Computer-aided screening system forcervical precancerous cells based onfield emission scanning electronmicroscopy and energy dispersive x-rayimages and spectra

Yessi JusmanSiew-Cheok NgKhairunnisa HasikinRahmadi KurniaNoor Azuan Bin Abu OsmanKean Hooi Teoh

Yessi Jusman, Siew-Cheok Ng, Khairunnisa Hasikin, Rahmadi Kurnia, Noor Azuan Bin Abu Osman, KeanHooi Teoh, “Computer-aided screening system for cervical precancerous cells based on field emission scanningelectron microscopy and energy dispersive x-ray images and spectra,” Opt. Eng. 55(10), 103110 (2016),doi: 10.1117/1.OE.55.10.103110.

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

Computer-aided screening system for cervical precancerouscells based on field emission scanning electron microscopyand energy dispersive x-ray images and spectra

Yessi Jusman,a,b,* Siew-Cheok Ng,a Khairunnisa Hasikin,a Rahmadi Kurnia,c Noor Azuan Bin Abu Osman,a andKean Hooi Teohd

aUniversity of Malaya, Department of Biomedical Engineering, Faculty of Engineering, Jalan Universiti, Wilayah Persekutuan,Kuala Lumpur 50603, MalaysiabUniversitas Abdurrab, Department of Informatics Engineering, Faculty of Engineering, Jalan Riau Ujung No. 73, Pekanbaru, Riau 28291,IndonesiacAndalas University, Department of Electrical Engineering, Faculty of Engineering, Jl. Universitas Andalas, Limau Manis, Pauh, Kota Padang,Sumatera Barat 25163, IndonesiadUniversity of Malaya, Department of Pathology, Faculty of Medicine, Jalan Universiti, Wilayah Persekutuan, Kuala Lumpur 50603, Malaysia

Abstract. The capability of field emission scanning electron microscopy and energy dispersive x-ray spectros-copy (FE-SEM/EDX) to scan material structures at the microlevel and characterize the material with its elementalproperties has inspired this research, which has developed an FE-SEM/EDX-based cervical cancer screeningsystem. The developed computer-aided screening system consisted of two parts, which were the automaticfeatures of extraction and classification. For the automatic features extraction algorithm, the image and spectraof cervical cells features extraction algorithm for extracting the discriminant features of FE-SEM/EDX data wasintroduced. The system automatically extracted two types of features based on FE-SEM/EDX images and FE-SEM/EDX spectra. Textural features were extracted from the FE-SEM/EDX image using a gray level co-occur-rence matrix technique, while the FE-SEM/EDX spectra features were calculated based on peak heights andcorrected area under the peaks using an algorithm. A discriminant analysis technique was employed to predictthe cervical precancerous stage into three classes: normal, low-grade intraepithelial squamous lesion (LSIL),and high-grade intraepithelial squamous lesion (HSIL). The capability of the developed screening system wastested using 700 FE-SEM/EDX spectra (300 normal, 200 LSIL, and 200 HSIL cases). The accuracy, sensitivity,and specificity performances were 98.2%, 99.0%, and 98.0%, respectively. © 2016 Society of Photo-Optical InstrumentationEngineers (SPIE) [DOI: 10.1117/1.OE.55.10.103110]

Keywords: cells; imaging technique; signal processing; feature extraction; computer vision; intelligent screening.

Paper 160044 received Jan. 10, 2016; accepted for publication Sep. 21, 2016; published online Oct. 25, 2016.

1 IntroductionCervical cancer is the third most commonly diagnosed cancerand the fourth leading cause of cancer death in females world-wide, which contributed 12% (528,000) of the total newcancer cases and 7.5% (266,000) of the total cancer deathsamong females in 2012.1 Cervical cancer develops over aperiod of two to three decades, providing sufficient time fora precursor screening.2 Therefore, early detection and routinescreening can reduce the incidence and mortality rates relatedto this disease. A multidisciplinary guideline panel and sys-tematic reviews of randomized controlled trials and observa-tional studies were conducted to minimize the incident andmortality rates.3 Several recommendations have been intro-duced for screen-and-treat strategies to prevent cervicalcancer, including human papillomavirus (HPV) testing, cytol-ogy testing, and visual inspection with acetic acid.3

Computer-aided screening techniques have been devel-oped for this purpose to help cytotechnologists and pathol-ogists in cervical cancer diagnosis.4,5 Due to the recentadvancement of imaging technology, much progress has

been made in computer-aided screening systems based onThinPrep,6 colposcopy,7 cervicography,8 fluorescent in situhybridization (FISH),9 and Fourier transform infrared,10

which can reduce human error and processing time.Computer-aided cervical cancer screening systems based oncellular level data have achieved better results compared tocervicography and colposcopy, which are based on extractedtissues.11 Table 1 shows the comparisons among data acquis-ition techniques of cervical cancer computer-aided screeningsystems. The ThinPrep and FISH techniques only providequalitative images to show the nucleus and cytoplasm ofthe cells. Both techniques require expert input to diagnosethe abnormality of the cells based on the morphologicalcharacteristics such as the ratio of cytoplasm to the nuclearsize and the concentrations of cell nuclei in the images.Although the characteristics can be computerized and devel-oped into an automated system, there are limitations, such asthe need for complicated algorithms due to the poor imagesand complications in classification.6,12 The developed comput-erized systems have produced significant results; however,accuracies of the systems have yet to be improved.12,13

*Address all correspondence to: Yessi Jusman, E-mail: yessi.jusman@univrab.ac.id 0091-3286/2016/$25.00 © 2016 SPIE

Optical Engineering 103110-1 October 2016 • Vol. 55(10)

Optical Engineering 55(10), 103110 (October 2016)

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

Meanwhile, the spectroscopy techniques provide more objec-tive results and are independently operated without expertinput. The techniques measure the chemical compounds inthe sample specimen. Although they have achieved goodaccuracy with simple algorithms,14,15 their classification proc-esses are based on the evaluation of the tissues rather thanindividual cervical cells. Evaluation on each cervical cancercell is essential for cancer screening and important to deter-mine the state of the cancer.

FE-SEM/EDX is an electron microscopy technique thatproduces high-resolution images; it captures and scans thestructural surface of a material at the micro- and/or nanoscalelevel for organic (such as polymers, enzymes, cells, and mem-branes) or inorganic (such as ceramics, pigments, minerals,and composite materials) materials.16,17 It can be used as ananalytical technique to investigate molecular surface structuresfor material characterization, understanding its mechanismand mode of formation, as well as explaining/predicting itsproperties. Recently, some researchers have applied the FE-SEM/EDX to fungal cells,18 human carotid plaques,19 yeastcells,20 stem cells,21 and human brain tumors.22

FE-SEM/EDX produces both qualitative and quantitativedata concurrently. Higher resolution images and elementaldistributions of the prospective cells in the images areused for classification purposes. These representative dataare believed to provide more accurate results with minimumcomputational expenses and rapid processing time, asspecific parts in the images can provide objective results(elemental distribution) without an expert consideration.Therefore, cervical precancerous cell features can beextracted using FE-SEM/EDX to observe the morphologyand elemental composition features.23

The morphological features from images are presented byintensities, topology structures, colors, and texture as pre-sented in Ref. 24. The texture of the image is one of thesignificant features to be analyzed in image processing forvarious applications including automated inspection.Currently, a gray level co-occurrence matrix (GLCM) tech-nique has been applied to characterize the most malignanttumors, script identification, facial expression recognition,mammography, brain morphology, and so on.25–29 Findingsindicated that feature extraction based on the GLCM tech-niques have achieved better classification results than othermethods.30,31 Therefore, based on the capability of the

FE-SEM/EDX as a data acquisition and image processingtechnique, this study proposed an automatic system forcervical precancerous screening based on FE-SEM/EDXand GLCM.

2 Materials and MethodsThe medical ethics committee of the University of MalayaMedical Centre has approved the collection of the liquid-based cytology specimens (i.e., ThinPrep vial) from thepathology department. These specimens underwent cervicalscreening for cervical precancerous cells and were diagnosedby the pathologist. In this research, the collected sampleswere cervical cells in the vials from which the rest of thesamples were used for screening by the experts. For the nextprocess, these samples were sent to various laboratories atthe University of Malaya, Kuala Lumpur, for sample prepa-ration and capture of images using FE-SEM/EDX.

2.1 Sample Preparation

In the first step of sample preparation, the specimens werefixed in a McDowell-Trump fixative prepared in 0.1-M phos-phate buffer or cacodylate buffer (pH 7.2) at 4°C for 2 h andthen washed twice with 0.1% in phosphate-buffered salinefor 10 min each time. For the dehydration process, a seriesof ethanol dilution dehydrations was implemented at 50%,75%, and twice at 95% for 15 min each time, and threetimes at 100% with an equilibration step of 20 min eachtime. In the drying process, the dehydrated specimens wereimmersed in 1 to 2 mL of HMDS for 10 min, and the HMDSwas then decanted into the specimen vials, put into thedesiccator, and left to air dry at room temperature. The nextprocess was to mount the dried specimens on circular stain-less steel molds, coated with 10 nm of pure gold in a vacuumsputter coater and kept in a desiccator or under vacuum at alltimes before they were viewed under FE-SEM/EDX.

2.2 Field Emission Scanning Electron Microscopyand Energy Dispersive X-Ray SpectroscopySystem

This research used FE-SEM with field emission gun 250SEM. The capturing of the FE-SEM image was implementedat 10 mm working distance and operated at low voltage(10 kV). Elevation, tilt, and azimuth degrees in the data

Table 1 Characteristics of data acquisition techniques of computer-aided screening system for cervical precancerous cells.

Characteristics Thin prep FISH Optical spectroscopy FE-SEM/EDX

Property Light microscopy Flourecencesmicroscopy

Spectroscopy Electron microscopyand spectroscopy

Resolution Medium High — Very high

Data Qualitative image Qualitative image Quantitative spectrum Qualitative image andquantitative spectrum

View cell ability Cell (nucleusand cytoplasm)

Cell (nucleusand cytoplasm)

— Cell (surface)

Spectral absorption types — — Sample chemicalcompounds

Cell elementaldistributions

More fields Limited Limited Not limited Not limited

Optical Engineering 103110-2 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

acquisition were set up at 35 deg, 0.01 deg to 0.5 deg, and0 deg, respectively. Both inlens and Everhart–Thornleydetectors were used to create the images. The FE-SEM/EDX imaging system for data acquisition was presented indetail in our previous paper.17 A total of 82 ThinPrep vialswere used for the samples in this study. Based on the sam-ples, 150 FE-SEM images were captured from the samplesand the elemental distributions of certain areas in the imageswere measured by using the FE-SEM/EDX technique. Therewere a total of up to 20 cervical cells in each capturedFE-SEM image. Thus, the elemental distributions were mea-sured in up to 20 scanned areas.

3 Proposed Automatic Screening SystemAn automated screening system for cervical cancer was pro-posed. The proposed system was developed to improve ourprevious work.32 A flow chart of the proposed system isshown in Fig. 1. The data input for the system had twotypes of data: FE-SEM/EDX images and FE-SEM/EDXspectra. The system differentiated three classes of the cervi-cal cells to be normal, low-grade intraepithelial squamouslesion (LSIL), and high-grade intraepithelial squamouslesion (HSIL). In this section, images and spectra processingalgorithms for feature extraction were applied on the inputFE-SEM/EDX data. Details of the automated feature extrac-tion technique proposed are presented below.

3.1 Features Extraction Based Field EmissionScanning Electron Microscopy Images

For the image processing technique, a magnification of 5000for the FE-SEM/EDX images of the cervical cells was usedas the region of interest (ROI) for feature extraction pur-poses. A large ROI that did not interfere with the imagescale was needed to observe the texture and pattern at spe-cific intervals and directions. Cervical cancer is caused byHPV infection. In the used ROI for the feature extraction,

the existence of the virus in the abnormal cells (LSIL andHSIL) images can be viewed as referred by the arrow headin Figs. 2(b) and 2(c). Figure 2 shows the textural differencesamong three cervical cells images of the normal, LSIL, andHSIL classes. The textural differences were presented dueto the existence of the virus in abnormal (LSIL and HSIL)cervical cells.

However, the captured images mostly had nonuniform illu-mination. Basically, in order to obtain uniform illumination,enhancement was needed to improve the visibility of thefeatures for better extraction and classification. Therefore,based on the image condition, an image processing algorithmwas needed to improve and clarify the textural features,which will be extracted from the enhanced images.

In this study, two preprocessing steps were applied to theoriginal FE-SEM/EDX images. In the first step, contraststretching was applied to the original images for growthintensity distribution to enhance its quality. The detail oncontrast stretching is widely available in image processingtextbooks.33 Next, a morphological opening operation wasapplied to the enhanced contrast image as explainedpreviously.33 The goal of the morphological opening opera-tion in this preprocessing step was to increase the detectabil-ity of the cell features and decrease undesired features withinthe image.

Then, the textural features were extracted based on theresulting morphological opening images. In this study, weused a statistical approach, which extracted the second-order statistical information based on the GLCM to computethe properties of the texture. As presented in Ref. 34, GLCMhas four parameters for consideration: region size, quantiza-tion levels, displacement, and orientation value. The currenttechnique for feature calculation is based on the rotationinvariant,35 which is calculated for each angle and thenaveraged across all angles.

The properties of GLCM parameter used in this studywere a 924 × 540 region size, nine quantization levels,

Contrast stretching process

Morphological opening

operation

Gray level co-occurrence

matrix

Find peaks and base points

Peak height calculation

Corrected area calculation

Six features from ratio of

the peak height

Six features from ratio of the

corrected area

Preprocessing

Features extraction

FE-SEM/EDX data

FE-SEM/EDX spectra

FE-SEM/EDX images

Classification

Fig. 1 A computer-aided screening system for cervical precancerous based FE-SEM/EDX data.

Optical Engineering 103110-3 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

d ¼ 1 and 2 pixels of displacement and θ ¼ 0 deg, 45 deg,90 deg, 135 deg orientation. The region size was an optimalview of the textural contexts in the image to be differentiatedamong the cervical cell classes as presented in Fig. 2. Severalquantization levels (i.e., 5, 9, 15, and 32 quantization levels)were tested to obtain the optimum quantization levels forthe images, since the gray-level condition in our images(after preprocessing) were minimally scattered as presentedin the results section. Furthermore, a large displacementvalue will result in the inability of the co-occurrence matrixto capture the textural information in the image. Thus, dis-placements of 1 and 2 (d ¼ 1 and 2) pixels were selected toensure that the textural and pattern information can beextracted for classification purposes.

For the FE-SEM images of cervical cells, there was nosystematic pattern based on orientation. The cell imagesand the existence of the virus in the abnormal cervicalcell images rotated and positioned themselves in all possibleorientations. Therefore, we used θ ¼ 0 deg, 45 deg, 90 deg,and 135 deg angles to efficiently cover all directions, asshown in Sec. 3.2.

The joint probability of two pixels with specific intensitylevels displaced a distance d and calculated along a specifieddirection, h. The spatial dependence matrix was defined as

EQ-TARGET;temp:intralink-;e001;63;304Pd;θði; jÞ

¼XN−dx

x¼1

XM−dy

y¼1

�1; if Iðx; yÞ ¼ i and Iðxþ dx; yþ dyÞ ¼ j0; otherwise

;

(1)

where I is the input image, N andM are the height and widthof I, respectively, i and j are the two specific intensity levels,and d denotes the offset between the two pixels ðx; yÞ and(xþ dx, yþ dy).

Four directions of θ, 0 deg, 45 deg, 90 deg, and 135 deg,were considered in the distance, d, which varied from 1 to2 pixels. The image texture analysis was executed in a slice-by-slice basis. For each direction resulting from a certainneighborhood type and a given distance, d, we computedfour features among those most frequently used in the liter-ature: contrast, correlation, energy, and homogeneity. Thefour features were calculated using the following equations:based on the co-occurrence matrixes as presented in Fig. 4,the features were calculated as follows:

EQ-TARGET;temp:intralink-;e002;326;568 1. Contrast; f1 ¼XNg−1

n¼0

n2�XNg

i¼1

XNg

j¼1

pði; jÞ����ji − jj ¼ n

�;

(2)

EQ-TARGET;temp:intralink-;e003;326;512 2. Correlation; f2 ¼1

σxσy

Xi

Xj

½ðijÞpði; jÞ − μxμy�;

(3)

EQ-TARGET;temp:intralink-;e004;326;457

3. Energy; f3 ¼Xi

Xj

pði; jÞ2; (4)

EQ-TARGET;temp:intralink-;e005;326;419

4. Homogeneity; f4 ¼Xi

Xj

pði; jÞ1þ ði − jÞ2 ; (5)

where μx, μy, σx, and σy are the means and the standarddeviations of the corresponding distributions. p is the jointprobability of two pixels with specific intensity levels asdescribed in Eq. (1). Ng is the number of gray levels inthe image. i and j are the two specific intensity levels.

3.2 Feature Extraction-Based Energy DispersiveX-Ray Spectroscopy Spectra

In addition to the FE-SEM/EDX images, the scanning ofthe cervical cells specimen area using point EDX and/orline scanning EDX can produce a spectrum of the area. Thescanning time was determined for the same duration for eachproduced spectrum. In this study, as presented in Sec. 2, theelemental distributions of cells were scanned in a certain areaof the FE-SEM images. Figure 3 describes the visible spec-tral differences among normal, LSIL, and HSIL cells. In thenormal cells, the carbon (C), nitrogen (N), oxygen (O), andsodium (Na) elemental peaks were detected strongly bythe FE-SEM/EDX system. LSIL cells consistently had Cand O elements, and they sometimes had N and Na elements.Meanwhile, C and O elements were higher in the HSIL cellswhile the N and Na elements were not present. Therefore, C,N, O, and Na elements can be used as features to differentiateamong the normal, LSIL, and HSIL cells.

The elemental distribution in the cervical cells, as tabu-lated in Table 2, had standard peak positions in terms ofkeV based on the periodic table of elements. By using C,N, O, and Na elements as the discriminative features, a signal

(a) (b) (c)

Fig. 2 Cervical precancerous cells FE-SEM images; (a) normal, (b) LSIL, and (c) HSIL. *Note: arrowheads refer to viruses.

Optical Engineering 103110-4 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

AuNa

N

AuO

C

0 2 4 6 8 10keVFull scale 2011 cts cursor: 10.881 (0 cts)

Spectrum 1

AuNa

N

AuO

C

0 2 4 6 8 10keVFull scale 2021 cts cursor: 10.881 (0 cts)

Spectrum 2

AuNa

N

AuO

C

0 2 4 6 8 10keVFull scale 2013 cts cursor: 10.853 (0 cts)

Spectrum 3

AuNa

N

AuO

C

0 2 4 6 8 10keVFull scale 2043 cts cursor: 10.881 (0 cts)

Spectrum 4

Au

N

NaAuO

C

0 2 4 6 8 10keVFull scale 2018 cts cursor: 10.853 (0 cts)

Spectrum 5

AuNa

O Au

C

0 2 4 6Full scale 2030 cts cursor: 10.855 (0 cts)

Au

N

AuO

C

0 2 4 6Full scale 2033 cts cursor: 10.855 (0 cts)

AuO Au

C

0 2 4 6Full scale 2030 cts cursor: 10.828 (0 cts)

AuAuO

C

0 2 4 6Full scale 2003 cts cursor: 10.855 (0 cts)

AuNa AuO

C

0 2 4 6Full scale 2043 cts cursor: 10.855 (0 cts)

AuAuO

C

0 2 4 6Full scale 2018 cts cursor: 10.868 (0 cts)

AuO Au

C

0 2 4 6Full scale 2031 cts cursor: 10.868 (0 cts)

AuAuO

C

0 2 4 6Full scale 2026 cts cursor: 10.868 (0 cts)

AuO Au

C

0 2 4 6Full scale 2025 cts cursor: 10.868 (0 cts)

AuO

Au

C

0 2 4 6Full Scale 2014 cts Cursor: 10.868 (0 cts)

(A1) (B1)

(A2)

(A3)

(A4)

(A5)

(A6)

(C1)

(B2)

(B3)

(B4)

(B5)

(B6)

(C4)

(C3)

(C2)

(C5)

(C6)

10 µm 10 µm 10 µm

Fig. 3 Description of the visible spectral difference among normal, LSIL, and HSIL cellsimages *Noted: (A1) normal cells image with five EDX scanning area; (A2)–(A6) EDX spectra ofthe scanning area; (B1) LSIL cells image with five EDX scanning area; (B2)–(B6) EDX spectra ofthe scanning area; (C1) HSIL cells image with five EDX scanning area; (C2)–(C6) EDX spectraof the scanning area.

Optical Engineering 103110-5 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

processing algorithm was utilized in this study. The height ofthe peak elements, the corrected area under the peaks, theratio of peak elements, and the ratio of corrected area underthe peaks were calculated from the FE-SEM/EDX spectra.The maximum peak and base points algorithm were appliedto the spectra ½sðxÞ� to automatically detect the peaks andbase points, as shown in Fig. 4 (used in calculating the cor-rected area), in the range of the 0 to 10 keV region. Afterdetecting the peaks and the base points, calculation of theheight of the peaks and the corrected area under the peakswere performed and the ratio of the peaks and correctedarea were calculated.

An automatic features extraction algorithm, the image andspectra of cervical cells features extraction (ISCFE) algo-rithm for extracting the discriminant features in FESEM/EDX data, was introduced in this section. Based on theprevious section shown in Fig. 1, the ISCFE algorithm isa combination of image and spectra processing techniquesfor FE-SEM/EDX data as follows:

– Implemented two preprocessing algorithms to the origi-nal FE-SEM/EDX images of the cervical cells: contraststretching and morphological opening algorithms.

– Set four parameters of the GLCM technique: the regionsize, quantization levels, displacement, and orienta-tion value.

– Applied the parameters and calculated the co-occur-rence matrix using Eq. (1).

– Applied the GLCM technique to extract contrast,correlation, energy, and homogeneity features setusing Eqs. (2)–(5), respectively, then arranged indataset A.

– Detected the important peaks in the FE-SEM/EDXspectra of the cervical cells as tabulated in Table 2using the maximum height value between a and bfor the important peaks, where a and b were −3and þ3 of the location of the important peaks heights.

– Found the peak height of the important peaks.

Fig. 4 Scanning EDX area in the FE-SEM image and FE-SEM/EDX spectrum with its peak and basepoints.

Table 2 Elements in cervical cell samples and their peak positions byusing FE-SEM/EDX spectra.

Features based onelemental distribution Peak positions (keV)

Carbon (C) 0.27

Nitrogen (N) 0.39

Oxygen (O) 0.53

Sodium (Na) 1.04

Table 3 The features of cervical cells based on FE-SEM/EDXspectra.

Features Detail peaks and corrected area ratio

Features 1 p1∕p2

Features 2 p1∕p3

Features 3 p1∕p4

Features 4 p2∕p3

Features 5 p2∕p4

Features 6 P3∕p4

Features 7 ca1∕ca2

Features 8 ca1∕ca3

Features 9 ca1∕ca4

Features 10 ca2∕ca3

Features 11 ca2∕ca4

Features 12 ca3∕ca4

Optical Engineering 103110-6 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

– Calculated the ratio of the specific peaks (i.e.,p1; p2; : : : , and so on) as determined in Table 3where, p1, p2, p3, and p4 were peaks of C, N, O,and Na, respectively, and arranged in dataset B.

– Determined the prominent base points of the importantpeaks, as presented in Fig. 4, to be used in featurescalculation.

– Calculated the corrected area (ca) under the peaks (i.e.,ca1; ca2: : : , and so on) based on the detected basepoints according to Eq. (6)

EQ-TARGET;temp:intralink-;e006;63;636Ca ¼Zx2

x1

sðxÞ − ½sðx1Þ þ sðx2Þ� × ðx2 − x1Þ2

; (6)

where dðxÞ is the height of the spectrum, x1 and x2 arethe base points for specific important peaks, and sðx1Þand sðx2Þ are the height of the base points.

– Calculated the ratio of the area under the peaks as tabu-lated in Table 3, where ca1, ca2, ca3, and ca4 were the

corrected area under the peaks of C, N, O, and Na,respectively, and arranged in dataset C.

– Combined datasets A, B, and C as a matrix input datafor classification screening part.

Based on the FE-SEM/EDX images, there were four mainfeatures based on intensity variation (contrast and homo-geneity), intensity distribution (correlation), and intensitypower difference (energy). For the contrast feature, thevalue of contrast was high if the variation in image intensitywas high. On the contrary, the homogeneity feature was theopposite of the contrast value, where a high-contrast imageindicated a nonhomogeneous region. Meanwhile, for the cor-relation feature, the value of the correlation was high if theimage had strong gray level linear dependence between thepixels at the specified positions relative to each other. Then,the energy value was high if the pixel intensities were con-centrated in certain coordinates for the energy feature. Forexample, for normal, LSIL, and HSIL in Fig. 5, the contrastfeature in the normal image was lower than the LSIL andHSIL images, and the homogeneity of the normal imagewas higher. Meanwhile, the correlation for LSIL and

Original images Output of contrast stretching Output of morphological opening

Normal

LSIL

HSIL

Fig. 5 Presentation of normal, LSIL, and HSIL images and their contrast stretching and morphologicalopening output images.

Optical Engineering 103110-7 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

HSIL images was higher than normal images. In LSIL andHSIL images, the energy value was lower than the normalimage due to the existence of the virus in the LSIL andHSIL images, as the pixel intensities were not concentratedin certain coordinates.

Based on the selected parameters for the GLCM tech-nique, eight co-occurrences matrixes were created for thefour features calculation. Thus, eight values for each feature(i.e., contrast, correlation, energy, and homogeneity) wereobtained depending on the displacement d ¼ 1, 2, andthe orientation θ ¼ 0 deg, 45 deg, 90 deg, and 135 deg.Based on the rotational invariance of the technique, foreach distance, a one value for each feature was eventuallyobtained. The total number of features taken into account ofthe two displacements could be eight features on the FE-SEM/EDX images.

Based on the previous explanation, the features of cervicalcells based on the FE-SEM/EDX spectra consisted of peakheight and corrected area under the peaks. The ratio of thepeak height and the corrected area was used as a featurebased on spectra in this study. The total features based onthe ratios of the peak height and corrected area under thepeaks (i.e., C, N, O, and Na element peaks) was 12, as tabu-lated in Table 3. A total of 20 features were extracted fromthe FE-SEM/EDX images and spectra and fed as the input tothe automated screening system.

3.3 Classifications

Data from four quantization levels (i.e., 5, 9, 15, and 32) ofthe GLCM technique for images were compared to obtainthe optimal results. The quantization levels were used forthe ISCFE algorithm. The capability of the ISCFE algorithmfor the features extraction techniques was tested using a 10-fold cross-validation and applied for data distribution. In theclassification stage, a discriminant analysis (DA) techniquewas applied. This screening system used validity measuresof normal and abnormal (i.e., LSIL and HSIL) classes.Confusion matrices were used for evaluation purposes.Comparison of performance between this study and relatedpublished work was performed by calculating the accuracy,sensitivity, and specificity of the screening results. The cer-vical cell data from the systems were compared with those ofcytology (gold standard).

4 Results and DiscussionsBased on the previous explanation, the ISCFE algorithmextracted features from FESEM/EDX data (i.e., imagesand spectra). FE-SEM/EDX images provided statisticalfeature information based on pixel intensities, specifically,contrast, correlation, energy, and homogeneity. Furthermore,the FE-SEM/EDX spectra can provide information on theratio of the peaks and the ratio of the corrected areaunder the peaks.

Figure 3 shows the FE-SEM/EDX data consisting ofimages and spectra of the cells on the FE-SEM images.From 150 FE-SEM images, we have obtained 700 spectraof the cells on the images of 300 normal, 200 LSIL, and200 HSIL spectra. The features from the FE-SEM imagewere considered as an individual prospective cell for classi-fication purposes.

Based on the DA technique, four confusion matrixes of 10times 10 cross-validations for each tested quantization level

in the GLCM technique in the ISCFE algorithm were usedfor evaluation. The comparison results are presented inTable 4. Based on Table 4, the screening performances ofcervical precancerous cells achieve more than 97% of accu-racy, sensitivity, and specificity when the FE-SEM/EDX datawere used as input for the automated system. The results ofthe automated system were achieved when the ISCFE algo-rithm with nine quantization levels were 98.2%, 99.0%, and98.0% of accuracy, sensitivity, and specificity, respectively.Meanwhile, with 15 quantization levels, 98.0% accuracy,99.0% sensitivity, and 97.2% specificity were obtained.When the ISCFE algorithm with 32 quantization levels wasoperated, the system achieved 98.0%, 99.0%, and 97.2% ofaccuracy, sensitivity, and specificity, respectively. Amongthe results, the ISCFE algorithm with nine quantizationlevels was chosen for an optimal result of the system. Theresults achieved were very stable. By using nine quantizationlevels in the GLCM, it still represented well the informationof the images, and it did not incur as much computationalload as other higher quantization levels (i.e., 15 and 32 quan-tization level). Meanwhile, if we chose a low enough numberof quantization levels, we would not be able to preserveenough information in the images.

Based on the screening system results in this study, theFE-SEM/EDX image and spectra features, ISCFE algorithm,and DA classifier were used for a screening system as asecond opinion for human experts to classify cervical precan-cerous cells. To show the capability of our proposed system,the performance results were compared with other publishedcervical precancerous screening systems. The comparison ofperformance of the cervical screening system and severalpreviously mentioned approaches are tabulated in Table 5in terms of accuracy, sensitivity, and specificity.

Five systems, systems A,13 B,8 C,36 D,14 and E (proposedsystem) were included for comparison. In the automated

Table 5 Comparison between the published systems and the pro-posed system results. Note: system A13, B8, C36, D14, and E (proposedsystem).

Performances

Comparison of systems (%)

A B C D E

Accuracy 93.5 83.5 75.0 85.4 98.2

Sensitivity 93.2 73.0 70.0 84.5 99.0

Specificity 94.1 94.0 80.0 86.4 98.0

Table 4 Comparison of DA classification result with several quanti-zation levels of GLCM technique for both images and FESEM/EDXspectra features.

Systemperformances (%)

Quantization level

5 9 15 32

Accuracy 97.1 98.2 98.0 98.0

Sensitivity 97.0 99.0 99.0 99.0

Specificity 98.0 98.0 97.2 97.2

Optical Engineering 103110-8 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

systems, the first three groups used images8,13,36 while therest used spectra.14 Among the systems based on the images,FISH13 presented a better classification performance with93.5%, 93.2%, and 94.1%, of accuracy, sensitivity, andspecificity, respectively. Meanwhile, as shown in Table 5,system B, which used cervicography, had better resultsthan system C, which used colposcopy. Meanwhile, the sys-tem using the Raman spectra14 applied preprocessing usinga Savitzky–Golay smoothing filter, principal componentanalysis, and linear discriminant analysis-based multivariateanalysis for classification purposes. The system achieved85.4%, 84.5%, and 86.4% of accuracy, sensitivity, and speci-ficity, respectively.

Therefore, based on the aforementioned explanation, oursystem can better differentiate the cervical cells as it appliedthe combination of image and signal data. By using bothimage and spectra data, there was a significant increase ininformation, so the tendency to produce accurate resultswas higher. Our proposed screening system had better clas-sification performances where it achieved 98.2%, 99.0%,and 98.0% of accuracy, sensitivity, and specificity, respec-tively. In addition, our system ran the 10 times 10 cross-val-idation technique to determine its classification accuracy inan unbiased manner. Therefore, based on the results, our sys-tem is better than other published systems, as presented inTable 5.

5 ConclusionsIn this paper, the capability of FE-SEM/EDX to scanmaterial structures at the microlevel and characterize thematerial with its elemental properties inspired the develop-ment of an FE-SEM/EDX-based cervical cancer screeningsystem. The developed computer-aided screening systemconsisted of two parts: automatic features extraction andclassification. The automatic feature extraction applied anISCFE algorithm, which extracted significant informationfrom both FE-SEM/EDX image and spectra data.Classification-based DA was employed. The performanceof this system is better than other current published systems.Better results had been achieved due to more informationbased on images and spectra compared to other systems.Therefore, a cervical precancerous screening system basedon FE-SEM/EDX data can provide accurate results as indi-cated by the high accuracy, sensitivity, and specificity values.

AcknowledgmentsThis research was supported by UM Postgraduate ResearchFund PG083-2013B, University of Malaya Research Grant,UMRG RP040C-15HTM from the Ministry of HigherEducation, Malaysia. I thank Universitas Abdurrab for thechanges to develop research on the next step.

References

1. IARC: GLOBOCAN, Estimated Cancer Incidence, Mortality andWorldwide in 2012, GLOBOCAN IARC, World HealthOrganization, France (2012).

2. H. S. Cronjé, “Screening for cervical cancer in the developing world,”Best Pract. Res. Clin. Obstet. Gynaecol. 19, 517–529 (2005).

3. N. A. Seman et al., “Classification of malaria parasite species based onthin blood smears using multilayer perceptron network,” Int. J. Comput.Internet Manage. 16, 46–52 (2008).

4. R. Lozano, “Comparison of computer-assisted and manual screening ofcervical cytology,” Gynecol. Oncol. 104, 134–138 (2007).

5. B. Rekhi et al., “Spectrum of neuroendocrine carcinomas of theuterine cervix, including histopathologic features, terminology,

immunohistochemical profile, and clinical outcomes in a series of50 cases from a single institution in India,” Ann. Diagn. Pathol. 17,1–9 (2013).

6. N. M. Harandi et al., “An automated method for segmentation ofepithelial cervical cells in images of ThinPrep,” J. Med. Syst. 34,1043–1058 (2010).

7. V. van Raad, X. Zhiyun, and L. Holger, Lesion Margin Analysis forAutomated Classification of Cervical Cancer Lesions, STI MedicalSystems, Honolulu (2006).

8. A. Alush, H. Greenspan, and J. Goldberger, “Automated and interactivelesion detection and segmentation in uterine cervix images,” IEEETrans. Med. Imaging 29, 488–501 (2010).

9. B. Kajtár et al., “Automated fluorescent in situ hybridization (FISH)analysis of t(9;22)(q34;q11) in interphase nuclei,” Cytometry, Part A69, 506–514 (2006).

10. Y. Jusman et al., “Automated cervical precancerous cells screeningsystem based on Fourier transform infrared spectroscopy features,”J. Biomed. Opt. 21, 075005 (2016).

11. Y. Jusman, S. C. Ng, and N. A. Abu Osman, “Intelligent screeningsystems for cervical cancer,” Sci. World J. 2014, 1–15 (2014).

12. N. A. Mat-Isa, M. Y. Mashor, and N. H. Othman, “An automatedcervical pre-cancerous diagnostic system,” Artif. Intell. Med. 42, 1–11(2008).

13. X. Wang et al., “Automated detection and analysis of fluorescent in situhybridization spots depicted in digital microscopic images of Pap-smearspecimens,” J. Biomed. Opt. 14, 021002 (2009).

14. S. Rubina et al., “Raman spectroscopic study on classification ofcervical cell specimens,” Vib. Spectrosc. 68, 115–121 (2013).

15. S. G. El-Tawil et al., “Comparative study between Pap smear cytologyand FTIR spectroscopy: a new tool for screening for cervical cancer,”Pathology 40, 600–603 (2008).

16. J. Pawley, “The development of field-emission scanning electronmicroscopy for imaging biological surfaces,” Scanning 19, 324–336(1997).

17. Y. Jusman, S. C. Ng, and N. A. Abu Osman, “Investigation of CPD andHMDS sample preparation techniques for cervical cells in developingcomputer aided screening system based on FE-SEM/EDX,” Sci. WorldJ. 2014, 1–11 (2014).

18. S. G. Kaminskyj and T. E. Dahms, “High spatial resolution surface im-aging and analysis of fungal cells using SEM and AFM,” Micron 39,349–361 (2008).

19. T. Congiu et al., “Scanning electron microscopy examination of endo-thelium morphology in human carotid plaques,” Micron 41, 532–536(2010).

20. P. Kilonzo, A. Margaritis, and M. Bergougnou, “Effects of surfacetreatment and process parameters on immobilization of recombinantyeast cells by adsorption to fibrous matrices,” Bioresour. Technol.102, 3662–3672 (2011).

21. S. Parameswaran and R. S. Verma, “Scanning electron microscopypreparation protocol for differentiated stem cells,” Anal. Biochem.416, 186–190 (2011).

22. R. Wahab et al., “Fabrication and growth mechanism of ZnO nanostruc-tures and their cytotoxic effect on human brain tumor U87, cervicalcancer HeLa, and normal HEK cells,” J. Biol. Inorg. Chem. 16,431–442 (2011).

23. C. P. Taylor et al., “Oxygen/glucose deprivation in hippocampal slices:altered intraneuronal elemental composition predicts structural andfunctional damage,” J. Neurosci. 19, 619–629 (1999).

24. Y. Jusman and N. A. M. Isa, “Denoising-based cascaded algorithms forsmoothing of different level additive white gaussian noise-corruptedspectra,” in Proc. of the 2011 3rd Int. Conf. on Software Technologyand Engineering (ICSTE 2011), pp. 425–430 (2011).

25. H. Kobatake et al., “Computerized detection of malignant tumors ondigital mammograms,” IEEE Trans. Med. Imaging 18, 369–378 (1999).

26. A. Busch, W. W. Boles, and S. Sridharan, “Texture for script identifi-cation,” IEEE Trans. Pattern Anal. Mach. Intell. 27, 1720–1732 (2005).

27. B. Hernández et al., “Visual learning of texture descriptors for facialexpression recognition in thermal imagery,” Comput. Vis. ImageUnderstanding 106, 258–269 (2007).

28. A. Oliver et al., “A review of automatic mass detection and segmen-tation in mammographic images,” Med. Image Anal. 14, 87–110(2010).

29. R. Maani, S. Kalra, and Y.-H. Yang, “Robust volumetric texture clas-sification of magnetic resonance images of the brain using local fre-quency descriptor,” IEEE Trans. Image Process. 23, 4625–4636 (2014).

30. R. W. Conners and C. A. Harlow, “A theoretical comparison of texturealgorithms,” IEEE Trans. Pattern Anal. Mach. Intell. PAMI-2, 204–222(1980).

31. P. Gong, D. J. Marceau, and P. J. Howarth, “A comparison of spatialfeature extraction algorithms for land-use classification with SPOTHRV data,” Remote Sens. Environ. 40, 137–151 (1992).

32. Y. Jusman et al., “A system for detection of cervical precancerous infield emission scanning electron microscope images using texturefeatures,” J. Innov. Opt. Health Sci. 10, 1650045 (2016).

33. F.-M. E. Uzoka, J. Osuji, and O. Obot, “Clinical decision supportsystem (DSS) in the diagnosis of malaria: a case comparison of two

Optical Engineering 103110-9 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

soft computing methodologies,” Expert Syst. Appl. 38, 1537–1553(2011).

34. L. K. Soh and C. Tsatsoulis, “Texture analysis of SAR Sea ice imageryusing gray level co-occurrence matrices,” IEEE Trans. Geosci. RemoteSens. 37, 780–795 (1999).

35. I. Hassan et al., “Radiomic texture analysis mapping predicts areas oftrue functional MRI activity,” Sci. Rep. 6, 25295 (2016).

36. S. Y. Park et al., “Domain-specific image analysis for cervical neoplasiadetection based on conditional random fields,” IEEE Trans. Med.Imaging 30, 867–878 (2011).

Yessi Jusman is a lecturer in the Informatics Engineering Depart-ment, Faculty of Engineering, Universitas Abdurrab, Pekanbaru,Riau, Indonesia. She received her degree in the Electrical Engineer-ing Department, Faculty of Engineering, Andalas University, Padang,Sumatera Barat, Indonesia. She received her master’s degree fromthe University of Science Malaysia, Penang, Malaysia. She startedher PhD in the Faculty of Engineering, University of Malaya, KualaLumpur, Malaysia, in 2013. She has published 18 papers. Her fieldsof interests are signal and image processing, algorithms, neural net-work, biomedical engineering, and intelligent systems.

Ng Siew-Cheok is a senior lecturer in the Department of BiomedicalEngineering, Faculty of Engineering, University of Malaya, Malaysia.He obtained his bachelor’s degree, and postgraduate degrees fromthe University of Malaya. He has published more than 15 papersin his fields. His fields of interests are rehabilitations, biomedical engi-neering, and biomedical signal processing.

Khairunnisa Hasikin is a senior lecturer in the Department ofBiomedical Engineering, Faculty of Engineering, University of

Malaya, Malaysia. She obtained her bachelor’s degree and master’sdegree from the University of Malaya. Her PhD was from UniversitiSains Malaysia, Penang, Malaysia. Her fields of research are medicalimage processing and analysis, expert systems, and medical infor-matics. She has published more than 17 papers.

Rahmadi Kurnia is a senior lecturer in electrical engineering,Andalas University, Indonesia. He received his degree and master’sdegree from the Indonesia University (UI), Indonesia. His PhD wasfrom Saitama University, Japan. His fields of research are telecom-munications, computer vision, and robotics. He has published morethan 19 papers.

Noor Azuan Bin Abu Osman is currently a professor in theDepartment of Biomedical Engineering, Faculty of Engineering,University of Malaya, Malaysia. He received his bachelor’s degreefrom Bradford University, UK, his master’s degree from StrathclydeUniversity, Scotland, and his PhD from Strathclyde University,Scotland. He has published more than 400 papers.

Kean Hooi Teoh is medical lecturer in the Pathology Department,University of Malaya. He received his bachelor’s degree from theRoyal College of Surgeons in Ireland, UK. His master’s degree isfrom the Faculty Medicine University of Malaya, Kuala Lumpur,Malaysia. His fields of research are surgeons and pathology. Hehas published more than 10 papers.

Optical Engineering 103110-10 October 2016 • Vol. 55(10)

Jusman et al.: Computer-aided screening system for cervical precancerous cells. . .

Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 11/04/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx