Research Article Automatic Classification of Normal and ...Research Article Automatic Classification...

Research ArticleAutomatic Classification of Normal and Cancer Lung CT ImagesUsing Multiscale AM-FM Features

Eman Magdy Nourhan Zayed and Mahmoud Fakhr

Computer and Systems Department Electronic Research Institute Giza 12611 Egypt

Correspondence should be addressed to Eman Magdy emanmagdyeriscieg

Received 1 May 2015 Revised 24 August 2015 Accepted 1 September 2015

Academic Editor Tiange Zhuang

Copyright copy 2015 Eman Magdy 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

Computer-aided diagnostic (CAD) systems provide fast and reliable diagnosis for medical images In this paper CAD systemis proposed to analyze and automatically segment the lungs and classify each lung into normal or cancer Using 70 differentpatientsrsquo lung CT datasetWiener filtering on the original CT images is applied firstly as a preprocessing step Secondly we combinehistogram analysis with thresholding and morphological operations to segment the lung regions and extract each lung separatelyAmplitude-Modulation Frequency-Modulation (AM-FM) method thirdly has been used to extract features for ROIs Then thesignificant AM-FM features have been selected using Partial Least Squares Regression (PLSR) for classification step Finally 119870-nearest neighbour (119870NN) support vector machine (SVM) naıve Bayes and linear classifiers have been used with the selectedAM-FM featuresThe performance of each classifier in terms of accuracy sensitivity and specificity is evaluatedThe results indicatethat our proposed CAD system succeeded to differentiate between normal and cancer lungs and achieved 95 accuracy in case ofthe linear classifier

1 Introduction

Computed Tomography (CT) has outperformed conven-tional radiography in the screening of lungs because it gener-ates very detailed high-resolution images and can show early-stage lesions that are too small to be detected by conventionalX-ray CT has been widely used to detect numerous lungdiseases including pneumoconiosis pneumonia pulmonaryedema and lung cancer [1] Early detection of diseases is verycrucial for treatment planning However it is considered oneof the most challenging tasks performed by radiologists dueto the huge amount of data generated by CT scan Thereforecomputer-aided diagnostic (CAD) systems are needed toassist radiologists in the analysis and evaluation of CTscans

A CAD system analyzes medical images in several stepsfirst a preprocessing step for noise reduction and enhancingthe image quality and then segmentation step to differentiateregion of interest (ROI) from other structures in the imageAfter segmentation different features such as geometri-cal textural and statistical features are extracted Finally

a classificationevaluation step is done to evaluate and diag-nose the ROI based on extracted features

Many efforts have been made to provide computer-aided diagnosis for lung images Lung segmentation is anecessary step it has progressed from manual tracing tosemiautomated to fully automated segmentation Here someautomated lung segmentation studies are presented [2ndash9]Other studies present content-based image retrieval (CBIR)systems for lung images [10ndash15] Earlier work in classificationof lung cancer includes the work of Patil and Kuchanur [16]and Kuruvilla and Gunavathi [17] that used artificial neuralnetworks to classify lung cancer images based on the featuresextracted from lung segmented images Nevertheless PatilandKuchanur used geometrical features for classification andachieved only 83 accuracy of classification And Kuruvillaand Gunavathi used statistical parameters as features forclassification and achieved accuracy of 933 Another workby Depeursinge et al [18] classified different lung tissue pat-terns using discrete wavelet frames combined with gray-levelhistogram features However themain limitation of this workwas the lack of resolution in scales with the decomposition

Hindawi Publishing CorporationInternational Journal of Biomedical ImagingVolume 2015 Article ID 230830 7 pageshttpdxdoiorg1011552015230830

2 International Journal of Biomedical Imaging

Image preprocessing

ROI selection

Feature extraction

Feature selection

Classification

Normal Cancer

Figure 1 Block diagram of the proposed fully automated CADsystem

along with required feature weighting while merging featuresfrom different origins

In this paper we propose a CAD system for analysisautomatic segmentation and classification of lung imagesinto normal or cancer from CT dataset The system isbased on the multiscale Amplitude-Modulation Frequency-Modulation (AM-FM) approach The lungs are firstly seg-mented from CT images and next left and right lungs areseparated individually to be analyzed over a filterbankThenthe AM-FM features are extracted and reduced for theclassification step Different classifiers are used to classifythe images and the performance of each classifier has beenevaluated

2 Materials and Methods

Figure 1 shows the main block diagram of our proposed fullyautomated CAD system As seen in this figure the systemis composed of five main steps image preprocessing selectregion of interest (ROI) feature extraction using AM-FMapproach feature selection to find the significant features andfinally a classification The details of each step are discussedin the following sections

21 Dataset Data used in this research were obtained fromThe Cancer Imaging Archive (TCIA) sponsored by theSPIE NCINIH AAPM and the University of Chicago[19] A dataset of 83 CT images from 70 different patientswas included All images have a size of 512 times 512 pixelsand are stored in Digital Imaging and Communication inMedicine (DICOM) format An example of dataset is shownin Figure 2(a) In this figure the right lung is abnormal as ithas a cancer (the rounded gray shape) while the left lung isa normal one For each lung CT image we separate the leftlung from the right lung automatically (as discussed later in

ROI Selection) and each separated lung is labelled as normalor cancer based on the dataset information

22 Image Preprocessing The objective of preprocessing stepis to remove unwanted noise and enhance image quality Wehave used a Wiener filter to remove noise while preservingthe edges and fine details of lungs The filter size of 3 times 3 isselected to avoid oversmoothing of the image The result ofWiener filtering is shown in Figure 2(b)

Wiener filtering [20] is based on estimating the localmean and variance from a local neighborhood of eachpixel Then it creates pixel-wise linear filtering using theseestimates

119865 (119898 119899) = 120583 +1205902minus V2

1205902(119868 (119898 119899) minus 120583) (1)

where 119868 and 119865 denote the original and filtered imagesrespectively 120583 and 120590 denote the mean and variance of a localneighborhood respectively and V is the noise variance

23 ROI Selection Lung segmentation is a necessary step forany lung CAD system We perform automatic segmentationof the lungs using successive steps Then the resultingsegmented image is used to extract each lung separately(ROIs) producing two images one for the left lung and theother for the right lung

In the CT image air appears in a mean intensity ofapproximately minus3024 Hounsfield units (HU) and the lungtissue is in the range of minus910HU to minus500HU while otherstructures are above minus500HUThe goal of segmentation stepis to separate the lungs from both background and nonlungregions To accomplish this we propose a hybrid techniqueresulting from a combination of histogram analysis thresh-olding and morphological operations for automatic lungsegmentation

To simplify the segmentation process the thorax regionis firstly segmented from the background A gray-leveldistribution (histogram) of the Wiener-filtered image is usedto identify different regions in the image The histogram hasone peak corresponding to lung region and another twopeaksfor fat and muscle of thorax region and lung mediastinumIn addition there is a spike at minus3024HU corresponding tobackground pixels Figure 3 shows all peaks except for thebackground spike

The threshold value is then computed from this histogramaccording to the following equation

119879 =119868FM minus 1198681198712+ 119868119871 (2)

where 119868119871denotes the peak intensity value of lung region and

119868FM denotes the average intensity value of fatmuscle peaksThen a binary image (Figure 4(a)) for segmented thorax

region is created where the pixels with gray level greater thanthe selected threshold are set to ldquoonerdquo and other pixels areldquozerordquo

After the thorax region is segmented we perform a fillingoperation to fill the holes inside the binary image so that thepixel values of lungs change from zero to one and produce

International Journal of Biomedical Imaging 3

Cancerlung

Normallung

(a) (b)

Figure 2 Image preprocessing (a) original image and (b) filtered image

0 2000

20

40

60

80

100

120

Pixe

l cou

nt Lung region

minus1000 minus800 minus600 minus400 minus200

Gray level (HU)

Fat + Muscle +mediastinum

mediastinum

Figure 3 Histogram of the Wiener-filtered image

a filled image (Figure 4(b)) Finally the thorax binary imageis subtracted from the filled image to obtain the segmentedlungs as shown in Figure 4(c)

Once the image of lung segmentation is obtained we usedit to locate the left and right lungs in the filtered image as seenin Figure 4(d)Then this image is divided into two images forboth lungs separately each one covering the region of the lungas shown in Figures 4(e) and 4(f) From the 83 CT images weobtained 166 different lung images after each image divisionwhere 83 of them are normal lung images and the other 83images are cancer lung images

24 Feature Extraction In this step we apply the Amplitude-Modulation Frequency-Modulation (AM-FM) modelingtechniques to extract features from lung images that will beused further in classification A lot of research work has beenmade on AM-FMmodels [21ndash23]

241 AM-FM Methods AM-FM is a technique that modelsnonstationary signals Unlike Fourier transforms that provide

the frequency content of signal AM-FM methods providepixel-based information in terms of instantaneous amplitude(IA) instantaneous frequency (IF) and instantaneous phase(IP) And it does not have the main limitation found onwavelet when used to segment the lungs which was the lackof resolution in scales with the decomposition

In 2D AM-FM model a nonstationary image is repre-sented by a sum of AM-FM components as [22]

119868 (1198961 1198962) =

119872

sum

119899=1

119886119899(1198961 1198962) cos120593

119899(1198961 1198962) (3)

where 119899=1 2 119872 denotes the different AM-FM harmon-ics 119886119899(1198961 1198962) denotes the instantaneous amplitude functions

(IA) and 120593119899(1198961 1198962) denotes the instantaneous phase func-

tions (IP)For each AM-FM component the instantaneous fre-

quency (IF) is defined as the gradient of phase nabla120593119899(1198961

1198962) = (120597120593

119899(1198961 1198962)1205971198961 120597120593119899(1198961 1198962)1205971198962) Here the AM-FM

demodulation problem is to estimate the IA IF and IP forthe given input image

In this work AM-FMdemodulation is achieved in severalsteps First we extend the input image to an analytic image byadding an imaginary part equal to the 2D Hilbert transformof the image [21] Given a real-valued image 119868(119896

1 1198962) the

analytic image 119868AS(1198961 1198962) is calculated as follows

119868AS (1198961 1198962) = 119868 (1198961 1198962) + 1198951198672D [119868 (1198961 1198962)] (4)

where 1198672D denotes a two-dimensional extension of one-

dimensional Hilbert transformThen the analytic image is processed through a collection

of band-pass filters (filterbank) (to be discussed in the nextsubsection) in order to isolate the AM-FM components

And from each filter response we can estimate the IA andIP straightforwardly using these following equations

119886 (1198961 1198962) =1003816100381610038161003816119868AS (1198961 1198962)

1003816100381610038161003816

120593 (1198961 1198962) = tanminus1

imag (119868AS (1198961 1198962))real (119868AS (1198961 1198962))

(5)


(a) (b) (c)

(d) (e) (f)

Figure 4 Segmentation and ROI selection (a) Thorax binary image (b) Filled image (c) Lung segmentation (d) Rectangular ROI of bothlungs (e) Right lung (cancer) (f) Left lung (normal)

To estimate the IF we used a variable spacing local linearphase (VS-LLP) method as described in [23]

242 Filterbank Design Thepurpose ofmultiscale filterbankis to isolate the AM-FM image components in model (3)prior to performing demodulation Here we use a four-scalefilterbank developed by Murray [22] (see Figure 5)

In Figure 5 the frequency range of filterbank is depictedFilter 1 is a low-pass filter (LPF) filters 2ndash7 are high frequencyfilters (H) filters 8ndash13 are medium frequency filters (M)filters 14ndash19 are low frequency filters (L) and filters 20ndash25are very low frequency filters (VL) It can be noticed that thebandwidth is decreased by a factor of 12 for each added scale

In this paper we used different combinations of scalesto extract the dominant AM-FM features Here are thecombinations used (1) VL L M and H (2) LPF (3) VL (4)L (5) M (6) LPF VL L M and H (7) LPF VL (8) VL L(9) L M (10)M H and (11) H And for each combinationof scales we estimate the IA IP and |IF| using (5) and theequations in [23]

243 Histogram Processing For each combination of scaleswe produce a histogram for AM-FM estimates IA IP and|IF| And all the computed histograms are normalized sothat the area of each histogram is equal to one Then foreach combination of scales we create a 96-bin feature vector

from the IA IP and |IF| histograms with 32 bins for IA 32bins for IF magnitude and 32 bins for IP (centered at themaximum value) Therefore each image produces 11 featurevectors corresponding to the 11 combinations of scales Weneed to obtain a combined feature vector for each case byselecting the optimal and signification features from all overscales Thus we use Partial Least Squares Regression (PLSR)to achieve that

244 Feature Selection Feature selection is an importantstep that provides the significant features which are usedto differentiate between different classes accurately We usedPartial Least Squares Regression (PLSR) [24] which is alinear regression method that finds the relation between thepredicted variables and observationsThe regression problemis defined as

119910 = 119883120573 + 120576 (6)

where119883 is an 119899 times 119901matrix of the extracted AM-FM features(119899 is the number of images and 119901 is the number of features)and 119910 is an 119899 times 1 vector of response or labels We used label0 for normal case and label 1 for abnormal case 120573 is 119901 times 1vector of the regression parameters and 120576 is 119899times1 vector of theresiduals

We apply PLSR to determine the optimal number offeatures to be used We select the PLSR factors number that


7 436

91213 105 2

24 21 2217 141615181911 8

20 01 1714819181516 11

2 5131210 9

3 64 7

0

2324 25

202122

2325

minus120587

minus120587

120587

120587

minus1205872

minus1205872

1205872

1205872(a)

19

17

18

14

15 16

16 15 18 19

14

17

22

20

21

22

20

21

23

25 24

24 25

231

0

0

minus1205874

minus1205874

1205874

1205874

(b)

Figure 5 Four-scale filterbank [22] (a) Complete frequency range of the filterbank (b) Zoom on the low frequency filters (to be easilyreadable)

0 5 10 15 20 25 3065

70

75

80

85

90

95

100

Number of PLSR factors

Varia

nce e

xpla

ined

iny

()

Figure 6 Plotting the number of PLSR factors with the cumulativevariance percentage in the response variable 119910

produces percentage of variance in response variable morethan 90 In Figure 6 we plot the percentage of variance inthe response versus the number of PLSR factors The plotshows that nearly 90 of the variance in 119910 is given by thefirst eleven factors

Once the optimal number of features is obtained we forma feature vector to represent the selected features

25 Classification The final step of the proposed system isto correctly discriminate between normal and cancer lungimages The input to classification stage is the feature matrixfrom the previous step and the labeled vector (where 0 =normal and 1 = cancer)

Here we have used four different classifiers 119870-nearestneighbor (119870NN) [25] support vector machine (SVM) [26]naıve Bayes [25] and linear classifier [25] The basic idea ofall of these classifiers depends on supervised learning that iseach classifier takes a set of labeled images as a training setto build a model that is used further to assign new images(testing set) into classes Out of 166 lung images 100 imagesare selected as a training dataset and 66 other images areselected as a testing dataset

3 Results and Discussion

The four classifiers are trained and their performance isevaluated with leave-119872-out cross validation We change thevalue of119872 to generate different sizes of testing and trainingsets and for each119872 value the classification performance isevaluated by computing these different measures

Sensitivity () = TPTP + FN

times 100

Specificity () = TNFP + TN

times 100

Accuracy () = TP + TNTP + FN + TN + FP

times 100

(7)

where TP TN FN and FP denote true positive true negativefalse negative and false positive respectively [27]

Figure 7 shows the computed accuracy sensitivity andspecificity respectively for the four classifiers with the changein size of testing set It can be noticed that the classifiersperformances in terms of accuracy sensitivity and specificityare much better in case of small size of the testing set (whenthe classifiers got trained with large size of the trainingset) However the performances of all classifiers decrease


0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM

LinearNaiumlve Bayes

Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear

0 20 40 60 80 100 120Size of testing set

04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)

Figure 7 Comparison between accuracy sensitivity and specificityfor the four classifiers with changing size of testing set

Table 1Classification performancemeasures for the four classifiers

Classifier Accuracy Sensitivity SpecificityKNN 64 55 72SVM 90 85 97Naıve Bayes 82 82 82Linear 95 94 97

with increasing testing set size From this figure it is easilyobserved that the performance of SVM classifier is the leaststable with increasing testing set size while the other threeclassifiers are more stable Moreover it can be concluded thatthe linear classifier is the best one to discriminate betweennormal and cancer lungs

Table 1 summarizes the performance measures for thefour classifiers when the size of training set relative to thetesting set is 60 to 40 of the total dataset size respectively(ie training set = 100 images and test set = 66 images)As shown in Table 1 the linear classifier gives the bestclassification with 95 accuracy 94 sensitivity and 97specificity On the other hand the119870NN classifier is the worstclassifier achieving only 64 accuracy 55 sensitivity and72 specificity

It is worth noting that the proposed CAD system isdeveloped using MATLAB R2010a on Intel Core i5 25 GHzCPU 6GB RAM Windows 7 64-bit PC

4 Conclusions

In this paper we develop a lung CAD system that analyzesand automatically segments the lungs and classifies eachlung to normal or cancer The system consists of five mainsteps preprocessing ROI selection feature extraction fea-ture selection and classification AM-FM approach has beenused to extract new features in terms of IA IP and IF AndPLSR is then used to reduce the large number of featuresand select the optimal and significant ones Four classifiersare used and the performance of each classifier has beenevaluated It has been found that the linear classifier was thebest one to discriminate between normal and cancer lungswith 95 accuracy

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

The authors acknowledge the SPIE the NCI the AAPM andThe University of Chicago for making public access to thelung cancer dataset

References

[1] N Hollings and P Shaw ldquoDiagnostic imaging of lung cancerrdquoEuropean Respiratory Journal vol 19 no 4 pp 722ndash742 2002


[2] W Li S DNie and J J Cheng ldquoA fast automaticmethod of lungsegmentation in CT images using mathematical morphologyrdquoinWorld Congress on Medical Physics and Biomedical Engineer-ing 2006 vol 14 of IFMBE Proceedings pp 2419ndash2422 SpringerBerlin Germany 2007

[3] S Hu E A Hoffman and J M Reinhardt ldquoAutomatic lungsegmentation for accurate quantitation of volumetric X-ray CTimagesrdquo IEEE Transactions on Medical Imaging vol 20 no 6pp 490ndash498 2001

[4] J K Leader B Zheng R M Rogers et al ldquoAutomated lung seg-mentation in X-ray computed tomography development andevaluation of a heuristic threshold-based schemerdquo AcademicRadiology vol 10 no 11 pp 1224ndash1236 2003

[5] M N Prasad M S Brown S Ahmad et al ldquoAutomaticsegmentation of lung parenchyma in the presence of diseasesbased on curvature of ribsrdquo Academic Radiology vol 15 no 9pp 1173ndash1180 2008

[6] JWang F Li andQ Li ldquoAutomated segmentation of lungs withsevere interstitial lung disease in CTrdquo Medical Physics vol 36no 10 pp 4592ndash4599 2009

[7] S Armato and H MacMahon ldquoAutomated lung segmentationand computer-aided diagnosis for thoracic CT scansrdquo Interna-tional Congress Series vol 1256 pp 977ndash982 2003

[8] S G Armato III and W F Sensakovic ldquoAutomated lungsegmentation for thoracic CT impact on computer-aided diag-nosisrdquo Academic Radiology vol 11 no 9 pp 1011ndash1021 2004

[9] Y Guo Y Feng J Sun et al ldquoAutomatic lung tumor segmen-tation on PETCT images using fuzzy markov random fieldmodelrdquo Computational andMathematical Methods in Medicinevol 2014 Article ID 401201 6 pages 2014

[10] M Lam T Disney M Pham D Raicu J Furst and RSusomboon ldquoContent-based image retrieval for pulmonarycomputed tomography nodule imagesrdquo in Medical Imaging2007 PACS and Imaging Informatics vol 6516 of Proceedingsof SPIE International Society for Optics and Photonics March2007

[11] A K Dhara C K Chama S Mukhopadhyay and N Khan-delwal ldquoContent-based image retrieval system for differentialdiagnosis of lung cancerrdquo Indian Journal of Medical Informaticsvol 6 no 1 article 1 2012

[12] J B Nirmala and S Gowri ldquoA content based CT lung imageretrieval by DCTmatrix and feature vector techniquerdquo Interna-tional Journal of Computer Science Issues vol 9 no 2 2012

[13] H Muller N Michoux D Bandon and A Geissbuhler ldquoAreview of content-based image retrieval systems in medicalapplicationsmdashclinical benefits and future directionsrdquo Interna-tional Journal of Medical Informatics vol 73 no 1 pp 1ndash232004

[14] C-T Liu P-L Tai A Y-J Chen C-H Peng T Lee and J-S Wang ldquoA content-based CT lung image retrieval system forassisting differential diagnosis images collectionrdquo inProceedingsof the IEEE International Conference on Multimedia and Expo(ICME 2001) pp 174ndash177 August 2001

[15] Y Song W Cai S Eberl M J Fulham and D Feng ldquoAcontent-based image retrieval framework for multi-modalitylung imagesrdquo in Proceedings of the 23rd IEEE InternationalSymposiumonComputer-BasedMedical Systems (CBMS rsquo10) pp285ndash290 October 2010

[16] S A Patil and M B Kuchanur ldquoLung cancer classificationusing image processingrdquo International Journal of Engineeringand Innovative Technology vol 2 no 3 2012

[17] J Kuruvilla and K Gunavathi ldquoLung cancer classificationusing neural networks for CT imagesrdquo Computer Methods andPrograms in Biomedicine vol 113 no 1 pp 202ndash209 2014

[18] A Depeursinge D Sage A Hidki et al ldquoLung tissue classifi-cation using wavelet framesrdquo in Proceedings of the 29th AnnualInternational Conference of IEEE Engineering in Medicine andBiology Society (EMBC rsquo07) pp 6259ndash6262 IEEE Lyon FranceAugust 2007

[19] SPIE-AAPM-NCI Lung Nodule Classification Challenge DatasetThe Cancer Imaging Archive 2015

[20] H Shao L Cao and Y Liu ldquoA detection approach for solitarypulmonary nodules based on CT imagesrdquo in Proceedings of the2nd International Conference on Computer Science and NetworkTechnology (ICCSNT rsquo12) pp 1253ndash1257 Changchun ChinaDecember 2012

[21] J Havlicek AM-FM image models [PhD thesis] University ofTexas at Austin Austin Tex USA 1996

[22] V Murray AM-FM methods for image and video processing[PhD thesis] University of New Mexico Albuquerque NMUSA 2008

[23] V Murray P Rodriguez and M S Pattichis ldquoMultiscale AM-FM demodulation and image reconstruction methods withimproved accuracyrdquo IEEE Transactions on Image Processing vol19 no 5 pp 1138ndash1152 2010

[24] R Rosipal and N Kramer ldquoOverview and recent advances inpartial least squaresrdquo in Subspace Latent Structure and FeatureSelection vol 3940 pp 34ndash51 Springer 2006

[25] T MitchellMachine Learning 1997[26] C Campbell and Y Ying Learning with Support Vector

Machines 2011[27] N A Papadopoulos E M Plissiti and I D Fotiadis ldquoMedical-

image processing and analysis for CAD systemsrdquo in MedicalImage Analysis Methods L Costaridou Ed pp 51ndash86 Tayloramp Francis CRC Press Boca Raton Fla USA 2005

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



Image preprocessing

ROI selection

Feature extraction

Feature selection

Classification

Normal Cancer

Figure 1 Block diagram of the proposed fully automated CADsystem

along with required feature weighting while merging featuresfrom different origins

In this paper we propose a CAD system for analysisautomatic segmentation and classification of lung imagesinto normal or cancer from CT dataset The system isbased on the multiscale Amplitude-Modulation Frequency-Modulation (AM-FM) approach The lungs are firstly seg-mented from CT images and next left and right lungs areseparated individually to be analyzed over a filterbankThenthe AM-FM features are extracted and reduced for theclassification step Different classifiers are used to classifythe images and the performance of each classifier has beenevaluated

2 Materials and Methods

Figure 1 shows the main block diagram of our proposed fullyautomated CAD system As seen in this figure the systemis composed of five main steps image preprocessing selectregion of interest (ROI) feature extraction using AM-FMapproach feature selection to find the significant features andfinally a classification The details of each step are discussedin the following sections

21 Dataset Data used in this research were obtained fromThe Cancer Imaging Archive (TCIA) sponsored by theSPIE NCINIH AAPM and the University of Chicago[19] A dataset of 83 CT images from 70 different patientswas included All images have a size of 512 times 512 pixelsand are stored in Digital Imaging and Communication inMedicine (DICOM) format An example of dataset is shownin Figure 2(a) In this figure the right lung is abnormal as ithas a cancer (the rounded gray shape) while the left lung isa normal one For each lung CT image we separate the leftlung from the right lung automatically (as discussed later in

ROI Selection) and each separated lung is labelled as normalor cancer based on the dataset information

22 Image Preprocessing The objective of preprocessing stepis to remove unwanted noise and enhance image quality Wehave used a Wiener filter to remove noise while preservingthe edges and fine details of lungs The filter size of 3 times 3 isselected to avoid oversmoothing of the image The result ofWiener filtering is shown in Figure 2(b)

Wiener filtering [20] is based on estimating the localmean and variance from a local neighborhood of eachpixel Then it creates pixel-wise linear filtering using theseestimates

119865 (119898 119899) = 120583 +1205902minus V2

1205902(119868 (119898 119899) minus 120583) (1)

where 119868 and 119865 denote the original and filtered imagesrespectively 120583 and 120590 denote the mean and variance of a localneighborhood respectively and V is the noise variance

23 ROI Selection Lung segmentation is a necessary step forany lung CAD system We perform automatic segmentationof the lungs using successive steps Then the resultingsegmented image is used to extract each lung separately(ROIs) producing two images one for the left lung and theother for the right lung

In the CT image air appears in a mean intensity ofapproximately minus3024 Hounsfield units (HU) and the lungtissue is in the range of minus910HU to minus500HU while otherstructures are above minus500HUThe goal of segmentation stepis to separate the lungs from both background and nonlungregions To accomplish this we propose a hybrid techniqueresulting from a combination of histogram analysis thresh-olding and morphological operations for automatic lungsegmentation

To simplify the segmentation process the thorax regionis firstly segmented from the background A gray-leveldistribution (histogram) of the Wiener-filtered image is usedto identify different regions in the image The histogram hasone peak corresponding to lung region and another twopeaksfor fat and muscle of thorax region and lung mediastinumIn addition there is a spike at minus3024HU corresponding tobackground pixels Figure 3 shows all peaks except for thebackground spike

The threshold value is then computed from this histogramaccording to the following equation

119879 =119868FM minus 1198681198712+ 119868119871 (2)

where 119868119871denotes the peak intensity value of lung region and

119868FM denotes the average intensity value of fatmuscle peaksThen a binary image (Figure 4(a)) for segmented thorax

region is created where the pixels with gray level greater thanthe selected threshold are set to ldquoonerdquo and other pixels areldquozerordquo

After the thorax region is segmented we perform a fillingoperation to fill the holes inside the binary image so that thepixel values of lungs change from zero to one and produce


Cancerlung

Normallung

(a) (b)


0 2000

20

40

60

80

100

120

Pixe

l cou

nt Lung region


Gray level (HU)


mediastinum








119868 (1198961 1198962) =

119872

sum

119899=1

119886119899(1198961 1198962) cos120593

119899(1198961 1198962) (3)





1198962) = (120597120593

119899(1198961 1198962)1205971198961 120597120593119899(1198961 1198962)1205971198962) Here the AM-FM



1 1198962) the


119868AS (1198961 1198962) = 119868 (1198961 1198962) + 1198951198672D [119868 (1198961 1198962)] (4)





119886 (1198961 1198962) =1003816100381610038161003816119868AS (1198961 1198962)

1003816100381610038161003816

120593 (1198961 1198962) = tanminus1

imag (119868AS (1198961 1198962))real (119868AS (1198961 1198962))

(5)


(a) (b) (c)

(d) (e) (f)









119910 = 119883120573 + 120576 (6)




7 436

91213 105 2

24 21 2217 141615181911 8

20 01 1714819181516 11

2 5131210 9

3 64 7

0

2324 25

202122

2325

minus120587

minus120587

120587

120587

minus1205872

minus1205872

1205872

1205872(a)

19

17

18

14

15 16

16 15 18 19

14

17

22

20

21

22

20

21

23

25 24

24 25

231

0

0

minus1205874

minus1205874

1205874

1205874

(b)


0 5 10 15 20 25 3065

70

75

80

85

90

95

100


Varia

nce e

xpla

ined

iny

()









times 100


times 100


times 100

(7)




0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM


Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear


04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)







4 Conclusions




Acknowledgments


References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Cancerlung

Normallung

(a) (b)


0 2000

20

40

60

80

100

120

Pixe

l cou

nt Lung region


Gray level (HU)


mediastinum








119868 (1198961 1198962) =

119872

sum

119899=1

119886119899(1198961 1198962) cos120593

119899(1198961 1198962) (3)





1198962) = (120597120593

119899(1198961 1198962)1205971198961 120597120593119899(1198961 1198962)1205971198962) Here the AM-FM



1 1198962) the


119868AS (1198961 1198962) = 119868 (1198961 1198962) + 1198951198672D [119868 (1198961 1198962)] (4)





119886 (1198961 1198962) =1003816100381610038161003816119868AS (1198961 1198962)

1003816100381610038161003816

120593 (1198961 1198962) = tanminus1

imag (119868AS (1198961 1198962))real (119868AS (1198961 1198962))

(5)


(a) (b) (c)

(d) (e) (f)









119910 = 119883120573 + 120576 (6)




7 436

91213 105 2

24 21 2217 141615181911 8

20 01 1714819181516 11

2 5131210 9

3 64 7

0

2324 25

202122

2325

minus120587

minus120587

120587

120587

minus1205872

minus1205872

1205872

1205872(a)

19

17

18

14

15 16

16 15 18 19

14

17

22

20

21

22

20

21

23

25 24

24 25

231

0

0

minus1205874

minus1205874

1205874

1205874

(b)


0 5 10 15 20 25 3065

70

75

80

85

90

95

100


Varia

nce e

xpla

ined

iny

()









times 100


times 100


times 100

(7)




0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM


Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear


04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)







4 Conclusions




Acknowledgments


References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)

(d) (e) (f)









119910 = 119883120573 + 120576 (6)




7 436

91213 105 2

24 21 2217 141615181911 8

20 01 1714819181516 11

2 5131210 9

3 64 7

0

2324 25

202122

2325

minus120587

minus120587

120587

120587

minus1205872

minus1205872

1205872

1205872(a)

19

17

18

14

15 16

16 15 18 19

14

17

22

20

21

22

20

21

23

25 24

24 25

231

0

0

minus1205874

minus1205874

1205874

1205874

(b)


0 5 10 15 20 25 3065

70

75

80

85

90

95

100


Varia

nce e

xpla

ined

iny

()









times 100


times 100


times 100

(7)




0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM


Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear


04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)







4 Conclusions




Acknowledgments


References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










7 436

91213 105 2

24 21 2217 141615181911 8

20 01 1714819181516 11

2 5131210 9

3 64 7

0

2324 25

202122

2325

minus120587

minus120587

120587

120587

minus1205872

minus1205872

1205872

1205872(a)

19

17

18

14

15 16

16 15 18 19

14

17

22

20

21

22

20

21

23

25 24

24 25

231

0

0

minus1205874

minus1205874

1205874

1205874

(b)


0 5 10 15 20 25 3065

70

75

80

85

90

95

100


Varia

nce e

xpla

ined

iny

()









times 100


times 100


times 100

(7)




0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM


Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear


04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)







4 Conclusions




Acknowledgments


References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0 20 40 60 80 100 12005

06

07

08

09

055

065

075

085

095

1

Size of testing set

Accu

racy

()

KNNSVM


Naiumlve Bayes

Naiumlve Bayes

KNNSVM

Linear

KNNSVM

Linear


04

05

06

07

09

08

1

Sens

itivi

ty (

)

04

05

06

07

09

08

1


Spec

ifici

ty (

)







4 Conclusions




Acknowledgments


References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation






































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Research Article Automatic Classification of Normal and ...Research Article Automatic Classification...

Documents

Transcript of Research Article Automatic Classification of Normal and ...Research Article Automatic Classification...