2.020901.
[29] S. Saeed and A. Abdullah, “Recognition of brain cancer and cerebrospinal fluid due to the usage of different MRI image by utilizing
support vector machine,” Bulletin of Electrical Engineering and Informatics, vol. 9, no. 2, pp. 619–625, 2020, doi:
10.11591/eei.v9i2.1869.
[30] S. Hartini and Z. Rustam, “The comparison study of kernel KC-means and support vector machines for classifying schizophrenia,”
Telkomnika (Telecommunication Computing Electronics and Control), vol. 18, no. 3, pp. 1643–1649, 2020, doi:
10.12928/TELKOMNIKA.v18i3.14847.
[31] S. A. Taie and W. Ghonaim, “A new model for early diagnosis of alzheimer’s disease based on bat-SVM classifier,” Bulletin of
Electrical Engineering and Informatics, vol. 10, no. 2, pp. 759–766, 2021, doi: 10.11591/eei.v10i2.2714.
[32] O. Y. Ogundepo, I. O. A. Omeiza, and J. P. Oguntoye, “Optimized textural features for mass classification in digital mammography
using a weighted average gravitational search algorithm,” International Journal of Electrical and Computer Engineering, vol. 12,
no. 5, pp. 5001–5013, 2022, doi: 10.11591/ijece.v12i5.pp5001-5013.
[33] B. S. Shukur and M. M. Mijwil, “Involving machine learning techniques in heart disease diagnosis: a performance analysis,”
International Journal of Electrical and Computer Engineering, vol. 13, no. 2, pp. 2177–2185, 2023, doi:
10.11591/ijece.v13i2.pp2177-2185.
[34] W. Cao, R. Wu, G. Cao, and Z. He, “A comprehensive review of computer-aided diagnosis of pulmonary nodules based on
computed tomography scans,” IEEE Access, vol. 8, pp. 154007–154023, 2020, doi: 10.1109/ACCESS.2020.3018666.
[35] V. A. Binson and M. Subramoniam, “Advances in early lung cancer detection: A systematic review,” 2018 International Conference
on Circuits and Systems in Digital Enterprise Technology, ICCSDET 2018, 2018, doi: 10.1109/ICCSDET.2018.8821188.
[36] X. Li et al., “Multi-resolution convolutional networks for chest X-ray radiograph based lung nodule detection,” Artificial
Intelligence in Medicine, vol. 103, 2020, doi: 10.1016/j.artmed.2019.101744.
[37] S. Anitha and T. R. Ganesh Babu, “An efficient method for the detection of oblique fissures from computed tomography images of
lungs,” Journal of Medical Systems, vol. 43, no. 8, 2019, doi: 10.1007/s10916-019-1396-0.
[38] S. G. Armato et al., “The lung image database consortium (LIDC) and image database resource initiative (IDRI): A completed
reference database of lung nodules on CT scans,” Medical Physics, vol. 38, no. 2, pp. 915–931, 2011, doi: 10.1118/1.3528204.
[39] Y. R. Zhao, X. Xie, H. J. De Koning, W. P. Mali, R. Vliegenthart, and M. Oudkerk, “NELSON lung cancer screening study,”
Cancer Imaging, vol. 11, no. SPEC. ISS. A, 2011, doi: 10.1102/1470-7330.2011.9020.
[40] S. G. Armato et al., “The reference image database to evaluate response to therapy in lung cancer (RIDER) project: A resource for
the development of change-analysis software,” Clinical Pharmacology and Therapeutics, vol. 84, no. 4, pp. 448–456, 2008, doi:
10.1038/clpt.2008.161.
[41] J. Gong, J. yu Liu, L. jia Wang, X. wen Sun, B. Zheng, and S. dong Nie, “Automatic detection of pulmonary nodules in CT images
by incorporating 3D tensor filtering with local image feature analysis,” Physica Medica, vol. 46, pp. 124–133, 2018, doi:
10.1016/j.ejmp.2018.01.019.
[42] “NLST datasets,” [Online]. Available: https://cdas.cancer.gov/nlst/.
[43] M. Dolejsi, J. Kybic, M. Polovincak, and S. Tuma, “The lung TIME: Annotated lung nodule dataset and nodule detection
framework,” Medical Imaging 2009: Computer-Aided Diagnosis, vol. 7260, p. 72601U, 2009, doi: 10.1117/12.811645.
[44] H. C. Shin et al., “Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and
transfer learning,” IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1285–1298, 2016, doi: 10.1109/TMI.2016.2528162.
[45] L. Kinahan, P. Muzi, M. Bialecki, B. Herman, and B. Coombs, “Data from the ACRIN 6668 trial NSCLC-FDG-PET [Data set].
The cancer imaging archive,” 2019, doi: 10.7937/tcia.2019.30il.
[46] M. Hershman et al., “Impact of interobserver variability in manual segmentation of non‐small cell lung cancer (Nsclc) applying
low‐rank radiomic representation on computed tomography,” Cancers, vol. 13, no. 23, 2021, doi: 10.3390/cancers13235985.
[47] O. Grove et al., “Quantitative computed tomographic descriptors associate tumor shape complexity and intratumor heterogeneity
with prognosis in lung adenocarcinoma,” PLOS ONE, vol. 10, no. 3, p. e0118261, Mar. 2015, doi: 10.1371/journal.pone.0118261.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1059

[48] K. Smith, “QIN LUNG CT.”

[49] G. Han et al., “The LISS - A public database of common imaging signs of lung diseases for computer-aided detection and diagnosis
research and medical education,” IEEE Transactions on Biomedical Engineering, vol. 62, no. 2, pp. 648–656, 2015, doi:
10.1109/TBME.2014.2363131.
[50] X. Huang, W. Sun, T. L. (Bill) Tseng, C. Li, and W. Qian, “Fast and fully-automated detection and segmentation of pulmonary
nodules in thoracic CT scans using deep convolutional neural networks,” Computerized Medical Imaging and Graphics, vol. 74,
pp. 25–36, 2019, doi: 10.1016/j.compmedimag.2019.02.003.
[51] S. Diciotti, G. Picozzi, M. Falchini, M. Mascalchi, N. Villari, and G. Valli, “3-D segmentation algorithm of small lung nodules in
spiral CT images,” IEEE Transactions on Information Technology in Biomedicine, vol. 12, no. 1, pp. 7–19, 2008, doi:
10.1109/TITB.2007.899504.
[52] H. A. H. Chaudhry et al., “UniToChest: A lung image dataset for segmentation of cancerous nodules on CT scans,” Lecture Notes
in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 13231
LNCS, pp. 185–196, 2022, doi: 10.1007/978-3-031-06427-2_16.
[53] S. Widodo, R. N. Rohmah, and B. Handaga, “Classification of lung nodules and arteries in computed tomography scan image using
principle component analysis,” Proceedings - 2017 2nd International Conferences on Information Technology, Information Systems
and Electrical Engineering, ICITISEE 2017, vol. 2018-January, pp. 153–158, 2018, doi: 10.1109/ICITISEE.2017.8285485.
[54] D. Jayaraj and S. Sathiamoorthy, “Random forest based classification model for lung cancer prediction on computer tomography
images,” Proceedings of the 2nd International Conference on Smart Systems and Inventive Technology, ICSSIT 2019, pp. 100–104,
2019, doi: 10.1109/ICSSIT46314.2019.8987772.
[55] A. Tariq and M. U. Akram, “Lung nodule detection in CT images using neuro fuzzy classifier,” Telkomnika, vol. 11, no. 2, pp. 331–
336, 2013, doi: 10.12928/telkomnika.v11i2.934.
[56] G. Kasinathan, S. Jayakumar, A. H. Gandomi, M. Ramachandran, S. J. Fong, and R. Patan, “Automated 3-D lung tumor detection
and classification by an active contour model and CNN classifier,” Expert Systems with Applications, vol. 134, pp. 112–119, 2019,
doi: 10.1016/j.eswa.2019.05.041.
[57] R. V. M. da Nóbrega, P. P. Rebouças Filho, M. B. Rodrigues, S. P. P. da Silva, C. M. J. M. Dourado Júnior, and V. H. C. de
Albuquerque, “Lung nodule malignancy classification in chest computed tomography images using transfer learning and
convolutional neural networks,” Neural Computing and Applications, vol. 32, no. 15, pp. 11065–11082, 2020, doi: 10.1007/s00521-
018-3895-1.
[58] A. O. De Carvalho Filho, W. B. De Sampaio, A. C. Silva, A. C. de Paiva, R. A. Nunes, and M. Gattass, “Automatic detection of
solitary lung nodules using quality threshold clustering, genetic algorithm and diversity index,” Artificial Intelligence in Medicine,
vol. 60, no. 3, pp. 165–177, 2014, doi: 10.1016/j.artmed.2013.11.002.
[59] G. Zhang et al., “Automatic nodule detection for lung cancer in CT images: A review,” Computers in Biology and Medicine, vol.
103, pp. 287–300, 2018, doi: 10.1016/j.compbiomed.2018.10.033.
[60] C. Pérez-Benito, S. Morillas, C. Jordán, and J. A. Conejero, “Smoothing Vs. sharpening of colour images: Together or separated,”
Applied Mathematics and Nonlinear Sciences, vol. 2, no. 1, pp. 299–316, 2017, doi: 10.21042/amns.2017.1.00025.
[61] S. Ukil and J. M. Reinhardt, “Smoothing lung segmentation surfaces in three-dimensional X-ray CT images using anatomic
guidance,” Academic Radiology, vol. 12, no. 12, pp. 1502–1511, 2005, doi: 10.1016/j.acra.2005.08.008.
[62] U. S. Govindarajulu, D. Spiegelman, S. W. Thurston, B. Ganguli, and E. A. Eisen, “Comparing smoothing techniques in Cox models
for exposure-response relationships,” Statistics in Medicine, vol. 26, no. 20, pp. 3735–3752, 2007, doi: 10.1002/sim.2848.
[63] U. S. Govindarajulu, E. J. Malloy, B. Ganguli, D. Spiegelman, and E. A. Eisen, “The comparison of alternative smoothing methods
for fitting non-linear exposure-response relationships with Cox models in a simulation study,” International Journal of Biostatistics,
vol. 5, no. 1, 2009, doi: 10.2202/1557-4679.1104.
[64] D. Surya Prabha and J. Satheesh Kumar, “Performance analysis of image smoothing methods for low level of distortion,” 2016
IEEE International Conference on Advances in Computer Applications, ICACA 2016, pp. 372–376, 2017, doi:
10.1109/ICACA.2016.7887983.
[65] J. Kuruvilla and K. Gunavathi, “Lung cancer classification using neural networks for CT images,” Computer Methods and Programs
in Biomedicine, vol. 113, no. 1, pp. 202–209, 2014, doi: 10.1016/j.cmpb.2013.10.011.
[66] H. Zhu, F. H. Y. Chan, and F. K. Lam, “Image contrast enhancement by constrained local histogram equalization,” Computer Vision
and Image Understanding, vol. 73, no. 2, pp. 281–290, 1999, doi: 10.1006/cviu.1998.0723.
[67] P. Perona and J. Malik, “Scale-space and edge detection using anisotropic diffusion,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 12, no. 7, pp. 629–639, 1990, doi: 10.1109/34.56205.
[68] A. E. Ilesanmi and T. O. Ilesanmi, “Methods for image denoising using convolutional neural network: A review,” Complex and
Intelligent Systems, vol. 7, no. 5, pp. 2179–2198, 2021, doi: 10.1007/s40747-021-00428-4.
[69] C. Anam et al., “Noise reduction in CT images using a selective mean filter,” Journal of Biomedical Physics and Engineering, vol.
10, no. 5, pp. 623–634, 2020, doi: 10.31661/jbpe.v0i0.2002-1072.
[70] A. Lucas, M. Iliadis, R. Molina, and A. K. Katsaggelos, “Using deep neural networks for inverse problems in imaging: Beyond
analytical methods,” IEEE Signal Processing Magazine, vol. 35, no. 1, pp. 20–36, 2018, doi: 10.1109/MSP.2017.2760358.
[71] A. Khan, A. Sohail, U. Zahoora, and A. S. Qureshi, “A survey of the recent architectures of deep convolutional neural networks,”
Artificial Intelligence Review, vol. 53, no. 8, pp. 5455–5516, 2020, doi: 10.1007/s10462-020-09825-6.
[72] Z. Guo, Y. Sun, M. Jian, and X. Zhang, “Deep residual network with sparse feedback for image restoration,” Applied Sciences
(Switzerland), vol. 8, no. 12, 2018, doi: 10.3390/app8122417.
[73] C. Tian, L. Fei, W. Zheng, Y. Xu, W. Zuo, and C. W. Lin, “Deep learning on image denoising: An overview,” Neural Networks,
vol. 131, pp. 251–275, 2020, doi: 10.1016/j.neunet.2020.07.025.
[74] A. R. Pulagam, G. B. Kande, V. K. R. Ede, and R. B. Inampudi, “Automated lung segmentation from HRCT scans with diffuse
parenchymal lung diseases,” Journal of Digital Imaging, vol. 29, no. 4, pp. 507–519, 2016, doi: 10.1007/s10278-016-9875-z.
[75] E. M. Van Rikxoort, B. De Hoop, S. Van De Vorst, M. Prokop, and B. Van Ginneken, “Automatic segmentation of pulmonary
segments from volumetric chest CT scans,” IEEE Transactions on Medical Imaging, vol. 28, no. 4, pp. 621–630, 2009, doi:
10.1109/TMI.2008.2008968.
[76] Q. Wei, Y. Hu, G. Gelfand, and J. H. MacGregor, “Segmentation of lung lobes in high-resolution isotropic CT images,” IEEE
Transactions on Biomedical Engineering, vol. 56, no. 5, pp. 1383–1393, 2009, doi: 10.1109/TBME.2009.2014074.
[77] P. Campadelli, E. Casiraghi, and D. Artioli, “A fully automated method for lung nodule detection from postero-anterior chest
radiographs,” IEEE Transactions on Medical Imaging, vol. 25, no. 12, pp. 1588–1603, 2006, doi: 10.1109/TMI.2006.884198.
[78] D. C. Tseng and C. H. Chang, “Color segmentation using perceptual attributes,” Proceedings - International Conference on Pattern
Recognition, vol. 3, pp. 228–231, 1992, doi: 10.1109/ICPR.1992.201967.
[79] H. Da Cheng and Y. Sun, “A hierarchical approach to color image segmentation using homogeneity,” IEEE Transactions on Image
A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)
1060  ISSN: 2252-8938

Processing, vol. 9, no. 12, pp. 2071–2082, 2000, doi: 10.1109/83.887975.

[80] F. Spagnolo, S. Perri, and P. Corsonello, “An efficient hardware-oriented single-pass approach for connected component analysis,”
Sensors (Switzerland), vol. 19, no. 14, 2019, doi: 10.3390/s19143055.
[81] A. Sheta, M. S. Braik, and S. Aljahdali, “Genetic algorithms: A tool for image segmentation,” Proceedings of 2012 International
Conference on Multimedia Computing and Systems, ICMCS 2012, pp. 84–90, 2012, doi: 10.1109/ICMCS.2012.6320144.
[82] A. Choudhury, R. R. Subramanian, and G. Sunder, “A novel approach for tumor segmentation for lung cancer using multi-objective
genetic algorithm and connected component analysis,” Advances in Intelligent Systems and Computing, vol. 828, pp. 367–376,
2019, doi: 10.1007/978-981-13-1610-4_37.
[83] J. R. F. da Silva Sousa, A. C. Silva, A. C. de Paiva, and R. A. Nunes, “Methodology for automatic detection of lung nodules in
computerized tomography images,” Computer Methods and Programs in Biomedicine, vol. 98, no. 1, pp. 1–14, 2010, doi:
10.1016/j.cmpb.2009.07.006.
[84] Y. Chunran and W. Yuanyuan, “Nodule on CT Images,” pp. 2–6, 2018.
[85] T. Messay, R. C. Hardie, and T. R. Tuinstra, “Segmentation of pulmonary nodules in computed tomography using a regression
neural network approach and its application to the lung image database consortium and image database resource initiative dataset,”
Medical Image Analysis, vol. 22, no. 1, pp. 48–62, 2015, doi: 10.1016/j.media.2015.02.002.
[86] S. P. Sankar and D. E. George, “Regression neural network segmentation approach with LIDC-IDRI for lung lesion,” Journal of
Ambient Intelligence and Humanized Computing, vol. 12, no. 5, pp. 5571–5580, 2021, doi: 10.1007/s12652-020-02069-w.
[87] H. Shaziya and K. Shyamala, “Pulmonary CT images segmentation using CNN and UNet models of deep learning,” 2020 IEEE
Pune Section International Conference, PuneCon 2020, pp. 195–201, 2020, doi: 10.1109/PuneCon50868.2020.9362463.
[88] R. Arora, I. Saini, and N. Sood, “Modified UNet++ model: A deep model for automatic segmentation of lungs from chest X-ray
images,” ICSCCC 2021 - International Conference on Secure Cyber Computing and Communications, pp. 166–169, 2021, doi:
10.1109/ICSCCC51823.2021.9478101.
[89] K. L. Hua, C. H. Hsu, S. C. Hidayati, W. H. Cheng, and Y. J. Chen, “Computer-aided classification of lung nodules on computed
tomography images via deep learning technique,” OncoTargets and Therapy, vol. 8, pp. 2015–2022, 2015, doi:
10.2147/OTT.S80733.
[90] W. Ausawalaithong, A. Thirach, S. Marukatat, and T. Wilaiprasitporn, “Automatic lung cancer prediction from chest X-ray images
using the deep learning approach,” BMEiCON 2018 - 11th Biomedical Engineering International Conference, 2019, doi:
10.1109/BMEiCON.2018.8609997.
[91] D. Ardila et al., “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,”
Nature Medicine, vol. 25, no. 6, pp. 954–961, 2019, doi: 10.1038/s41591-019-0447-x.
[92] I. Nazir, I. U. Haq, M. M. Khan, M. B. Qureshi, H. Ullah, and S. Butt, “Efficient pre-processing and segmentation for lung cancer
detection using fused CT images,” Electronics (Switzerland), vol. 11, no. 1, 2022, doi: 10.3390/electronics11010034.
[93] M. Osadebey, H. K. Andersen, D. Waaler, K. Fossaa, A. C. T. Martinsen, and M. Pedersen, “Three-stage segmentation of lung
region from CT images using deep neural networks,” BMC Medical Imaging, vol. 21, no. 1, 2021, doi: 10.1186/s12880-021-00640-
1.
[94] M. Chavan, V. Varadarajan, S. Gite, and K. Kotecha, “Deep neural network for lung image segmentation on chest X-ray,”
Technologies, vol. 10, no. 5, p. 105, 2022, doi: 10.3390/technologies10050105.
[95] R. Tandon, S. Agrawal, A. Chang, and S. S. Band, “VCNet: Hybrid deep learning model for detection and classification of lung
carcinoma using chest radiographs,” Frontiers in Public Health, vol. 10, 2022, doi: 10.3389/fpubh.2022.894920.
[96] Q. Wang, F. Shen, L. Shen, J. Huang, and W. Sheng, “Lung nodule detection in CT images using a raw patch-based convolutional
neural network,” Journal of Digital Imaging, vol. 32, no. 6, pp. 971–979, 2019, doi: 10.1007/s10278-019-00221-3.
[97] A. Masood et al., “Automated decision support system for lung cancer detection and classification via enhanced RFCN with
multilayer fusion RPN,” IEEE Transactions on Industrial Informatics, vol. 16, no. 12, pp. 7791–7801, 2020, doi:
10.1109/TII.2020.2972918.
[98] S. Kido, Y. Hirano, and N. Hashimoto, “Detection and classification of lung abnormalities by use of convolutional neural network
(CNN) and regions with CNN features (R-CNN),” 2018 International Workshop on Advanced Image Technology, IWAIT 2018, pp.
1–4, 2018, doi: 10.1109/IWAIT.2018.8369798.
[99] 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 Transactions on Biomedical Engineering, vol. 64, no. 7, pp. 1558–1567, 2017, doi: 10.1109/TBME.2016.2613502.
[100] M. Anthimopoulos, S. Christodoulidis, L. Ebner, A. Christe, and S. Mougiakakou, “Lung pattern classification for interstitial lung
diseases using a deep convolutional neural network,” IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1207–1216, 2016,
doi: 10.1109/TMI.2016.2535865.
[101] R. Bhakkiyalakshmi, P. Ponnammal, M. K. Srilekha, and K. Abhishikt Sai, “Fast and adaptive detection of pulmonary nodules in
thoracic CT images using a contextual clustering based region growing,” International Journal of Engineering and
Technology(UAE), vol. 7, no. 2, pp. 106–108, 2018, doi: 10.14419/ijet.v7i2.24.12010.
[102] L. Böröczky, L. Zhao, and K. P. Lee, “Feature subset selection for improving the performance of false positive reduction in lung
nodule CAD,” IEEE Transactions on Information Technology in Biomedicine, vol. 10, no. 3, pp. 504–511, 2006, doi:
10.1109/TITB.2006.872063.
[103] T. Jia, D. Z. Zhao, Y. Wei, X. H. Zhu, and X. Wang, “Computer-aided lung nodule detection based on CT images,” 2007
IEEE/ICME International Conference on Complex Medical Engineering, CME 2007, pp. 816–819, 2007, doi:
10.1109/ICCME.2007.4381854.
[104] A. R. Talebpour, H. R. Hemmati, and M. Z. Hosseinian, “Automatic lung nodules detection in computed tomography images using
nodule filtering and neural networks,” 22nd Iranian Conference on Electrical Engineering, ICEE 2014, pp. 1883–1887, 2014, doi:
10.1109/IranianCEE.2014.6999847.
[105] R. Golan, C. Jacob, and J. Denzinger, “Lung nodule detection in CT images using deep convolutional neural networks,” Proceedings
of the International Joint Conference on Neural Networks, vol. 2016-October, pp. 243–250, 2016, doi:
10.1109/IJCNN.2016.7727205.
[106] H. Shi, Z. Peng, and H. Wan, “Pulmonary nodules detection based on CNN multi-scale feature fusion,” 2019 IEEE 11th
International Conference on Advanced Infocomm Technology, ICAIT 2019, pp. 86–90, 2019, doi: 10.1109/ICAIT.2019.8935936.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1044-1061

Int J Artif Intell ISSN: 2252-8938  1061

BIOGRAPHIES OF AUTHORS

Sekar Sari received her B.Eng. degree in Electrical Engineering from

Universitas Ahmad Dahlan, Yogyakarta, Indonesia, in 2019. After graduating, she pursued a
Master's degree in the Department of Electrical and Information Engineering at Universitas
Gadjah Mada, Indonesia, and received her M.Eng. in 2022. Her research interests include
electronics design, deep learning, biomedical engineering, image processing, and pattern
recognition. She can be contacted at [email protected].

Tole Sutikno is a lecturer in the Electrical Engineering Department at the

Universitas Ahmad Dahlan (UAD), Yogyakarta, Indonesia. He received his B.Eng., M.Eng.,
and Ph.D. degrees in Electrical Engineering from Universitas Diponegoro, Universitas
Gadjah Mada, and Universiti Teknologi Malaysia, in 1999, 2004, and 2016, respectively. He
has been an Associate Professor at UAD, Yogyakarta, Indonesia since 2008. He is currently
the Editor-in-Chief of the Bulletin of Electrical Engineering and Informatics and the Head of
the Embedded Systems and Power Electronics Research Group. His research interests include
the fields of digital design, industrial applications, industrial electronics, industrial
informatics, power electronics, motor drives, renewable energy, FPGA applications,
embedded systems, artificial intelligence, intelligent systems, image processing, information
systems, and digital libraries. He can be contacted by email at [email protected].

Indah Soesanti received the M.Eng. and Ph.D. degrees from the Department of
Electrical Engineering, Universitas Gadjah Mada, Yogyakarta, Indonesia. She is currently a
Lecturer with the Department of Electrical and Information Engineering, Faculty of
Engineering, Universitas Gadjah Mada. Her research interests include artificial intelligence,
optimization, signal processing, image processing, information systems, ICT-based systems,
biomedical engineering, and pattern classification. She can be contacted at
[email protected].

Noor Akhmad Setiawan received his Bachelor and Master degree in Electrical
Engineering from Universitas Gadjah Mada in 1998 and 2003 respectively. He received his
PhD degree in Electrical and Electronics Engineering from Universiti Teknologi
PETRONAS in 2009. He is with the Department of Electrical and Information Engineering
Universitas Gadjah Mada. His research interest includes machine learning, soft computing,
data mining, big data, medical engineering and informatics, and electrical engineering. He is
a member of IEEE, ACM, IRSS, IAENG and IACSIT. He can be contacted by email at
[email protected].

A review of convolutional neural network-based computer-aided lung nodule detection system (Sekar Sari)