research article detection of gastric cancer with fourier transform...

Hindawi Publishing CorporationBioMed Research InternationalVolume 2013 Article ID 942427 4 pageshttpdxdoiorg1011552013942427

Research ArticleDetection of Gastric Cancer with Fourier Transform InfraredSpectroscopy and Support Vector Machine Classification

Qingbo Li1 Wei Wang1 Xiaofeng Ling2 and Jin Guang Wu3

1 School of Instrumentation Science and Opto-Electronics Engineering Precision Opto-Mechatronics Technology Key Laboratory ofEducation Ministry Beihang University Xueyuan Road Number 37 Haidian District Beijing 100191 China

2Department of General Surgery Third Hospital Peking University Beijing 100083 China3 College of Chemistry and Molecular Engineering Peking University Beijing 100871 China

Correspondence should be addressed to Qingbo Li qblee126com

Received 30 April 2013 Accepted 19 July 2013

Academic Editor Joanna Domagala-Kulawik

Copyright copy 2013 Qingbo Li et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Early diagnosis and earlymedical treatments are the keys to save the patientsrsquo lives and improve the living quality Fourier transforminfrared (FT-IR) spectroscopy can distinguish malignant from normal tissues at the molecular level In this paper programs weremadewith pattern recognitionmethod to classify unknown samples Spectral datawere pretreated by using smoothing and standardnormal variate (SNV) methods Leave-one-out cross validation was used to evaluate the discrimination result of support vectormachine (SVM) method A total of 54 gastric tissue samples were employed in this study including 24 cases of normal tissuesamples and 30 cases of cancerous tissue samples The discrimination results of SVM method showed the sensitivity with 100specificity with 833 and total discrimination accuracy with 922

1 Introduction

Cancer is a disease that does great harms to the health ofhuman beings The development of tumor is a multistepcomplex process which is affected by many factors By farthere have been few effective ways to cure the cancer Thusthe survival of patients depends largely on the detection ofcancer at an early stage It is of great importance to explorethe early cancer diagnosis method But before the changesin cell morphology can be seen under light microscopemillions of cancer cells have already existed In the processof carcinogenesis nuclear acids proteins carbohydratesand other biomolecules generate significant changes in theirmolecular structures Fourier transform infrared (FT-IR)spectroscopy as an effective tool for investigating chemicalchanges at molecular level has been utilized to detect car-cinoma This method could detect human tissues directlywithout sample pretreatment so the operation is simple timesaving and convenient compared with the gene expression-based method At present with the development of biospec-troscopy and spectral analysis technology the applicationof FT-IR spectroscopy in distinguishing malignant tissues

from normal ones has become a focus [1ndash5] Also greatprogresses have beenmade in the research of cancer detectionusing FT-IR spectroscopy Some institutes such as College ofChemistry and Molecular Engineering in Peking Universityin China have got several achievements to a certain extent[6ndash11]

In this paper the spectra of gastric tissues were collectedby FT-IR spectrometer In order to eliminate the high-frequency random noise baseline drift and light scatteringspectra preprocessing methods of smoothing and standardnormal variate (SNV)were used After spectra preprocessingthe spectra are presented with a high signal-to-noise ratioIn order to achieve a high discrimination accuracy of gastriccancer diagnosis SVM method was adopted to classify thespectra of normal gastric tissues and cancerous gastric tissues

2 Materials and Methods

21 Tissue Specimens A total of 54 gastric tissues wereobtained from the Surgery Department of theThird Hospitalof Peking University China The samples were washed by

2 BioMed Research International

distilled water and divided into two equal parts one partfor pathological study the other part for FT-IR spectroscopicmeasurement According to the results from the pathologydiagnosis the studied samples consisted of 30 cases of cancertissues and 24 cases of normal tissues

22 Acquisition of FT-IR Spectra of Gastric Tissues The FT-IR spectra of samples were obtained on a Nicolet Magna 750II FT-IR spectrometer Spectra-Tech mid-IR optical fiber wasutilized Scan range is from 4000 cmminus1 to 900 cmminus1 with aresolution of 4 cmminus1

23 Spectra Preprocessing Method First smoothing is uti-lized to filter high-frequency noise in FT-IR spectra Thenthe absorption band of CO

2

is substituted with a straight linewhich contains little useful information for measurementLast standard normal variate (SNV) method [12 13] isadopted to weaken the effect to the accuracy of modelingfrom the differences of sample shape size density andnonspecific scatter at the surface of the samples

Each spectrum was preprocessed by SNV the spectro-scopic data of sample 119894 at wavenumber 119896 could be standardnormalized as (1)

119909119894119896SNV =

119909119894119896

minus 119909119894

radicsum119901

119896=1

(119909119894119896

minus 119909119894

)2

(119901 minus 1)12

(1)

where 119909119894

represents the average of spectroscopic data ofsample 119894 while 119901 is the number of wavelength and (119901 minus 1)is the freedom degrees

24 Discrimination Analysis Method Support vectormachine (SVM) is a kernel-based learning method rootedin structural risk minimization [14] and it is based on theconcept of decision planes that define decision boundariesA decision plane is one that separates between a set of objectshaving different class memberships SVM has shown tobe capable of handling high-dimensional data well as itsperformance bounds on classification error do not explicitlydepend on the dimensionality of the input data [15] Theprinciple is explained as follows [16]

For labeled training data of the form (119909119894

119910119894

) 119894 isin 1 119897where 119909

119894

is an 119899-dimensional feature vector and 119910 isinminus1 1 the labels a decision function is found representinga separating hyperplane defined as (2)

119891 (119909) = (⟨119908 120601 (119909119894

)⟩ + 119887) (2)

where119908 is theweight vector 119887 is the bias value and120601(119909) is thekernel function By projecting the data using amapping 120601(119909)nonlinear decision boundaries in the input data space can beobtained The separating hyperplane is found by maximizingdistances to its closest data points embedding it in a largemargin which is defined by support vectors Finding the

hyperplane whilemaximizing themargin is formulated as thefollowing optimization problem

min12119908119879

119908 + 119862

119873

sum

119894=1

120585119894

subject to 119910119894

(⟨119908 120601 (119909119894

)⟩ + 119887) ge 1 minus 120585119894

120585119894

ge 0 (119894 = 1 119873)

(3)

where 119862 is the cost parameter constant 120585119894

is parameter forhandling nonseparable data and the index 119894 labels the 119873training cases Note that 119910 isin minus1 1 is the class labels and119909119894

is the independent variablesThe optimization problem can be efficiently solved using

an equivalent formulation to (3) using Lagrangian multipli-ers In this formulation data are represented exclusively bydot products which can be replaced with a kernel function119870(119909 119909

1015840

) = 120601(119909)119879

120601(1199091015840

) allowing large margin separation inthe kernel spaceThe choice of the kernel function determinesthe mapping 120601(119909) of the input data into a higher dimensionalfeature space in which the linear separating hyperplane isfound (see also (2)) In this paper nonlinear classificationusing the Gaussian kernel as (4) was investigated

119870(119909 119909119894

) = exp(minus1003817100381710038171003817119909 minus 1199091198941003817100381710038171003817

2

120590) (4)

It comprises one free parameter the kernel width 120590 whichcontrols the amount of local influence of support vectors onthe decision boundary

3 Results and Discussion

31 FT-IR Spectra of Gastric Tissues The original FT-IRspectra (Figure 1(a)) were got from it it would be seenthat the spectra obtained from two different experimentsdivided into two clusters Contrast to Figure 1(b) whichshowed the spectra preprocessed with SVN it would befound that low-frequency noise and baseline drift had beencorrected with preprocessing After spectra preprocessingthe differences between FT-IR spectra of normal gastrictissues and cancerous gastric tissues are more pronouncedand the boundary of two clusters is also clearer The relevantinformation of FT-IR spectra has been explained in theprevious paper [7 10]

32 Discrimination Analysis Before discrimination resam-pling half mean (RHM) [17] method is adopted to eliminatethe outliers and the time of resampling process is set 500TheRHM scores of 54 gastric tissue samples are shown in Table 1it could be seen that the samples Number 11 18 and 19 wereoutliers which would be removed

After removing the outliers discrimination model wasbuilt for classification of cancer and normal tissue samplesand leave-one-out cross validation (LOOCV) was utilizedto evaluate the discrimination results of SVM method TheLOOCVmethod attempts to predict the data of the unknownsample with the data of training sample set One sample was

BioMed Research International 3

3400 2920 2440 1960 1480 1000

0

02

04

06Ab

sorb

ance

(au

)

Normal

Cancer

3400 2920 2440 1960 1480 1000

Wavenumber (cmminus1)


minus02

0

02

04

06

Abso

rban

ce (a

u)

minus02

(a)

Normal

Cancer

3400 2920 2440 1960 1480 1000

0

2

4

Abso

rban

ce (a

u)

3400 2920 2440 1960 1480 1000

0

1

2

3

Abso

rban

ce (a

u)



minus2

minus1

(b)

Figure 1 The FTIR spectra of normal and malignant gastric tissues (a) Original spectra (b) preprocessed spectra with smoothing and SNV

Table 1 RHM scores of gastric tissue samples

Number Score Number Score Number Score Number Score1 0 15 0 29 9 43 02 0 16 0 30 0 44 03 0 17 0 31 0 45 04 24 18 471 32 0 46 05 0 19 325 33 0 47 06 15 20 0 34 0 48 07 0 21 0 35 0 49 08 0 22 0 36 0 50 09 0 23 0 37 0 51 010 0 24 42 38 0 52 011 289 25 0 39 0 53 012 0 26 102 40 0 54 013 0 27 0 41 014 1 28 0 42 0

randomly selected and excluded from the training set Theselected sample regarded is as an unknown one and then thesample is classified using themodel built with the rest trainingsamples Record the result Repeat the above course until allsamples had been selected for once and only for once

The discrimination results were shown in Table 2 whichdisplays that among the 24 cases of normal tissue samples 20cases were correctly distinguished while 4 samples were mis-judged among the 27 cases of cancer tissue samples all caseswere well judged The total accuracy was 922

From Table 2 it could be determined that the sensitivityof cancer diagnosis was 100 specificity of cancer diagnosiswas 833 prediction value of a positive test was 871 andprediction value of a negative test was 100

Table 2 Discrimination results using SVMmethod

Pathologic analyzing result SVMmethod resultPositive (T+) Negative (Tminus)

Cancer (D+) 27 0Normal (Dminus) 4 20

4 Conclusions

In this paper the FT-IR spectra of normal gastric tissuesand cancerous gastric tissues are acquired After spectrapreprocessing of smoothing and SNV the spectra withhigh signal-to-noise ratio are presented and achieve a highdiscrimination accuracy with 922 by using SVM methodThis indicates that FT-IR spectroscopy with chemometricsmethod is reliable practical and could be easily implementedin gastric cancer diagnosis

Acknowledgments

This work is supported by Programs for Changjiang Scholarsand Innovative Research Team (PCSIRT) in University ofChina (IRT0705) and National Natural Science Foundation(60708026)

References

[1] J K Pijanka N Stone G Cinque et al ldquoFTIR microspec-troscopy of stained cells and tissues Application in cancerdiagnosisrdquo Spectroscopy vol 24 no 1-2 pp 73ndash78 2010

[2] P D Lewis K E Lewis R Ghosal et al ldquoEvaluation of FTIRSpectroscopy as a diagnostic tool for lung cancer using sputumrdquoBMC Cancer vol 10 article 6404 2010


[3] J Soh A Chueng A Adio A J Cooper B R Birch and B ALwaleed ldquoFourier transform infrared spectroscopy imaging oflive epithelial cancer cells under non-aqueous mediardquo Journalof Clinical Pathology vol 66 no 4 pp 312ndash318 2013

[4] L B Mostaco-Guidolin L S Murakami M R Batistuti ANomizo and L Bachmann ldquoMolecular and chemical charac-terization by Fourier transform infrared spectroscopy of humanbreast cancer cells with estrogen receptor expressed and notexpressedrdquo Spectroscopy vol 24 no 5 pp 501ndash510 2010

[5] T Hu Y H Lu C G Cheng and X Sun ldquoStudy on theearly detection of gastric cancer based on discrete wavelettransformation feature extraction of FT-IR spectra combinedwith probability neural networkrdquo Spectroscopy vol 26 no 3 pp155ndash165 2011

[6] J G Wu Y Z Xu C W Sun et al ldquoDistinguishing malignantfrom normal oral tissues using FTIR fiber-optic techniquesrdquoBiopolymersmdashBiospectroscopy Section vol 62 no 4 pp 185ndash1922001

[7] P R Griffiths H S Yang Q B Li et al ldquoDiscrimination ofnormal and malignant gastric tissues with FTIR spectroscopyand principal component analysisrdquo Spectroscopy and SpectralAnalysis vol 24 no 9 pp 1025ndash1027 2004

[8] Q B Li X J Sun Y Z Xu et al ldquoUse of Fourier-transforminfrared spectroscopy to rapidly diagnose gastric endoscopicbiopsiesrdquo World Journal of Gastroenterology vol 11 no 25 pp3842ndash3845 2005

[9] Q B Li X J Sun Y Z Xu et al ldquoDiagnosis of gastricinflammation and malignancy in endoscopic biopsies based onfourier transform infrared spectroscopyrdquo Clinical Chemistryvol 51 no 2 pp 346ndash350 2005

[10] Q B Li X Li Y Z Xu et al ldquoCancer investigation atPeking University using Fourier transform infrared (FT-IR)spectroscopy FT-IR separation of gastric cancer from normalstomach using chemometicsrdquo Gastroenterology vol 130 ppA473ndashA473 2006

[11] X Li Q B Li G J Zhang et al ldquoIdentification of colitisand cancer in colon biopsies by Fourier transform infraredspectroscopy and chemometricsrdquo The Scientific World Journalvol 2012 Article ID 936149 4 pages 2012

[12] R J Barnes M S Dhanoa and S J Lister ldquoStandard normalvariate transformation and de-trending of near-infrared diffusereflectance spectrardquoApplied Spectroscopy vol 43 no 5 pp 772ndash777 1989

[13] W Wu Q Guo D Jouan-Rimbaud and D L Massart ldquoUsingcontrasts as data pretreatment method in pattern recognitionof multivariate datardquo Chemometrics and Intelligent LaboratorySystems vol 45 no 1-2 pp 39ndash53 1999

[14] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[15] N Cristianini and J Shawe-Taylor An Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[16] S S Keerthi and C J Lin ldquoAsymptotic behaviors of supportvector machines with gaussian kernelrdquo Neural Computationvol 15 no 7 pp 1667ndash1689 2003

[17] W J Egan and S L Morgan ldquoOutlier detection in multivariateanalytical chemical datardquo Analytical Chemistry vol 70 no 11pp 2372ndash2379 1998

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


distilled water and divided into two equal parts one partfor pathological study the other part for FT-IR spectroscopicmeasurement According to the results from the pathologydiagnosis the studied samples consisted of 30 cases of cancertissues and 24 cases of normal tissues

22 Acquisition of FT-IR Spectra of Gastric Tissues The FT-IR spectra of samples were obtained on a Nicolet Magna 750II FT-IR spectrometer Spectra-Tech mid-IR optical fiber wasutilized Scan range is from 4000 cmminus1 to 900 cmminus1 with aresolution of 4 cmminus1

23 Spectra Preprocessing Method First smoothing is uti-lized to filter high-frequency noise in FT-IR spectra Thenthe absorption band of CO

2

is substituted with a straight linewhich contains little useful information for measurementLast standard normal variate (SNV) method [12 13] isadopted to weaken the effect to the accuracy of modelingfrom the differences of sample shape size density andnonspecific scatter at the surface of the samples

Each spectrum was preprocessed by SNV the spectro-scopic data of sample 119894 at wavenumber 119896 could be standardnormalized as (1)

119909119894119896SNV =

119909119894119896

minus 119909119894

radicsum119901

119896=1

(119909119894119896

minus 119909119894

)2

(119901 minus 1)12

(1)

where 119909119894

represents the average of spectroscopic data ofsample 119894 while 119901 is the number of wavelength and (119901 minus 1)is the freedom degrees

24 Discrimination Analysis Method Support vectormachine (SVM) is a kernel-based learning method rootedin structural risk minimization [14] and it is based on theconcept of decision planes that define decision boundariesA decision plane is one that separates between a set of objectshaving different class memberships SVM has shown tobe capable of handling high-dimensional data well as itsperformance bounds on classification error do not explicitlydepend on the dimensionality of the input data [15] Theprinciple is explained as follows [16]

For labeled training data of the form (119909119894

119910119894

) 119894 isin 1 119897where 119909

119894

is an 119899-dimensional feature vector and 119910 isinminus1 1 the labels a decision function is found representinga separating hyperplane defined as (2)

119891 (119909) = (⟨119908 120601 (119909119894

)⟩ + 119887) (2)

where119908 is theweight vector 119887 is the bias value and120601(119909) is thekernel function By projecting the data using amapping 120601(119909)nonlinear decision boundaries in the input data space can beobtained The separating hyperplane is found by maximizingdistances to its closest data points embedding it in a largemargin which is defined by support vectors Finding the

hyperplane whilemaximizing themargin is formulated as thefollowing optimization problem

min12119908119879

119908 + 119862

119873

sum

119894=1

120585119894

subject to 119910119894

(⟨119908 120601 (119909119894

)⟩ + 119887) ge 1 minus 120585119894

120585119894

ge 0 (119894 = 1 119873)

(3)

where 119862 is the cost parameter constant 120585119894

is parameter forhandling nonseparable data and the index 119894 labels the 119873training cases Note that 119910 isin minus1 1 is the class labels and119909119894

is the independent variablesThe optimization problem can be efficiently solved using

an equivalent formulation to (3) using Lagrangian multipli-ers In this formulation data are represented exclusively bydot products which can be replaced with a kernel function119870(119909 119909

1015840

) = 120601(119909)119879

120601(1199091015840

) allowing large margin separation inthe kernel spaceThe choice of the kernel function determinesthe mapping 120601(119909) of the input data into a higher dimensionalfeature space in which the linear separating hyperplane isfound (see also (2)) In this paper nonlinear classificationusing the Gaussian kernel as (4) was investigated

119870(119909 119909119894

) = exp(minus1003817100381710038171003817119909 minus 1199091198941003817100381710038171003817

2

120590) (4)

It comprises one free parameter the kernel width 120590 whichcontrols the amount of local influence of support vectors onthe decision boundary

3 Results and Discussion

31 FT-IR Spectra of Gastric Tissues The original FT-IRspectra (Figure 1(a)) were got from it it would be seenthat the spectra obtained from two different experimentsdivided into two clusters Contrast to Figure 1(b) whichshowed the spectra preprocessed with SVN it would befound that low-frequency noise and baseline drift had beencorrected with preprocessing After spectra preprocessingthe differences between FT-IR spectra of normal gastrictissues and cancerous gastric tissues are more pronouncedand the boundary of two clusters is also clearer The relevantinformation of FT-IR spectra has been explained in theprevious paper [7 10]

32 Discrimination Analysis Before discrimination resam-pling half mean (RHM) [17] method is adopted to eliminatethe outliers and the time of resampling process is set 500TheRHM scores of 54 gastric tissue samples are shown in Table 1it could be seen that the samples Number 11 18 and 19 wereoutliers which would be removed

After removing the outliers discrimination model wasbuilt for classification of cancer and normal tissue samplesand leave-one-out cross validation (LOOCV) was utilizedto evaluate the discrimination results of SVM method TheLOOCVmethod attempts to predict the data of the unknownsample with the data of training sample set One sample was


3400 2920 2440 1960 1480 1000

0

02

04

06Ab

sorb

ance

(au

)

Normal

Cancer

3400 2920 2440 1960 1480 1000



minus02

0

02

04

06

Abso

rban

ce (a

u)

minus02

(a)

Normal

Cancer

3400 2920 2440 1960 1480 1000

0

2

4

Abso

rban

ce (a

u)

3400 2920 2440 1960 1480 1000

0

1

2

3

Abso

rban

ce (a

u)



minus2

minus1

(b)










4 Conclusions


Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















3400 2920 2440 1960 1480 1000

0

02

04

06Ab

sorb

ance

(au

)

Normal

Cancer

3400 2920 2440 1960 1480 1000



minus02

0

02

04

06

Abso

rban

ce (a

u)

minus02

(a)

Normal

Cancer

3400 2920 2440 1960 1480 1000

0

2

4

Abso

rban

ce (a

u)

3400 2920 2440 1960 1480 1000

0

1

2

3

Abso

rban

ce (a

u)



minus2

minus1

(b)










4 Conclusions


Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















research article detection of gastric cancer with fourier transform...

Documents