A Support Vector Machine based Prediction Model for

Discrimination of Malignant Pulmonary Nodules from Benign

Nodules

Yan Wu

1,2

, Emmanuel Zachariah

, Judith K. Amorosa

, Anjani Naidu

, Mina L. Labib

Jamil Shaikh

, Donna Eckstein

, Sinae Kim

, John E. Langenfeld

, Joseph Aisner

, John L. Nosher

Robert S. DiPaola

and David J. Foran

Rutgers Cancer Institute of New Jersey, Rutgers, The State University of New Jersey,

195 Little Albany Street, New Brunswick, NJ 08903, U.S.A.

Department of Radiology, Rutgers Robert Wood Johnson Medical School, Rutgers, The State University of New Jersey,

1 Robert Wood Johnson Place, New Brunswick, NJ 08901, U.S.A.

Department of Biostatistics, Rutgers School of Public Health, Rutgers, The State University of New Jersey,

683 Hoes Lane West, Piscataway, NJ 08854, U.S.A.

Keywords: Support Vector Machine, Malignant Nodules, Benign Nodules, Pulmonary Nodules, Prediction Model.

Abstract: Lung cancer is the leading cause of cancer death in the United States and worldwide. Most patients are

diagnosed at an advanced stage, usually stage III or IV. Identification of lung cancer patients at an early

stage might enable oncologists to surgically remove the tumors. Currently, low dose CT scans are used to

identify the malignant nodules in high risk patients. However, screening CT scans yield a high rate of false-

positive results. A prediction model was developed for improved discrimination of malignant nodules from

benign nodules in patients who underwent lung screening CT. CT images and clinical outcomes of 39

patients were obtained from the National Lung Screening Trial (NLST), National Cancer Institute, National

Institute of Health. Images were analyzed to extract computational features relevant to malignancy

prediction. A Support Vector Machine (SVM) based model was developed to predict the malignancy of

nodules. During pilot studies, our model achieved the following prediction performance: accuracy of 0.74,

sensitivity of 0.85, and specificity of 0.61.

1 INTRODUCTION

Lung cancer is the leading cause of cancer death in

the United States and worldwide. The mortality rate

from lung cancer is greater than the number of

deaths from breast, colon, and prostate cancer

combined. In 2015, there will be an estimated

226,000 new cases of lung cancer diagnosed in the

United States and over 160,000 individuals are

expected to die from this disease (Cancer Facts and

Figures 2015). Most patients will be diagnosed with

locally advanced (stage III) or metastatic (stage IV)

disease and the expected 5‐year survival rate is only

~15% (Sozzi and Boeri 2014). Adenocarcinoma, the

most common form of lung cancer, presents as a

solitary pulmonary nodule which is easier to detect

on CT than on a chest X-ray. Identification of

patients at an early stage could potentially enable

oncologists to surgically remove the tumors.

Presently, there is no simple screening protocol

for lung cancer that yields discriminatory results

similar to those realized in breast (mammogram) and

colon cancers (colonoscopy). The National Lung

Screening Trial (NLST), initiated during 2002,

established the potential utility of low dose CT

(LDCT) screening to reduce lung cancer specific

mortality in the high-risk population of current and

former smokers (55–74 years of age, cigarette

smokers with a history of at least 30 pack-years, and

former smokers who quit smoking within the past 15

years). The data showed that there was a significant

reduction (20%) in the death rates from lung cancer

in participants who had LDCT compared to

participants who had standard chest X-ray (National

Lung Screening Trial Research Team, 2011);

however, LDCT scans yield a high rate of false-

positive results (National Lung Screening Trial

Wu, Y., Zachariah, E., Amorosa, J., Naidu, A., Labib, M., Shaikh, J., Eckstein, D., Kim, S., Langenfeld, J., Aisner, J., Nosher, J., DiPaola, R. and Foran, D.

A Support Vector Machine based Prediction Model for Discrimination of Malignant Pulmonary Nodules from Benign Nodules.

DOI: 10.5220/0005824101290133

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 2: BIOIMAGING, pages 129-133

ISBN: 978-989-758-170-0

129

Research Team, 2011), (Wood et al., 2012),

(Arenberg and Kazerooni 2012). In the NLST CT

group, almost 40% of participants had at least 1

positive CT result during the study, but more than

96% of the positive test results in the CT group of

the NLST were false-positive. Consequently, it may

lead to unnecessary and costly biopsies or follow up

Positron emission tomography–computed

tomography (PET-CT) scans or follow up CT.

RECIST (Response Evaluation Criteria In Solid

Tumors) criteria is being used to identify and follow

nodules in lung cancer patients (Eisenhauer et al.,

2009). However, it has limited capability to

distinguish malignant nodules from benign lesions.

In our study, we developed and tested a Support

Vector Machine (SVM) based risk prediction model

to investigate it performance in discriminating

between malignant and benign nodules in

asymptomatic high risk patients.

2 MATERIAL AND METHODS

2.1 CT Scan Image Analysis

CT images from 39 patients who participated in the

NLST were obtained from National Cancer Institute

(NCI), National Institute of Health. A trial-wide

database is available for download at

https://biometry.nci.nih.gov/cdas/, Cancer Data

Access Systems (CDAS), NCI. The data set for each

patient acquired at each year includes 100–300 slices

of dicom images. An approval was obtained from

Rutgers Internal Review Board to process the

images.

Images were analyzed using software developed

at the Center for Biomedical Imaging and

Informatics, Rutgers Cancer Institute of New Jersey

(http://pleiad.umdnj.edu). For each patient, a

pulmonary nodule was first segmented on the Vitrea

workstation (Vitrea 2; Vital Images, Plymouth,

MN). Within the tumor region, texture analysis was

performed using the Local Binary Pattern (LBP)

method (Ojala et al., 1996). The local binary pattern

of a single pixel is determined by its signal intensity

relative to its neighbors (if a neighbor has signal

intensity higher than that of the central pixel, then 1”

is assigned; otherwise “0” is assigned). In this way, a

binary sequence is generated for the pixel, indicating

the relative variation in its signal intensity compared

with its neighbors. In the binary sequence, the

number of consecutive “1”s is counted as the LBP

for the pixel, if all "1”s are consecutive; otherwise,

the LBP is set as N+1 and not differentiated further.

With local binary pattern of each pixel obtained, the

histogram of LBP is collected within the tumor

region and normalized to eliminate the influence of

tumor size. Although the local binary pattern can be

measured at different scales (to capture the texture

characteristics from fine to coarse), a single scale

LBP was measured in this study due to the small

size of some pulmonary nodules, where a small

radius (1 pixel, 8 neighbors) was used. For each

nodule, the volume of the nodule and its histogram

of signal intensity were also calculated

automatically.

Other features were specified by diagnostic

radiologists in the Department of Radiology, Rutgers

Robert Wood Johnson Medical School. Those

features include the location of nodules (left, right;

superior, middle, inferior), the margin of nodules

(smooth, lobular, irregular, spiculated), the shape of

nodules (round, oval, complex), the attenuation of

nodules (soft tissue, ground glass, mixed, solid,

calcium), the attachment of nodules (none, fissural,

pleural), and whether the patient has emphysema.

All the features specified or extracted from images

were used as the input to statistical prediction model.

2.2 SVM based Risk Prediction Model

The prediction of malignancy of nodules was

accomplished by establishing and using the SVM

model. Support Vector Machine is a supervised

learning algorithm (Cortes and Vapnik, 1995). The

SVM classifier was established during the training

procedure based on the clinical outcomes (malignant

vs benign) and image features (extracted from CT

images or specified by radiologists as described

above). The optimal decision boundary was

determined by maximizing the distance in features

between two different classes. After a SVM

prediction model was established, the prediction of

tumor progression of a new patient could be

automatically made based on his own image

features. The prediction was compared with the

clinical outcome (truth) for evaluation of accuracy,

sensitivity, and specificity. Due to the limited

sample size, we adopted leave-one-out approach to

build the prediction model.

In the SVM classifier developed for this study,

soft margin was used to allow for mislabeled

samples, where training errors were incorporated

into the cost function, and optimization became a

tradeoff between a large margin and a small error

penalty controlled by a parameter C. By using a

nonlinear kernel, the SVM classifier permitted

nonlinear decision boundary that fits data more

BIOIMAGING 2016 - 3rd International Conference on Bioimaging

130

closely. When Gaussian Radial Basis function was

used (as one of the most popular kernels), a

parameter sigma determines how smoothly input

features varied. The choice of parameters (C for soft

margin and sigma for GRB) was determined by

cross validation – the parameters with the best cross

validation accuracy were selected during the training

procedure and would be used for future predictions

during the test procedure.

3 RESULTS

CT images obtained from 39 patients were analyzed

in the present study. Pulmonary nodules were first

segmented by diagnostic radiologists. Then texture

features were automatically calculated using the

Local Binary Pattern method, and 10D local binary

histograms were generated, as shown in Figure 1.

Values were assigned for other imaging features

(including tumor location, attachment, shape,

attenuation, margin, and emphysema) depending on

its possibility of malignancy, as shown in Table 1.

The SVM prediction model was used with RBF

(radius basis function) kernel and soft margin. The

prediction result is shown as follows:

Accuracy= 0.74, sensitivity = 0.85, specificity =

0.61.

Figure 1: Tumor segmentation and texture feature

extraction.

Table 1: Image features specified by radiologists.

Location

Right Upper Lobe

1.0

Right Middle Lobe

1.5

Right Lower Lobe

2.0

Left Lower Lobe

1.1

Left Upper Lobe

2.1

Attachment

None

Fissure

Pleural

Shape

Oval

Round

Complex

Attenuation

Soft Tissue

Ground glass

Mixed

Solid Density

Margins

Smooth

Lobulated

Irregular

Spiculated

Calcification

None

Central

Diffuse

Emphasyma

None

Yes

4 DISCUSSION

Several lung cancer risk models have been

established during the past ten years (Tammemagi,

2015). Risk factors were used to predict the

probability that a person was likely to develop lung

cancer (Bach et al., 2003), (Spitz et al., 2007),

(Cassidy et al., 2008). More recently, radiology

image features were incorporated to predict

malignancy of pulmonary nodules (Sluimer et al.,

2006) (Maldonado et al., 2013). In these studies,

logistic regression model was widely used.

However, in several recent clinical studies and

simulations, SVM demonstrated better performance

in various classiﬁcation applications than the logistic

regression model (Entezari-Maleki et al. 2009),

(Wang and Huang 2011), (Salazar et al., 2012). In

the present pilot study, we developed a SVM based

model to discriminate malignant nodules from

benign nodule.

While we recognize that the present study

utilized a relatively small number of subjects,

incorporation of CT scan data from more subjects

could potentially increase the prediction accuracy of

our present standalone model. In addition, we

hypothesize that the incorporation of molecular

biomarkers with the CT scan image data for both

risk assessment and early diagnosis will further

improve the diagnostic accuracy of LDCT (Rutman

A Support Vector Machine based Prediction Model for Discrimination of Malignant Pulmonary Nodules from Benign Nodules

131

and Kuo 2009), (Zander et al., 2011), (Gevaert et al.,

2012), (Tammemagi, et al., 2013). There are a few

reports on the integration of genomic data with CT

scan image data (Showe et al., 2009) however, there

is no commercial test available using the integrated

approach to discriminate nodules in lung cancer.

Further studies are in progress to analyze CT scan

images and genomic data collected from high risk

individuals.

ACKNOWLEDGEMENTS

This research was funded, in part, by grants from the

National Institute of Health through contract

5R01CA161375-03 from the National Cancer

Institute; and contract 5R01LM009239-06 from the

National Library of Medicine (D.F.).

COMPETING INTEREST

E.Z. holds position and shares in OncoPath

Genomics, Inc.

REFERENCES

American Cancer Society. (2015) Cancer Facts and

Figures. http://www.cancer.org/research/cancerfacts

statistics/cancerfactsfigures2015/ Accessed on

10/10/2015.

Arenberg, D., and Kazerooni, E.A., (2012). Setting up a

lung cancer screening program. Journal of the

National Comprehensive Cancer Network, 10(2):277–

285.

Bach, P.B., Kattan, M.W., Thornquist, M.D., Kris, M.G.,

Tate, R.C., Barnett, M.J., Hsieh, L.J., and Begg, C.B.

(2003).Variations in lung cancer risk among smokers.

Journal of the National Cancer Institute, 95(6): 470–

478.

Cassidy, A., Myles, J.P., Van-Tongeren, M., Page, R.D.,

Liloglou, T., Duffy, S.W. and Field, J.K. (2008). The

LLP risk model: an individual risk prediction model

for lung cancer. British Journal of Cancer, 98(2):270–

276.

Cortes, C., and Vapnik, V., (1995). Support-vector

networks. Machine Learning, 20:273-297.

Eisenhauer, E.A., Therasse, P., Bogaerts, J., et al. (2009).

New response evaluation criteria in solid tumors:

Revised RECIST guideline (version 1.1). European

Journal of Cancer, 45:228-247.

Entezari-Maleki, R., Rezaei, A., a nd Minaei-Bidgoli, B.,

(2009). Comparison of Classification Methods Based

on the Type of Attributes and Sample Size. Journal of

Convergence Information Technology, 4(3):94-102.

Gevaert, O., Xu, J., Hoang, C.D., Leung, A.N., Xu, Y.,

Quon, A., Rubin, D.L., Napel, S., and Plevritis, S.K.

(2012). Non–Small Cell Lung Cancer: Identifying

prognostic imaging biomarkers by leveraging public

gene expression microarray data-Methods and

preliminary results. Radiology, 264(2):387-396.

Maldonado, F., Boland, J.M., Raghunath, S., et al. (2013).

Non-invasive Characterization of the Histopathologic

Features of Pulmonary Nodules of the Lung

Adenocarcinoma Spectrum using Computer Aided

Nodule Assessment and Risk Yield (CANARY) – a

Pilot Study. Journal of Thoracic Oncology, 8(4): 452-

460.

National Lung Screening Trial Research Team. (2011).

Reduced lung-cancer mortality with low-dose

computed tomographic screening. New England

Journal of Medicine, 365(5):395-409.

Ojala, T., Pietikainen, M., and Harwood, D., (1996). A

comparative study of texture measures with

classification based on featured distributions. Pattern

recognition, 29(1):51-59.

Rutman, A.M., and Kuo, M.D., (2009). Radiogenomics:

creating a link between molecular diagnostics and

diagnostic imaging. European Journal of Radiology,

70(2):232–241.

Salazar, D.A., Velez, J.I., and Salazar, J.C., (2012).

Comparison between SVM and Logistic Regression:

Which One is better to Discriminate? Revista

Colombiana de Estadística Número especial en

Bioestadística, 35:223-237.

Showe, M.K., Vachani, A., Kossenkov, A.V., et al.

(2009). Gene Expression Profiles in Peripheral Blood

Mononuclear Cells Can Distinguish Patients with

Non-Small-Cell Lung Cancer from Patients with Non-

Malignant Lung Disease. Cancer Research,

69(24):9202–9210.

Sluimer, I., Schilham, A., Prokop, M., and Van-Ginneken,

B. (2006). Computer Analysis of Computed

Tomography Scans of the Lung: A Survey. IEEE

Transactions on Medical Imaging, 25(4):385-405.

Sozzi, G., and Boeri, M., (2014). Potential biomarkers for

lung cancer screening. Transl. Lung Cancer Research,

3(3):139-148.

Spitz, M.R., Hong, W.K., Amos, C.I., Wu, X., Schabath,

M.B., Dong, Q., Shete, S. and Etzel, C.J. (2007). A

risk model for prediction of lung cancer. Journal of the

National Cancer Institute, 99(2):715–726.

Tammemagi, M.C., Katki, H.A., Hocking, W,G., et al.

(2013). Selection criteria for lungcancer screening.

New England Journal of Medicine, 368(8):728–736.

Tammemagi, M.C., (2015). Application of Risk Prediction

Models to Lung Cancer Screening: a review. Journal

of Thoracic Imaging, 30(2):88–100.

Wang, H., and Huang, G., (2011). Application of support

vector machine in cancer diagnosis. Medical

Oncology, 28(1):613-618.

Wood, D.E., Eapen, G.A., Ettinger, D.S., et al. (2012).

Lung cancer screening. Journal of the National

Comprehensive Cancer Network, 10(2):240–265.

Zander, T., Hofmann, A., Staratschek-Jox, A., et al.

BIOIMAGING 2016 - 3rd International Conference on Bioimaging

132

(2011). Blood-Based Gene Expression Signatures in

Non–Small Cell Lung Cancer. Clinical Cancer

Research, 17(10):3360-3367.

A Support Vector Machine based Prediction Model for Discrimination of Malignant Pulmonary Nodules from Benign Nodules

133