Journal of Experimental & Theoretical Artificial

Intelligence

ISSN: 0952-813X (Print) 1362-3079 (Online) Journal homepage: www.tandfonline.com/journals/teta20

Intensity-based statistical features for

classification of lungs CT scan nodules using
artificial intelligence techniques

Sheeraz Akram, Muhammad Younus Javed, Ayyaz Hussain, Farhan Riaz & M.
Usman Akram

To cite this article: Sheeraz Akram, Muhammad Younus Javed, Ayyaz Hussain, Farhan Riaz &
M. Usman Akram (2015) Intensity-based statistical features for classification of lungs CT scan
nodules using artificial intelligence techniques, Journal of Experimental & Theoretical Artificial
Intelligence, 27:6, 737-751, DOI: 10.1080/0952813X.2015.1020526

To link to this article: https://doi.org/10.1080/0952813X.2015.1020526

Published online: 19 Mar 2015.

Submit your article to this journal

Article views: 524

View related articles

View Crossmark data

Citing articles: 8 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=teta20
Journal of Experimental & Theoretical Artificial Intelligence, 2015
Vol. 27, No. 6, 737–751, http://dx.doi.org/10.1080/0952813X.2015.1020526

Intensity-based statistical features for classification of lungs CT scan

nodules using artificial intelligence techniques
Sheeraz Akrama*, Muhammad Younus Javeda, Ayyaz Hussainb,
Farhan Riaza and M. Usman Akrama
a
Department of Computer Engineering (DCE), College of Electrical and Mechanical Engineering
(CEME), National University of Science and Technology (NUST), Islamabad, Pakistan; bDepartment of
Computer Science & Software Engineering, Faculty of Basic & Applied Sciences, International Islamic
University, Islamabad, Pakistan
(Received 28 July 2014; accepted 2 November 2014)

A computer-aided diagnostic (CAD) system for effective and accurate pulmonary

nodule detection is required to detect the nodules at early stage. This paper proposed a
novel technique to detect and classify pulmonary nodules based on statistical features
for intensity values using support vector machine (SVM). The significance of the
proposed technique is, it uses the nodules features in 2D & 3D and also SVM for the
classification that is good to classify the nodules extracted from the image. The lung
volume is extracted from Lung CT using thresholding, background removal, hole-filling
and contour correction of lung lobe. The candidate nodules are extracted and pruned
using the rules based on ground truth of nodules. The statistical features for intensity
values are extracted from candidate nodules. The nodule data are up-samples to reduce
the biasness. The classifier SVM is trained using data samples. The efficiency of
proposed CAD system is tested and evaluated using Lung Image Consortium Database
(LIDC) that is standard data-set used in CAD Systems for Lungs Nodule classification.
The results obtained from proposed CAD system are good as compare to previous CAD
systems. The sensitivity of 96.31% is achieved in the proposed CAD system.
Keywords: computer-aided diagnostic (CAD) system; support vector machine (SVM);
computerised tomographic images; pulmonary nodules detection; statistical features
for intensity values; nodule classification

1. Introduction
The humans are suffering from different diseases. The cancer is the most dangerous of all
diseases. More number of people are suffering from lung cancer (Greenlee, Murray, Bolden, &
Wingo, 2000) and the number of people who died from lung cancer is high than any other
cancers (Jung et al., 2011). The survival rate of lung cancer patients can be increased by more
than 50% by early detection of the lung cancer, which is only 14% at present (Jung et al., 2011).
The survival rate has significantly improved, but there is need to increase this survival rate.
In order to detect lung cancer at an early stage, there is a need to have an inner view of the body.
In our case, we have the inner view of body using CT scan which provides 3D lung images. The
lung CT scans are not of good quality at the initial stage, so there is a need to enhance these
images so that the any lesion can be identified.

*Corresponding author. Email: [email protected]

q 2015 Taylor & Francis
7238 S. Akram et al.

The cancer is caused by different lesions that are produced in the different body parts. Such
lesions are referred to as nodule if they cause cancer, otherwise non-nodule. The main task in the
design of a CAD system is segmentation of the volume of a particular body part, such as lung
volume, to be separated from the complete image so that we can keep our focus on the object of
interest. The next task is to separate the objects from the lung volume which are not part of the
lungs. These objects are unwanted lesions. The unwanted lesions are potential nodules.
The potential nodules are classified into nodules and non-nodules which is the next task of the
CAD system.
The lung CT scans are observed by the experts for their opinion but the medical experts with
same expertise are not available at every place. The need of certain guidance of such medical
experts, the field of medical imaging introduced computer-aided diagnostic (CAD) systems
which help the medical specialist to identify and categorise the problem.
In this paper, this section gives an introduction, the next section is literature review related to
proposed technique, and the proposed method is discussed briefly in Section 3. Section 4 is
preprocessing to process the lung CT scans, Section 5 is candidate nodule and feature extraction
and nodule pruning, Section 6 is on Candidate Nodule up-sampling. Section 7 is Support Vector
Machine (SVM)-based Classification, Section 8 is on Results and Discussion and the last section
concludes the work.

2. Literature review
The lesions are detected automatically by scanning of the lung images. The various methods are
introduced to classify these lesions as nodules and non-nodules. In Ozekes (2007), the density
value of each pixel is calculated, the rule-based lung region segmentation is performed and the
Regions of Interest (ROIs) are extracted using 8-directional search. The classification is
performed by Location Change Measurement and the later nodules are searched using trained
Genetic Algorithm from the ROIs images. In the lung segmentation by the Genetic Cellular
Neural Network, the ROIs are extracted based on 8-directional search. The nodules are detected
by searching through 3D image with 3D template using convolution-based filter. The Fuzzy
Rule-Based Thresholding is used to further refine the detected nodules in Ozekes, Osman, and
Ucan (2008). In Xujiong Ye, Lin, Dehmeshki, Slabaugh, and Beddoe (2009), the lung
segmentation is performed, and the 3D nodules are extracted by anti-geometric diffusion,
volumetric shape features, Gaussian filtering and multi-scale dot enhancement filtering. The 3D
potential nodules are segmented. The 2D and 3D features are calculated, and the Rule-Based
filtering and weighted SVM are used for Nodule classification. In Retico et al. (2009), the
pleural regions are identified by directional-gradient concentration and morphological opening.
The ROIs are extracted from segmented pleura region. The features are extracted and candidate
nodules are classified using Feed-forward Neural Network. The region growing algorithm is
used to identify the lung parenchyma. The rolling-ball methodology is used to correct the
boundaries of pleura; the region growing algorithm is used again to identify the lung nodule.
The SVM is used to reduce the false positives in Da Silva Sousa, Silva, Paiva, and Nunes
(2010). In Lee, Kouzani, and Hu (2010), the nodule classification is achieved by classification
of training data-set into nodules and non-nodules using clustering. The nodules and non-nodules
obtained from clustering are used for training of SVM. This is ensemble classification aided by
clustering. In Maeda et al. (2011), the consecutive CT images are temporally subtracted to
detect candidate nodules. The features of candidate nodules are calculated and the candidates
are refined using rule-based feature analysis. The feature space is reduced using PCA. The
Artificial Neural network is used for nodule classification. The isotropic resampling of CT
 Journal of Experimental & Theoretical Artificial Intelligence 7393

image is performed to change the resolution of image. The lung is segmented using the
established techniques of segmentation. The nodule centre is estimated based on divergence of
normalised gradient. The multi-scale nodule and vessel enhancement filtering is used to
segment nodule clusters. The invariant, shape and regional descriptor are calculated and the mix
of ANN, GA (FD-NEAT) and SVM is used for feature selection and nodule classification (Tan,
Deklerck, Jansen, Bister, & Cornelis, 2011). In Choi and Choi (2012), the lung volume is
extracted by thresholding, contouring correction and morphological operation. The candidate
nodules are obtained by multiple thresholding technique from lung volume. The extracted
candidate nodules are pruned using rule-based pruning method. The rules are designed
according to the features of a lung nodule. The features are extracted from the candidate
nodules. The genetic programming classifier is trained and used for the classification of nodules
and non-nodules. Choi and Choi (2013) introduce the hierarchical block classification approach
using SVM for nodule classification. The CT image is split into blocks. The non-informative
blocks are discarded. The block image is enhanced and the object is segmented in the block
image. The location of the block is adjusted. The features are extracted from nodule candidate
block images. The SVM is used to classify candidate nodules as nodules and non-nodules.
In Tartar, Kilic, and Akan (2013), the features are selected from the images of nodules and non-
nodules. The features included 2D-PCA values with minimum Redundancy Maximum
Relevance, Statistical values of 2D-PCA with minimum Redundancy Maximum Relevance, and
2D-Geometric shape features with minimum Redundancy Maximum Relevance. The best
features are selected from the extracted features. The ANN, RF, Bagging and Adaboost are used
for training and testing. In El-Baz et al. (2013), the lung nodules, arteries, veins, bronchi and
bronchioles are isolated from the surrounding structures. The 2D and 3D deformable templates
are used to describe the geometry of nodules and the genetic optimisation algorithm is also used
in the detection. Later, the false positives are reduced from the detected nodules. In Rebouc
as
Filho, Cortez, and Albuquerque (2013), the segmentation in lung CT images is performed using
region growing, 3D region growing variations and multi-thresholding to segment the blood
vessels, lung emphysema and bones. The work done previously focuses separately on different
parts such as segmentation, feature extraction and classification. There is a need of work that
focuses on all parts at once place.

3. Proposed method
In this paper, the classification of nodules using SVM is proposed. The block diagram is given
in Figure 1. The proposed classification consists of the following steps. First, the Lung CT Scan
is processed to obtain the lung lobe region. In next step, the candidate nodules are obtained and
pruning is performed to reduce the false positive. Later, the features are extracted from
candidate nodules and the required features are selected. The number of candidate nodules is
increased by up-sampling and SVM classifier is used for the classification of nodules and
non-nodules. The proposed method improves the accuracy of nodule classification. The
proposed method is evaluated on Lung Image Database Consortium Database (LIDC) (Reeves
et al., 1997).

4. Preprocessing
The lung CT image contains values in HU (Hounsfield units). The Lung CT scans are 3D
images. Each image contains slices ranges from 100 to 250. The size of each slice is 512 rows
and 512 columns. The lung CT scans are preprocessed using steps given next.
7440 S. Akram et al.

Figure 1. Proposed pulmonary nodule detection system’s block diagram.

4.1 Thresholding
The HU values in each Lung CT scan range from 2 2000 to þ 2000 HU. The lung area is a low
density area ranging from 2 1000 to 2 450 HU, called non-body area. The CT scanner area is
also part of the non-body area of Lung CT scan. The body area contains the surrounding of lung
lobes. The lungs are in the non-body area, so threshold it at 2 500 HU (Brown et al., 1997; Hu,
Hoffman, & Reinhardt, 2001; Jemal et al., 2009). The voxels value below 2 500 HU contains
lung volume, and voxels values above þ 500 HU contain the body volume. Figure 2(b) shows
the result of thresholding.
If voxel value is less than 2 500 HU then
Voxel value ¼ minimum HU value in lung CT
Else
Voxel value ¼ maximum HU value in lung CT

4.2 Background removal

The Thresholded Lung CT scan contains body and non-body area. The black area is body voxels
and the white area is non-body voxels. The non-body area contains lung lobes and the CT
 Journal of Experimental & Theoretical Artificial Intelligence 741
5

Figure 2. Lung volume segmentation.

scanner producing a cylinder around the lungs and the body area. There is a need to remove the
cylinder. The 3D connected component approach is applied to make all the component
boundaries accurate (Messay, Hardie, & Rogers, 2010; Suárez-Cuenca et al., 2009). The non-
body component touching the sides of Lung CT image is removed and the voxel values are set
to background values, i.e. the value of the body voxels. Figure 2(c) shows the Background
Removed Lung CT image as a result of Background Removal.

4.3 Hole filling of lung lobes

The background-removed lung CT image obtained in the previous step contains holes in the lung
lobe, that are either potential nodules or vessels. It is important to include them in the lung
lobe area. The hole-filling morphological operators are used to fill the holes (Paik et al., 2004;
Suárez-Cuenca et al., 2009). Figure 2(d) shows the hole-filled lung CT image as a result of the
hole-filling process.

4.4 Contour correction of lung image

The hole-filled lung CT image obtained may contain nodules or vessels at the boundary of lung
lobe known as juxta-pleural. These juxta-pleural need to be included in the lung lobe area.
Initially, the lung contour is obtained and corrected using chain codes (Jusoh & Zain, 2009).
The critical points are detected using differential chain coding. The required critical points
are connected through straight lines in order to include the critical part at the contour of the
lung lobe. Figure 2(e) shows contour-corrected lung CT images as a result of contour-corrected
image.

5. Candidate nodules and feature extraction

The preprocessing step produces the lung volume. This lung volume contains the candidate
nodules. The candidate nodules are extracted, pruned and the features are extracted from pruned
candidate nodules.

5.1 Candidate nodule extraction

The preprocessing step produced the 3D lung mask that is used to extract the lung volume from
the Lung CT scan. The lung volume contains both vessels and nodules. The density of nodules,
vessels and lungs is different from each other. In other words, the nodules and vessels are denser
than the lungs. The lung volume is extracted from the Lung CT image using a lung mask. The
optimal thresholding method is used to extract the ROIs. The threshold is calculated on the
7642 S. Akram et al.

median slice. Multiple threshold values are calculated as the vessels and nodules have different
densities depending on the type of nodule. These ROIs contain both nodules and vessels.

5.2 Pruning of candidate nodule

The nodules in the data-set have diameter ranges from 3 to 30 mm. The ROIs having
diameter smaller than 3 mm are ignored as noise and the ROIs having diameter greater than
30 mm are pruned as lesion/vessels. The property of elongation is used to detect the vessels in
ROIs. The pruned nodules present in the 3D lung images are mapped onto 2D image, shown
in Figure 3.

5.3 Features extraction from pruned candidate nodules

The candidate nodules extracted can be nodules and non-nodules. These candidate nodules are
3D object. The following features are extracted as given in Table 1.

Figure 3. Candidate nodules extracted.

Table 1. List of features extracted.

2D Statistical features for intensity values 3D Statistical features for intensity values
Minimum value inside Minimum value inside
Mean inside Mean inside
Mean outside Mean outside
Variance inside Variance inside
Skewness inside Skewness inside
Kurtosis inside Kurtosis inside
Eigen values (8)
 Journal of Experimental & Theoretical Artificial Intelligence 743
7

The 2D statistical features for intensity values are extracted from the median slice of the
segmented object. The 3D statistical features for intensity values are extracted from the 3D
segmented object.
The statistical features for intensity values are calculated as given in Equations (1) –(4)
Pn
 ¼ i¼1 X i
Mean ðXÞ ; ð1Þ
n

Pn
2 X 2 Þ2
i¼1 ðX i
Variance ðs 2 Þ ¼ ; ð2Þ
n21

Pn
i¼1 ðX i
2 XÞ 4
Kurtosis ¼ ; ð3Þ
ðn 2 1Þs 4

Pn
i¼1 ðX i
2 XÞ  3
Skewness ¼ : ð4Þ
ðn 2 1Þ 3

6. Candidate nodule up-sampling

The Lung Image Database Consortium (LIDC) is a standard data-set that is used to develop and
test the Lung CAD systems. It consists of 84 CT Scans of different patients. The 47 CT scans that
contain nodules are used in the system. The nodules are obtained and pruned. The number of
nodules and non-nodules is not the same. This creates an unbalanced data-set for classification.
The appropriate number of non-nodules from the data-set is selected and then the nodules are up-
sampled by repeating the candidate nodules in order to make the data-set balanced, i.e. the
number of nodules and non-nodules is equal.

7. Support vector machine based classification

The SVM produces good results for binary classification for image data. The Gaussian Radial
Basis Function kernel is used to train the SVM classifier. There is no biasness in the data, i.e. the
number of nodules and non-nodules is equal in both training data and testing data. The candidate
nodules are classified into nodules and non-nodules. The data-set has 2D Statistical Features for
Intensity Values and 3D Statistical Features for Intensity Values. The features are divided into
the following sets:
. 2D statistical features for intensity values
. 3D statistical features for intensity values
. 2D and 3D statistical features for intensity values.
The data-set is divided into the following for training and testing purposes:
. 30– 70, The 30% is used for training and 70% is used for testing
. 50– 50, The 50% is used for training and 50% is used for testing
. 70– 30, The 70% is used for training and 30% is used for testing.
All the candidate nodules are categorised into True Positive, True Negative, False Positive,
and False Negative as given in Table 2.
7844 S. Akram et al.

Table 2. Terms of nodule categorisation.

Determined from standard truth

Result of CAD system Nodule Non-nodule Total (row)
Nodule True positive (TP) False positive (FP) TP þ FP
Non-nodule False negative (FN) True negative (TN) FN þ TN
Total (column) TP þ FN FP þ TN TP þ FP þ FN þ TN

Table 3. Results of 2D statistical features for intensity values.

Training-testing Accuracy (%) Sensitivity (%) Specificity (%) AUC
30 – 70 87.49 89.95 85.03 0.9853
50 – 50 93.37 94.92 91.81 0.9967
70 – 30 95.06 97.5 92.63 0.9981

The accuracy, sensitivity, specificity and AUC (Area under Receiver Operating Curve [ROC
curve]) are measured as given in Equations (5) – (7).

TN þ TP
Accuracy ¼ ; ð5Þ
TN þ TP þ FN þ FP

Figure 4. 2D Statistical features for intensity values ROC curves.

 Journal of Experimental & Theoretical Artificial Intelligence 745
9

Table 4. Results of 3D statistical features for intensity values.

Training-testing Accuracy (%) Sensitivity (%) Specificity (%) AUC

30 – 70 80.33 82.23 78.43 0.9841
50 – 50 84.90 87.65 82.15 0.9941
70 – 30 86.57 91.15 81.99 0.9967

TP
Sensitivity ¼ ; ð6Þ
TP þ FN

TN
Specificity ¼ : ð7Þ
TN þ FP

8. Results and discussion

The standard data-set LIDC is used for training and validation purpose. The Lung CT scan is a
3D lung image. Each slice is of size 512 £ 512. Each lung CT scan has around 150 slices and
4096 grey-level values in HU. The pixel size in the database ranges from 0.50 to 0.76 mm, and
the reconstruction interval ranges from 1 to 3 mm. The LIDC data-set is publically available.
The data is up-sampled to minimise the biasness of the classifier. The results of the Gaussian
radial basis function kernel SVM classifier for 2D Statistical Features for Intensity Values, 3D
Statistical Features for Intensity Values and statistical features for intensity values for both 2D &
3D are shown in the tables. The ROC curves are also drawn for each set of features.

Figure 5. 3D Statistical features for intensity values of ROC curves.

746
10 S. Akram et al.

Table 5. Results of 2D and 3D statistical features for intensity values.

Training-testing Accuracy (%) Sensitivity (%) Specificity (%) AUC

30 – 70 89.16 83.65 94.67 0.9779
50 – 50 96.54 96.31 96.77 0.9967
70 – 30 98.30 99.10 97.50 0.9971

Table 3 shows the result of 2D statistical features for intensity values. With 50– 50 training
testing ratio, 93.37% accuracy, 94.92% sensitivity and 91.81% specificity are achieved. Figure 4
shows the ROC curve for 2D statistical features for intensity values for 30– 70, 50 –50 and 70–
30 training – testing ratios. The AUC is highest for the 70 –30 training – testing ratio. Table 4
shows the result of the 3D statistical features for intensity values. With 50– 50 training – testing
ratio, 84.90% accuracy, 87.65% sensitivity and 82.15% specificity are achieved. Figure 5 shows
the ROC curve for 3D statistical features for intensity values for 30– 70, 50 – 50 and 70– 30
training –testing ratios. The AUC is highest for the 70– 30 training –testing ratio. Table 5 shows
the result of 2D and 3D statistical features for intensity values. With 50 –50 training testing ratio,
96.54% accuracy, 96.31% sensitivity and 96.77% specificity are achieved. Figure 6 shows the
ROC curve for 2D and 3D statistical features for intensity values for 30– 70, 50 –50 and 70– 30
training –testing ratios. The AUC is highest for the 70– 30 training – testing ratio.
Figures 7 and 8 show the scatter graphs of 2D statistical features for intensity values and 3D
statistical features for intensity values.
Suzuki, Armato, Li, Sone, and Doi (2003) worked for nodules of size 8– 20 mm with
sensitivity 80.3%. Rubin et al. (2004) worked for nodule size $ 3 with sensitivity of 76%.
Dehmeshki, Ye, Lin, Valdivieso, and Amin (2007) worked for nodule size 3– 20 mm with

Figure 6. 2D and 3D Statistical features for intensity values of ROC curves.

 Journal of Experimental & Theoretical Artificial Intelligence 747
11

Figure 7. Feature space of 2D statistical features for intensity values.

sensitivity of 90%. Suárez-Cuenca et al. (2009) worked for nodule sizes 4 –27 mm with
sensitivity of 80%. Opfer and Wiemker (2007) worked for nodule size $ 4 mm with sensitivity
of 74%. Sahiner et al. (2007) worked for nodule size 3– 36.4 mm with sensitivity of 79%.
Messay, Hardie, and Rogers (2010) worked for nodule size 3– 30 mm with sensitivity of
748
12 S. Akram et al.

Figure 8. Feature space of 3D statistical features for intensity values.

82.66%. Choi and Choi (2012) worked for nodule size 3 – 30 mm with sensitivity of 94.1%. Choi
and Choi (2013) worked for nodule size 3 –30 mm with sensitivity of 95.28%. The Proposed
Method work for nodule size 3 –30 mm with sensitivity of 96.31%, that is better than the earlier
 Journal of Experimental & Theoretical Artificial Intelligence 749
13

Table 6. Performance comparison of different CAD systems.

CAD systems Nodule size (mm) Sensitivity (%)

Suzuki et al. (2003) 8 – 20 80.3
Rubin et al. (2004) $3 76
Dehmeshki, Ye, Lin, Valdivieso, and Amin (2007) 3 – 20 90
Suárez-Cuenca et al. (2009) 4 – 27 80
Opfer and Wiemker (2007) $4 74
Sahiner et al. (2007) 3 – 36.4 79
Messay et al. (2010) 3 – 30 82.66
Choi and Choi (2012) 3 – 30 94.1
Choi and Choi (2013) 3 – 30 95.28
Proposed method 3 – 30 96.31

techniques. The performance comparison of earlier CAD systems and the proposed CAD system
is represented in Table 6.

9. Conclusion
In this paper, a novel technique based on intensity-based statistical feature using SVM for
automatic pulmonary nodule detection is presented. The thresholding, background removal,
hole-filling and contour correction is performed to extract the lung volume. The candidate
nodules are extracted from the lung volume. The candidate nodules are pruned using rules based
on information of nodules. The intensity-based statistical 2D and 3D features are extracted. The
SVM classifier is trained using nodule samples from the data-set. The classifier is evaluated by
selecting nodules from extracted nodules from data-set of LIDC. The classifier achieves the
sensitivity of 96.31% with accuracy of 96.54% that is improved than the existing CAD systems.
In future work, the sensitivity and accuracy can be further improved by reducing the false
positives by improving candidate nodule pruning algorithm.

Acknowledgement
This work is not financially supported by any funding agency.

Disclosure statement
No potential conflict of interest was reported by the authors.

References
Brown, M. S., McNitt-Gray, M. F., Mankovich, N. J., Goldin, J. G., Hiller, J., Wilson, L. S., & Aberie, D. R.
(1997). Method for segmenting chest CT image data using an anatomical model: Preliminary results.
IEEE Transactions on Medical Imaging, 16, 828– 839. doi:10.1109/42.650879
Choi, W. J., & Choi, T. S. (2012). Genetic programming-based feature transform and classification for the
automatic detection of pulmonary nodules on computed tomography images. Inform Sciences, 212,
57 – 78. doi:10.1016/j.ins.2012.05.008
Choi, W. J., & Choi, T. S. (2013). Automated pulmonary nodule detection system in computed tomography
images: A hierarchical block classification approach. Entropy, 15, 507– 523. doi:10.3390/
e15020507
750
14 S. Akram et al.

Da Silva Sousa, J. R. F. S., Silva, A. C., De Paiva, A. C., & Nunes, R. A. (2010). Methodology for automatic
detection of lung nodules in computerized tomography images. Computer Methods and Programs in
Biomedicine, 98(1), 1 – 14. doi:10.1016/j.cmpb.2009.07.006
Dehmeshki, J., Ye, X., Lin, X., Valdivieso, M., & Amin, H. (2007). Automated detection of lung nodules in
CT images using shape-based genetic algorithm. Computerized Medical Imaging and Graphics, 31,
408– 417. doi:10.1016/j.compmedimag.2007.03.002
El-Baz, A., Elnakib, A., El-Ghar, M. A., Gimel’farb, G., Falk, R., & Farag, A. (2013). Automatic detection
of 2D and 3D lung nodules in chest spiral CT scans. International Journal of Biomedical Imaging,
2013, 1 – 11.
Greenlee, R. T., Murray, T., Bolden, S., & Wingo, P. A. (2000). Cancer statistics, 2000. CA: A Cancer
Journal for Clinicians, 50, 7 – 33. doi:10.3322/canjclin.50.1.7
Hu, S., Hoffman, E. A., & Reinhardt, J. M. (2001). Automatic lung segmentation for accurate quantitation
of volumetric X-ray CT images. IEEE Transactions on Medical Imaging, 20, 490– 498. doi:10.1109/
42.929615
Jemal, A., Siegel, R., Ward, E., Hao, Y., Xu, J., & Thun, M. J. (2009). Cancer statistics, 2009. CA:
A Cancer Journal for Clinicians, 59, 225– 249. doi:10.3322/caac.20006
Jung, K. W., Won, Y. J., Park, S., Kong, H. J., Sung, J., Shin, H. R., . . . Lee, J. S. (2011). Cancer statistics
in Korea: Incidence, mortality and survival in 2005. Journal of Korean Medical Science, 43, 1 –11.
Jusoh, N. A., & Zain, J. M. (2009). Application of freeman chain codes: An alternative recognition
technique for Malaysian car plates. International Journal of Computer Science and Network
Security, 9, 222– 227.
Lee, S. L. A., Kouzani, A. Z., & Hu, E. J. (2010). Random forest based lung nodule classification aided by
clustering. Computerized Medical Imaging and Graphics, 34, 535–542. doi:10.1016/j.compmedi-
mag.2010.03.006
Maeda, S., Tomiyama, Y., Kim, H., Miyake, N., Itai, Y., Tan, . . . Yamamoto, A. (2011). Detection of lung
nodules in thoracic MDCT images based on temporal changes from previous and current images.
Journal of Advanced Computational Intelligence and Intelligent Informatics, 15, 707– 713.
Messay, T., Hardie, R. C., & Rogers, S. K. (2010). A new computationally efficient CAD system for
pulmonary nodule detection in CT imagery. Medical Image Analysis, 14, 390– 406. doi:10.1016/j.
media.2010.02.004
Opfer, R., & Wiemker, R. (2007, March). Performance analysis for computer-aided lung nodule detection
on LIDC data. In Medical imaging 2007: Image perception, observer performance, and technology
assessment, Vol. 6515 of Proceedings of the SPIE, San Diego, CA, 65151C.
Ozekes, S. (2007). Rule-based lung region segmentation and nodule detection via Genetic Algorithm
trained template matching. Istanbul Commerce University Journal of Science, 6, 17 –30.
Ozekes, S., Osman, O., & Ucan, O. N. (2008). Nodule detection in a lung region that’s segmented with
using genetic cellular neural networks and 3D template matching with fuzzy rule based thresholding.
Korean Journal of Radiology, 9(1), 1 – 9. doi:10.3348/kjr.2008.9.1.1
Paik, D. S., Beaulieu, C. F., Rubin, G. D., Acar, B., Jeffrey, Jr, R. B., Yee, J., Dey J. & Napel, S. (2004).
Surface normal overlap: A computer-aided detection algorithm with application to colonic polyps
and lung nodules in helical CT. IEEE Transactions on Medical Imaging, 23, 661– 675. doi:10.1109/
TMI.2004.826362
Rebouc
as Filho, P. P., Cortez, P. C., & de Albuquerque, V. H. C. (2013). 3D segmentation and visualization of
lung and its structures using CT images of the thorax. Journal of Biomedical Science and Engineering,
6, 1099– 1108. doi:10.4236/jbise.2013.611138
Reeves, A. P., Biancardi, A. M., Apanasovich, T. V., Meyer, C. R., MacMahon, H., Beek, E. J., . . . Clarke,
L. P. (1997). The Lung Image Database Consortium (LIDC): A comparison of different size metrics
for pulmonary nodule measurements. Academic Radiology, 14, 1475– 1485.
Retico, A., Fantacc, M. E., Gori, I., Kasae, P., Golosio, B., Piccioli, A., . . . Tangaro, S. (2009). Pleural
nodule identification in low-dose and thin-slice lung computed tomography. Computers in Biology
and Medicine, 39, 1137– 1144.
 Journal of Experimental & Theoretical Artificial Intelligence 751
15

Rubin, G. D., Lyo, J. K., Paik, D. S., Sherbondy, A. J., Chow, L. C., Leung, A. N., . . . Napel, S. (2004).
Pulmonary nodules on multi-detector row CT scans: Performance comparison of radiologists and
computer-aided detection. Radiology, 234, 274– 283.
Sahiner, B., Hadjiiski, L. M., Chan, H., Shi, J., Cascade, P. N., Kazerooni, E. A., . . . Poopat, C. (2007,
March). Effect of CAD on radiologists’ detection of lung nodules on thoracic CT scans: Observer
performance study. In Proceedings of SPIE 6515, medical imaging 2007: Image perception,
observer performance, and technology assessment, Vol. 6515 of Proceedings of the SPIE, San
Diego, CA, 65151D.
Suárez-Cuenca, J. J., Tahoces, P. G., Souto, M., Lado, M. J., Remy-Jardin, M., Remy, J., & Vidal, J. J.
(2009). Application of the iris filter for automatic detection of pulmonary nodules on computed
tomography images. Computers in Biology and Medicine, 39, 921– 933.
Suzuki, K., Armato, S. G., Li, F., Sone, S., & Doi, K. (2003). Massive training artificial neural network
(MTANN) for reduction of false positives in computerized detection of lung nodules in low-dose
computed tomography. Medical Physics, 30, 1602– 1617. doi:10.1118/1.1580485
Tan, M., Deklerck, R., Jansen, B., Bister, M., & Cornelis, J. (2011). A novel computer-aided lung nodule
detection system for CT images. Medical Physics, 38, 5630 –5645. doi:10.1118/1.3633941
Tartar, A., Kilic, N., & Akan, A. (2013). Classification of pulmonary nodules by using hybrid features.
Computational and Mathematical Methods in Medicine, 2013, 1 – 11. doi:10.1155/2013/148363
Xujiong, Y., Xinyu, L., Dehmeshki, J., Slabaugh, G., & Beddoe, G. (2009). Shape-based computer-aided
detection of lung nodules in thoracic CT images. IEEE Transactions on Biomedical Engineering, 56,
1810– 1820. doi:10.1109/TBME.2009.2017027