1-s2.0-S0169260714002521-main

Please citePrograms B

ARTICLE IN PRESSCOMM-3824; No. of Pages 13c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 4 ) xxxxxx

jo ur nal ho me p ag e: www.int l .e lsev ierhea l t h.com/ journa ls /cmpb

Gold- wohead segmentation

Violeta C a,c a,b c d

Nancy Ha Departmenb ORAND S.c Laboratory(CEDAI SpAFaculty of Md LaboratoryUniversity o

a r t i c

Article histor

Received 25

Received in

31 May 2014

Accepted 26

Keywords:

Infertility

Morphologic

Sperm head

Sperm head

Acrosome s

Nucleus seg

CorresponE-mail a

lsarabia@m

http://dx.do0169-2607/ this article in press as: V. Chang, et al., Gold-standard and improved framework for sperm head segmentation, Comput. Methodsiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

hang , Jose M. Saavedra , Victor Castaneda , Luis Sarabia ,itschfelda, Steffen Hrtel c,

t of Computer Science, University of Chile, Beauchef 851, 4th Floor, Santiago, ChileA., Estado 360, 7th Floor, Ofce 702, Santiago, Chile

for Scientic Image Analysis (SCIAN-Lab), Centro de Espermiograma Digital Asistido por Internet), Biomedical Neuroscience Institute (BNI), Program of Anatomy and Developmental Biology (ICBM),edicine, University of Chile, Independencia 1027, Santiago, Chile

of Spermiogram, Program of Anatomy and Developmental Biology (ICBM), Faculty of Medicine,f Chile, Independencia 1027, Santiago, Chile

l e i n f o

y:

November 2013

revised form

June 2014

al analysis

detection

segmentation

egmentation

mentation

a b s t r a c t

Semen analysis is the rst step in the evaluation of an infertile couple. Within this process,

an accurate and objective morphological analysis becomes more critical as it is based on the

correct detection and segmentation of human sperm components. In this paper, we present

an improved two-stage framework for detection and segmentation of human sperm head

characteristics (including acrosome and nucleus) that uses three different color spaces. The

rst stage detects regions of interest that dene sperm heads, using k-means, then candidate

heads are rened using mathematical morphology. In the second stage, we work on each

region of interest to segment accurately the sperm head as well as nucleus and acrosome,

using clustering and histogram statistical analysis techniques. Our proposal is also charac-

terized by being fully automatic, where a user intervention is not required. Our experimental

evaluation shows that our proposed method outperforms the state-of-the-art. This is sup-

ported by the results of different evaluation metrics. In addition, we propose a gold-standard

built with the cooperation of a referent expert in the eld, aiming to compare methods for

detecting and segmenting sperm cells. Our results achieve notable improvement getting

above 98% in the sperm head detection process at the expense of having signicantly fewer

false positives obtained by the state-of-the-art method. Our results also show an accurate

head, acrosome and nucleus segmentation achieving over 80% overlapping against hand-

segmented gold-standard. Our method achieves higher Dice coefcient, lower Hausdorff

distance and less dispersion with respect to the results achieved by the state-of-the-art

method. 2014 Elsevier Ireland Ltd. All rights reserved.

ding author. Tel.: +56 2 29786366.ddresses: [email protected] (V. Chang), [email protected] (J.M. Saavedra), [email protected] (V. Castaneda),ed.uchile.cl (L. Sarabia), [email protected] (N. Hitschfeld), [email protected] (S. Hrtel) .

i.org/10.1016/j.cmpb.2014.06.018 2014 Elsevier Ireland Ltd. All rights reserved.standard and improved frame rk for sperm

Please cite framPrograms B

ARTICLE IN PRESSCOMM-3824; No. of Pages 132 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 4 ) xxxxxx

1. Introduction

Infertility is a problem that affects up to 15% of couplesworldwide [1]. This condition has emotional and physiologicalimplicationtion [2]. A sthe rst stebasis for all[4]. A typicvitality, anaddition, than importasample [5].is a difcultis consider

Therefomalities, sdue to impin the semfor many da challengiratories deshowed thaysis was pnon repeata[5,7]. Despiphological analysis of lenge conceintra obser[811]. Theization of t[12,13]. A soovercome t

Overall,(size of theetc.) and ptail, coiledaccording tsperm clas

In this detecting aable detectall posterioapproach isprocessingIn additioninstead of a gold-standataset wain the eldhundred spstandard hwith the onthe past anmethod.

1 Available

Our main contribution is the application of a clusteringalgorithm for detecting sperm heads, combining differentcolor spacealgorithm

. Thiest fs papsearcmm

spetand

In Seng oued gn 5. T

Re

s secutom

nonrm. Tationed.

comeledan sent

of vs [10utomped eld [re arlly, orkmenorm.ity dct thrm h

witber o

Hourivedestimsi e

lowsper

thathm o-s

iecesectivtage,ethsomhat, ed in this article in press as: V. Chang, et al., Gold-standard and improved

s including stress, depression or sexual dysfunc-emen analysis according to standard criteria [3], isp in the evaluation of the male factor and sets the

posterior steps for medical treatment of the coupleal spermiogram considers concentration, motility,d/or the fragmentation of the spermatic DNA. Ine morphology of the sperm cells is considered asnt parameter to elucidate the potential fertility of a

The classication of abnormal sperm morphology task since the spectrum of possible malformationsably wide [6].re, it is important to objectively quantify abnor-uch as double-headed or multiple-tailed sperm,lications of the presence of these abnormalitiesen sample [5]. However, there has been evidenceecades that the aforementioned quantication isng task. In 1966, a comparative study in 47 labo-dicated to human sperm morphological analysist the traditional method of performing the anal-

ersonality oriented, as well as subjective, qualitative,ble and difcult to teach to students and technicianste of the fact that the classication rules for mor-semen analysis have been simplied [3], the visualsperm morphology still presents a substantial chal-rning reproducibility and objectivity, and inter andver variability still presents a well known problemre are many authors revealing a lack of standard-he methods used in laboratories in many countriesphisticated computational analysis might help tohese problems.

the evaluation of cellular and sub-cellular regions sperm head, tail length, residual cytoplasm area,attern recognition (multiple heads or tails, absent

tail, etc.) are required for categorizing defectso normal and abnormal sperm denitions in visualsication under the microscope [14].paper, we present an improved framework fornd segmenting human sperm heads, since a reli-ion and segmentation presents the rst step forr classication algorithms. This fully automatic

based on a clustering method as well as on image techniques especially adapted for this application., we propose to combine different color spaces,using only RGB color space. We also introducedard1 for head sperm parts segmentation. Thiss built with the cooperation of a referent expert

and contains twenty images with more than twoerm cells plus hand-segmented masks. This gold-as been used to evaluate and compare our resultsly reproducible method that has been published ind therefore presents our state-of-the-art reference

in http://morfologia.cedai.cl/public/.

pointsthe qu

Thithe reand cotion ofgold-stion 3.applyiproposSectio

2.

In thitially athoughof speapplicdescrib

Thebeen fuof humimplemdegreeniciansemi-adevelonary

Thematicaframewfor segTransfintensto selethe speimateda numing thethe deof the

Nacells inof the imagealgorit

A twmid-pan objrst sOtsu mThen, After tenclosiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

ework for sperm head segmentation, Comput. Methods

s. Another contribution is the proposal of a novelto determine which direction the sperm heads is a very important issue for posterior stages inor an accurate morphological analysis.er is organized as follows. In Section 2 we reviewh work in the area, focusing on scientic papersercial applications whose main goal is segmenta-rm cells. Our proposed framework as well as theard, are presented and described in detail in Sec-ction 4 we present the description of the results ofr approach and the state-of-the-art method to theold-standard, which we discuss in more depth inhe conclusions can be found in Section 6.

lated work

tion, the most prominent approaches for par-ated sperm detection and analysis are discussed,e of them presents a full morphological analysishe characteristics of the three major commercials for the morphological analysis of sperm are also

puter assisted sperm morphology assessment has by the inherent lack of objectivity in the evaluationperm morphology, the difculty in standardizing,ing and controlling manual methods, and the highariation within and between laboratories and tech-]. With the aim of providing more objective results,atic methods based on image analysis have been

[15,16,5,17,18], some of them applied to the veteri-1921].e few approaches to evaluate semen samples auto-even though none of them proposes a complete. The work of Park et al. [22] presents an approachtation of sperm heads using the strategic Hough

For each sperm in the image, the authors use theifference between the sperm head and backgrounde region of interest (ROI) for the segmentation ofead. The boundary of the sperm head was approx-

h an ellipse. The resulting ellipse is represented byf parameters that have been investigated by apply-gh Transform strategically. Finally, the authors use

boundary to calculate morphological parametersated sperm head.

t al. [4] proposed an algorithm for nding sperm contrast images, with the added value of detectionm tail for discarding or not some particles in the

could be similar in size with a sperm head. Thets ellipses for detecting the sperm heads.

tage method for segmentation of sperm heads and was presented by Carrillo et al. [23,24] looking fore analysis of human sperm morphology. At the

the objects obtained by thresholding using theod [25] are classied through histogram analysis.e particles are removed according to their size.each sperm cell detected (head and mid-piece) is

a bounding box. Then, each sperm cell is extracted


ARTICLE IN PRESSCOMM-3824; No. of Pages 13c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 4 ) xxxxxx 3

Fig. 1 Mo ight size: 277 th oand tail of com5 m long ce: 5

from the oauthors prapplying a nfusion metfollowed bystructed usof our knowand other r

Abbiramcal classicusing Matlaobject segmdetector.

Bijar et sperm acrotails. The Bayesian cmaximizatidenticatilarity indexresults thatexperimendation of th

In the cthe interfaassisted syCASA (Comtems are usreproductivwhen the order to gusystems aration, it is rand qualitydyes, type osen [3033,

mols of

(Ham CAB (

speSCA lies.alyzuenry reent

e syequirphology of the normal human sperm. (a) Representative br 144 pixels 58 31 m). (b) Manually segmented ground-trustained spermatozoa. (c) Schematic drawing of the principaland 3 m wide. Acrosome: 4070% of the head area. Mid-pie

riginal RGB color image. At the second stage, theoposed to segment the head and mid-piece byth-fusion method to the enhanced image. The nth-

hod is based on nth-level thresholding of an image intersection with n special growing masks, con-ing prior object morphological models. To the bestledge, this proposal is the state-of-the-art method,esearch works compare their results with it [26,27].y et al. [28] proposed a method for morphologi-ation of sperm cells either as normal or abnormalb. One of the steps in their proposal is regardingentation in microscopic images, using sobel edge

al. [27] proposed a method for segmentation of

Comsis tooSystem SpermedeaLreportswhile anomaical ansubseqhas veequipmof thescians r this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

some, nucleus, mid-piece and identication ofsegmentation step is performed by means of alassier which uses entropy based expectationion and a Markov random eld. For sperm tailon, the authors proposed to use a structural simi-

and local entropy techniques. The paper presented outperformed those of Carrillo et al., however thetal framework is so weak that it makes the vali-ose results very difcult.ontext of commercial applications, companies ince of research and development offer computerstems for semen analysis, usually referred to asputer Aided/Assisted Sperm Analysis). CASA sys-ed for research and routine analysis in the area ofe medicine (human or animal). It was in the 1990srst CASA systems appeared on the market [29,6]. Inarantee that the repeatability and validity of CASAe higher than any subjective morphological evalu-equired that the quality of the preparation, choice

of xing, thickness of the preparation, choice off light and adjustment of optics are carefully cho-

14].

techniques

3. Ma

In this sectWe then instage in dewe use to a

3.1. Go

3.1.1. SamSperm samtoxylin/eossperm morwith Ethan

2 Sperm saProgram of AMedicine, UInstitute (IDeld image of a normal human sperm (imagef the sperm: head, acrosome, nucleus, mid-pieceponents of a normal human sperm. Oval head:

m. Tail: 55 m.

n CASA systems that offer morphological analy- human sperm are IVOS Integrated Visual Opticmilton-Thorne Biosciences, Beverly, MA, USA), SCAlass Analyzer (Microptics, Barcelona, Spain) andMedical Technology MTG, Altdorf, Germany). IVOSrm head parameters such as elongation and area,also reports information about mid-piece and tail

medeaLAB offers the most complete morpholog-er, reporting the length of the sperm tail. It istly the most expensive CASA on the market andstrictive requirements for microscopic and camera. Difculties in rendering the sperm tail, the coststems, and the specialized equipment and techni-red to operate them limited the introduction of theework for sperm head segmentation, Comput. Methods

into andrological practice.

terial and methods

ion, we rst introduce the proposed gold-standard.troduce our two-stage framework, describing eachtail. Finally, we discuss the evaluation metrics thatssess the quality of our results.

ld-standard

ple preparationples2 were stained with a modied hema-

in assay, in order to distinguish different parts ofphology (Fig. 1). Briey, the sperm smear was xedol 70% and immersed in Harris hematoxylin for

mples obtained from: (1) Laboratory of Spermiogram,natomy and Developmental Biology (ICBM), Faculty of

-Chile, Santiago, Chile, and (2) Maternal Child ResearchIMI), San Borja Arriaran Hospital, Santiago, Chile.



Fig. 2 Det e witrepresents . (c) Rsperm cells size.stage. Imag f thereader is re

10 s for nucfor ten minimmersed in a pink-ople was wawere air drsamples fo

3.1.2. ImDigital ima(Axiostar Ptive (oil, NA(scA780-54with more segmented

3.2. Alg

The proposour goal is second stagsperm head

3.2.1. DeFirst, we trRGB and Lof different(Algorithmcoefcient hand-segming algoriththe backgrosperm cells

n oneckgrng imn thesideaimme s neection of sperm heads. (a) Original image in RGB color spac ROIs after applying k-means in RGB and L*a*b* color spaces

at border. (d) Yellow color represents ROIs after erasing by e size: 780 580 pixels 164 122 m. (For interpretation oferred to the web version of the article.)

lear staining. Slides were washed with tap waterutes to remove residual staining. Later, slides werein 1% eosin for two minutes to stain the acrosomerange color, mid-piece and tail. Finally, the sam-shed with distilled water for 1 min. Then, samplesied and xed. This staining procedure allows usingr more than one year.

tails) iand baresultistage i

Conrithm are soregion this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

age acquisitionges were acquired using an optical microscopelus, Carl Zeiss Inc., Wetzlar, Germany), a 63 objec-

1.4) with an adapter of 0.63 and a digital cameragc, Basler AG, Ahrensburg, Ger). Twenty imagesthan two hundred sperm cells were captured and

with the cooperation of an expert.

orithm

ed framework consists of two stages. In the rst,to identify the ROIs of sperm heads (Fig. 2a). In thee, we work on each ROI to accurately segment the

as well as the nucleus and acrosome.

tection of sperm headsansform the RGB color space to L*a*b*. We choose*a*b* after experimental evaluation of the impact

color spaces such as RGB, L*a*b*, YCbCr, and YQM 1, step 2). We evaluate Hausdorff distance and Dicevalues for each color space combination againstented masks. Then, we apply the k-means cluster-m looking for separation of the sperm cells fromund (Fig. 2b). We separate the pixels belonging to

(heads, mid-piece and/or residual cytoplasm, and

eliminating(Algorithm

In orderof sperm hpropose to of size r anvolution, wa threshold(Fig. 2c).

Algorithm imRgr: sizsumVerase

1: imLa2: data

imLa3: [clus4: clust5: noBo6: noTa7: returh resulting ROIs marked on it. (b) Blue colored color represents ROIs after erasing tails and

Yellow pixels constitute the nal ROIs of this references to color in this gure legend, the

cluster, and the pixels belonging other structuresound in a second cluster (Algorithm 1, step 3). Theage contains the ROIs that we need for the second

cluster of a smaller area (Algorithm 1, step 4).ring that our detection and segmentation algo-s for an accurate morphological analysis, thereconditions that we must meet. Therefore, theseed to be rened. This renement includesework for sperm head segmentation, Comput. Methods

sperm cells which touch the border of the image 1, step 5).

to eliminate most of the pixels that are not parteads, we use a binary morphology-based idea. Weuse a convolution process with a disk-shape kerneld unitary weight (Algorithm 1, step 6). After con-e remove all pixels with a resulting value below

sumV. We refer to this procedure as eraseTails

1. Detection of sperm heads.b: original imagee of neighborhood for eraseTails: threshold value for sum inside neighborhood in

Tailsb transformRGBtoLAB(imRgb)

[imRgb(1) imRgb(2) imRgb(3) imLab(1)b(2) imLab(3)]ter1,cluster2] kmeans(data,2)er chooseMinorCluster(cluster1,cluster2)rderImage eraseBorderSperms(cluster)ilImage eraseTails(noBorderImage,r,sumV)n nalImage



Fig. 3 Seg returk-means in e smthe head). ion tgrows acco f spe

In the nsperm headsperm headthe whole nucleus anthat we pro

3.2.2. SegFor each indetected caopening anbetween mwith the cowith the Cr2, step 3). W(Fig. 3b) torest (that cportion of tneed to enresidual cydetermine 7). We use step 6) to cpointing di

To deterpropose a tthe orientaX axis and rotate the axis of the3, step 2). portions (Fextreme po10/11). Theindicates thdirection athe whole (Algorithma growing angle [0, 2segmentedeliminate tmaxTs2 thr

hm hiteinT

eforaxT

eforaxTeadinT

fteraxT

fterhitehiteata clusmalmalanglmalaskeadnalIetur

hm whangmentation of sperm heads. (a) Detection of sperm head (as L*a*b* and YCbCr color spaces. (c) smallHead detected as th(d) Fitness of (a) with an ellipse (red) for getting which directrding to growing mask up to (a). (f) Contour superposition o

ext section, we present our approach to segments. After the ROI detection, we individualize each

and work separately with each one. We segmenthead and then process to identify the regions ofd acrosome. Afterward, we describe the algorithmpose for each step.

mentation of sperm headdividual sperm head (Fig. 3a), we rst rene thendidate head by means of applying morphologicald discarding objects whose size are out of rangeinTs1 and maxTs1 (Algorithm 2, step 2). We worklor-opponent dimensions of L*a*b* color space and

component of the YCbCr color space (Algorithme then apply k-means only in the particular ROI

separate the darkest part of the head from theould be acrosome and residual cytoplasm). As thishe head is smaller than the real head (Fig. 3c), welarge it, up to the region of interest, but withouttoplasm or mid-piece. Thus, it is important tothe front direction of the head (Algorithm 2, stepmaxTc as a maximum size threshold (Algorithm 2,onsider a set of pixels as a candidate head whoserection is important.mine the direction in which the head points, wewo-step method (Algorithm 3). First, we determinetion of the sperm head as the angle between the

Algoritwmbmbmhmama

1: w2: w3: d4: [5: s6: s7: [

s8: m9: h

10: 11: r

Algorit

1: this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

the major axis of the ROI, and using this angle, wehead to a horizontal position in which the major

tted ellipse is parallel to the X axis (AlgorithmWe then divide this major axis in three similarig. 3d) and calculate a tness value [34] of the twortions with the tted ellipse (Algorithm 3, steps

portion corresponding to the lowest tness valuee direction to where the head points. The pointingllows us to build a growing mask for segmentinghead, and not only the darkest part of the head

2, step 8). So, the next step consists in settingmask according to head pointing direction and) and apply it to the part of the head previously

(Algorithm 2, step 9). As a nal renement, wehe small and big objects according to minTs2 andeshold values (Fig. 3e).

2: wh3: img4: ma5: x1 6: x2 7: [xc,8: ma9: min

10: valmin

11: valmin

12: if v13: dire14: elsened from Algorithm 1). (b) ROI after applyingallest cluster (this constitutes the darkest part ofhe head points. (e) smallHead detected in (c)rm head segmentation.

2. Segmentation of one sperm head.: binary image with a candidate sperm detecteds1: minimum number of pixels of a sperm heade k-meanss1: maximum number of pixels of a sperm heade k-meansc: maximum number of pixels of a candidate

after kmeanss2: minimum number of pixels of a sperm head

growings2: maximum number of pixels of a sperm head

growing opening(white) eraseBySize(white,minTs1,maxTs1)

[imLab(2) imLab(3) imYCbCr(3)]ter1,cluster2] kmeans(data,2)lHead chooseMinorCluster(cluster1,cluster2)lHead eraseBySize(smallHead,0,maxTc)e,direction] getPointingDirection(white,lHead)

generateGrowingMask(white,angle,direction) makeHeadBigger(white,mask,smallHead)mage eraseBySize(head,minTs2,maxTs2)n nalImage

3. Sperm head direction.ite: binary image with a candidate sperm detectedle getOrientation(white)ework for sperm head segmentation, Comput. Methods

iteRot rotateImage(white,angle)Border getPerimeter(whiteRot)ximum number of columns of imgBorder ceil(maximum/3) 2*x1yc] getCentroid(whiteRot)jorAxis getMajorAxis(whiteRot)orAxis getMinorAxis(whiteRot)

ue1 tnessFunction([xc,yc,majorAxis,orAxis,0],imgBorder,a,x1)

ue2 tnessFunction([xc,yc,majorAxis,orAxis,0],imgBorder,x2,maximum)

alue1 < value2 thenction 0



15: direction 116: end if17: retur

The pargenerate eipossible orgrowing malowing.

1 0 0

1 0 0

1 0 0

1

1

0

A

1 1 1

0 0 0

0 0 0

0

1

1

A

3.2.3. SegThe segmetation of na statisticatics of bothof RGB colspace becanucleus anOtsu threshpixels of th4, step 2).

The nucbecause of in the lightregions arepixels in thuse our proacrosome a

Algorithm whiteimRGr: sizesumVin era

1: imRed2: thresh3: acros4: acros5: acros6: nucle7: nucle8: return

3.3. Evaluation metrics

Evaa gote outly detly dewronmbe

valuositi

ositi

osit

on =

Evaa gods toa is to thput

ed od on

sum

Re

aperd to imprcant

a puds ca

impl a sogmes sh this article in press as: V. Chang, et al., Gold-standard and improved

n angle and direction

ameters returned by Algorithm 3 can be used toght different growing masks, according to all theientations that a sperm head could present. Thesk is created using angle and direction, as the fol-

0 0 1

0 0 1

0 0 1

0 0 0

1 0 0

1 1 0

0 1

0 0

0 0

ngle = 0 Angle = 2

0 0 0

0 0 0

1 1 1

1 1 0

1 0 0

0 0 0

0 0

0 0

0 1

ngle = Angle = 32

mentation of sperm nucleus and acrosomented sperm head is used for the posterior segmen-ucleus and acrosome. To this end, we performedl analysis to determine the intensity characteris-

head components, working on the red channelor space. We use the R channel of the RGB coloruse R offers a better differentiation between spermd acrosome, compared to the G or B channels. Theold calculated in this region allows us to separatee nucleus from pixels of the acrosome (Algorithm

leus region is darker than the acrosome regiona staining effect. Therefore, we pick the biggest ROIer region as the acrosome (Fig. 4c). The segmented

obtained as a difference set operation between thee acrosome and pixels in the nucleus (Fig. 4d). Weposed procedure eraseTails to smooth the resultingnd nucleus.

4. Segmentation of nucleus and acrosome.: binary image with a segmented sperm headB: rgb image with a segmented sperm head

of neighborhood for eraseTails: threshold value for sum inside neighborhoodseTails

imRGB(:,:,1)old OtsuThreshold(imRed)

ome pixels in imRed whose value is > thresholdome eraseTails(acrosome,r,sumV)ome getBiggestROI(acrosome)us imageDifference(white,acrosome)us eraseTails(nucleus,r,sumV)

acrosome and nucleus

3.3.1. Given evaluacorreccorrecber of the nu

To efalse-p

True-p

False-p

Precisi

3.3.2. Given methothe idepared by combe bas[35]) anTable 1

4.

This pmethocould signiis notmetho

Weably asand sement aiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018


luation metrics for detectionld-standard, we have four possible scenarios tor algorithm: true positives (TPs), the number oftected objects, true negatives (TNs), the number oftected non objects, false positives (FPs), the num-gly detected non objects, and false negatives (FNs),r of wrongly non detected objects.

ate detection algorithms, we use true-positive rate,ve rate and precision as follows.

ve rate = TPTP + FN

ive rate = FPFP + TN

TPTP + FP

luation metrics for segmentationld-standard segmentation, we can use different

evaluate the quality of segmentation. In general,that the automatic segmentation S has to be com-e manually segmented image (gold-standard) G,ing some evaluation metrics. These metrics cann spatial overlap measures (e.g. Dice coefcient

distance measures (e.g. Hausdorff distance [36]).marizes the evaluation metrics used in this paper.

sults

presents a gold-standard, aiming to compare ourpreviously published methods [24] or methods thatove our approach in the future. This represents acontribution of our work, because at present thereblic standard ground-truth, so the few existingnnot be evaluated properly.emented Carrillos method [24], since it is not avail-urce code by the authors to compare our detectionntation precision. We obtain a signicant improve-own in the following sections.



Fig. 4 Seg erm(b) Image a 2). (ROI(a)ROI( e (rethis gure le.)

Table 1

Metric Hausdorff distance

Expression H(S, G) = max(h(S, G), h(G, S))h(S, G) = max

sS(min

gGd(s, g))

Range [0,1]Interpretat 0: maximum agreement

between perimeters of Sand G1: maximum disagreement between perimeters

We haveand for cousing Matla

4.1. Tes

So far, a gohuman speset consiststapered, 4526 normal)without no(Fig. 2a). Fohave been the eld. Wcell, consid

4.2. Par

In order toformed difparametersreferred tohood size underlyingfrom the Aare used inof sperm hThe paramstep beforeimum sizemask. The

3 Matlab R

2 Variation of parameters for our proposedod.

eter name Range Best value

2:1:4 3 15:20:55 35s1 70:10:90 80s1 800:100:1000 900c 1200:100:1500 1400s2 30:10:50 40s2 350:50:450 400mentation of nucleus and acrosome. (a) Segmentation of spfter applying statistical-dened threshold (Algorithm 4, stepc), (e) Contour superposition of nucleus (green) and acrosomlegend, the reader is referred to the web version of the artic

Metrics for quality of segmentation.

Dice coefcient

D = 2|SG||S|+|G|

[0,1] ion 0: no spatial overlap

between S and G

1: complete overlap

conducted experiments for parameter estimation,mparison of our approach and Carrillos methodb.3

ting data set

ld-standard data set for detecting and segmentingrm heads does not exist. Our gold-standard data

of 20 images with 264 sperm cells, where 210 (70 pyriform, 52 amorphous, 15 small, 2 round, and

are valid sperm cells (not at the border of image,

Tablemeth

Param

r sumVminTmaxTmaxTminTmaxT this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

ise on it, etc.). Each image has 780 580 pixelsr each of these images, hand-made ground-truthsdesigned under supervision of a referent expert ine generated segmentation masks for each spermering the nucleus and acrosome, among others.

ameter optimization

choose the best set of parameters, we have per-ferent experiments varying the values of seven

that are described as following. The rst two are the procedure eraseTails, related to the neighbor-(r) and the threshold value for the sum in the

neighborhood (sumV). This procedure is calledlgorithms 1 and 3. The remaining ve parameters

Algorithm 2 and are related to the allowed sizeeads at different stages of the segmenting process.eters minTs1 and maxTs1 are used for a renement

applying k-means, while maxTc indicates the max- of sperm heads allowed to grow with a growinglast two parameters, minTs2 and maxTs2 indicate

2013a 8.1.0.604.

Table 3

Paramete

aMin aMax gIntensity tMin tMax

size threshthe variatiochosen accand false n

In addiCarrillos mmentation.referred to allowed forused as a mtMin and tMthis sense, falls withinered a sper head in R channel (as returned by Algorithm 2).c) Only the biggest ROI is kept and smoothed. (d)d). (For interpretation of the references to color inework for sperm head segmentation, Comput. Methods

Variation of parameters for Carrillos method.

r name Range Best value

50:25:400 200500:100:1500 50090:10:150 1202000:500:6000 55002300:500:27,000 24,000

old values for the nal segmented heads. In Table 2n of parameters is shown. The nal values wereording to the best tradeoff between true positiveegative values, as well as Dice coefcient.tion, we have evaluated parameter values forethod. There are ve free parameters in our imple-

The rst two parameters, aMin and aMax arethe minimum and maximum size in pixels, to be

a head of a sperm cell. The parameter gIntensity isaximum threshold value. The two last parameters,ax, represent a range for sperm cell validation. In

if the sum of pixel values of a candidate sperm cell the range [tMin, tMax], the candidate is consid-m cell, in other case it is discarded. In Table 3, we



Fig. 5 ROand numbeour proposmethod (do

Table 4

Precision True positiFalse posit

show the vexperimen

4.3. Spe

The performROC curve.curve (AUCbetter the qcost in termis desired (calculated gold-standinstance of

Our appmethod acca correct donly 23 fala correct d41 false pos95% [24,26]with 39 falfalse positiated metho

4.4. Spe

We performof sperm

calculate two quality measures for our results: the Diceient to assess the accuracy of our results with hand-ntedEuclintatimen

wios mistare th

6 resu

ntatis), ws oof a

7 shropoient s acrringmetientsas fopliedd theme acleucantlerageC curves for sperm head detection. Detection rater of false positives according to the results ofed method (continuous line) and Carrillostted line) versus hand-segmented masks.

Detection accuracy.

Our proposedmethod

Carrillosmethod

97.6% 95.7%ves (TP) 205 201ives (FP) 23 39

coefcsegmed as segme

As resultsCarrilldorff dcompa

Fig.tation segmenucleuin termresult head.

Fig.(our pcoefcwell ameasuposed coefcdure wwe apculateacrosoand nusigniOur av this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

ariation of parameters for Carrillos method in ourts.

rm detection

ance of our detection results was determined by ROC curve takes into account the area under the) as a quality measure. The higher the AUC, theuality of a method. The ROC curve determines thes of false positives when a high correct detection

Fig. 5). To create each point of the ROC curve, wethe percentage of correct detection (according ourard) and the number of false positives for a given

the parameter values.roach achieves an AUC value of 0.88 while Carrillosomplishes 0.81. In addition, our proposal achievesetection rate over 97% at the expense of havingse positives. To achieve a comparable result withetection rate over 97%, Carrillos method achievesitives. Carrillos method reports a correct rate over. According to our experiments, this rate is achievedse positives. In Table 4 the relationship betweenves and correct detection is shown for both evalu-ds.

rm head segmentation

ed different experiments to assess segmentationhead, acrosome and nucleus. In each case, we

hand-segm0.83 and 0.respectivel

Fig. 8 shdensity fuing to acrtogether. Ting to our achieved bbution of v(with = 0.vides smal = 0.08).

Additiontance correof Carrillosis to assesshead partsmeasuringthe same pcomponentance valueachieved bdistance (inmask and less than 2with avera0.24 for hetively.ework for sperm head segmentation, Comput. Methods

masks and the Hausdorff distance (consideringdian distance) to assess the disagreement ofon against hand-segmented mask.tioned before, we compare our segmentation

th the results obtained by implementation ofethod. We also calculate Dice coefcient and Haus-nce, using the same testing images in order toe results.shows an image gallery with some segmen-lts, considering head, acrosome and nucleuson. For each segmentation (head, acrosome ande present our best, average and worst result,

f Dice coefcient. For each image, we show thepplying Carrillos method to the same sperm

ows Dice coefcients for both segmentation resultssed method and Carrillos method). The Diceassesses quality of sperm head segmentation asosome and nucleus segmentation, by means of

the overlap with ground-truth. We applied our pro-hod to the testing images and calculated Dices

for each segmented sperm head. The same proce-llowed in the case of acrosome and nucleus. Then,

Carrillos method to the same data set and cal- Dice coefcient for each segmented sperm head,nd nucleus. For every component (head, acrosomes), the Dice coefcients of our proposed method arey better than those achieved by Carrillos method.

results have more than 80% of overlapping againstented masks, with average Dice coefcients of 0.88,82 for head, acrosome and nucleus segmentation,y.ows a graphical representation of the probabilitynction (PDF) for Dice coefcients, correspond-osome, nucleus and sperm head segmentation,his is a comparison between the PDF correspond-results and the one corresponding to the resultsy Carrillos method. As we can observe, the distri-alues for Dice coefcient achieved by our method06) is shifted to higher Dice coefcients and pro-ler variance than that of Carrillos method (with

ally, we present a comparison of Hausdorff dis-sponding to our proposed framework and results

method, in Fig. 9. As in the previous case, our aim quality of segmentation of sperm heads and sperm

(acrosome and nucleus), but now by means of the disagreement with ground-truth. We followedrocedure as in Dice coefcients case. For every

t (head, acrosome and nucleus), the Hausdorff dis-s of our proposed framework are better than thosey Carrillos method, because ours show a smaller

average) between perimeters of hand-segmentedsegmentation results. Our average results have5% of disagreement with hand-segmented masks,ge Hausdorff distance values of 0.15, 0.20 andad, acrosome and nucleus segmentation, respec-



Fig. 6 Resfor averagecolumn), ogold-standproposed/Cweb versio

5. Di

In this papsegmentatito Carrillosposals so fa this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

ults of head, acrosome and nucleus segmentation. Upper row sh results, and last row for worst results. For each part (head, acrour result (second column) and Carrillos method result (third coluard, red presents our proposed/Carrillos method and yellow thearrillos method. (For interpretation of the references to color in n of the article.)

scussion

er we have presented a framework for sperm cellon achieving signicant improvement with respect

method. Our approach is different from the pro-r known (Section 2) in three different aspects:

1 Use of chave signtation. Cincludinpurposesfor imagthe reseaework for sperm head segmentation, Comput. Methods

ows representatives for best results, middle rowsome and nucleus), we present the original (rstmn). The blue color represents the

overlap between gold-standard and ourthis gure legend, the reader is referred to the

olor space combinations. Choices of color spaceicant inuences on the result of image segmen-heng et al. [37] compared several color spacesg RGB and L*a*b* for color image segmentation, and they stated that the selection of a color spacee processing is image/application dependent. Allrch works cited in Section 2 use RGB color space



Fig. 7 Dice coefcient for head, acrosome and nucleus. Oneach box, the edges are the 25th and 75th percentiles andthe whiskers extend to the most extreme data points thatare not outliers. For each box, we show the median value(horizontal line) and the sample mean (). Statisticallysignicant differences between our proposal (grey) andCarrillos method (white) using Wilcoxon rank sum test areindicated (*p < 0.05).

to segmeever, RGBbecause ponents similaritbecause a uniformIn this prcombininfor detec

Fig. 8 ProAcrosome,are showedachieved b = 0.85 0(dotted line

Fig. 9 HaOn each boand the whthat are novalue (horisignicantCarrillos mindicated (

r feae repient use eene be

L*a* mornt sperm cells, including Carrillos method. How- is not suitable for color segmentation and analysisof the high correlation among the R, G, B com-[38,39]. Besides, it is impossible to evaluate they of two colors from their distance in RGB spaceRGB space does not represent color differences in

scale.oposal, we choose to work with a hybrid color spaceg RGB, YCbCr and L*a*b* color spaces. Therefore,

colospacefcbecabetwtancthattion this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

tion stage (Section 3.2.1) we used six redundant

bability density function for Dice coefcient. nucleus and sperm head segmentation results

together, considering the Dice coefcientsy our proposed method (continuous line,.0026, s2 = 0.0036) and by Carrillos method, = 0.79 0.0035, s2 = 0.0065).

been shoothers inalong wievaluatiotion, forcombineponent othe coloexpressethe intenindependbeen ext

2 Use of csegmentstrongly tion metimage sefor colorand colothe viewimage isspace pocolor spaboth stagsperm heusdorff distance for head, acrosome and nucleus.x, the edges are the 25th and 75th percentilesiskers extend to the most extreme data pointst outliers. For each box, we show the medianzontal line) and the sample mean (). Statistically

differences between our proposal (grey) andethod (white) using Wilcoxon rank sum test are

*p < 0.05).

tures in RGB and L*a*b* color spaces. L*a*b* colorresents perceptual uniformity, and it is especiallyin the measurement of small color difference,this can be calculated as the Euclidian distance

two color points [37], and particularly as E dis-tween two color points [40]. For us, it was importantb* space can control color and intensity informa-e independently and simply than RGB. Also, it hasework for sperm head segmentation, Comput. Methods

wn that L*a*b* color space gives better results than color segmentation [41]. We decided to keep RGBth L*a*b* color space after intensive experimentaln, and regarding related work techniques. In addi-

segmentation stage (Section 3.2.2) we choose to L*a*b* and YCbCr color spaces. A chromatic com-f YCbCr was introduced due to two reasons: (a)

r difference of human perception can be directlyd by Euclidean distance in the color space, and (b)sity and chromatic components can be easily andently controlled. In general, YCbCr color space has

ensively used for skin color segmentation [42,43].lustering method. All the existing color imageation approaches are ad hoc, because they areapplication-dependent. Among several segmenta-hods, clustering has been widely used for colorgmentation [4447]. This is due to the fact that

images, a color space is a natural feature space,rs tend to form clusters in the color space. Frompoint of color clustering, it is desired that the

represented by color features which constitute assessing uniform characteristics such as the L*a*b*ce [44]. We choose to apply k-means clustering ines of our proposal: detection and segmentation ofads. As a traditional clustering algorithm, k-means



Fig. 10 N becaSperm hea noth

is populait can beThis, comus an exction app

3 Identicatributiondeterminnot beenreviewedfront direhead anplasm ardetectionthat authcal analyused it toerates eipossible our knowing Carrfront dircould hemorphol

Our resuaspects deby Carrilloachieves a and less diCarrillos min Fig. 6. Wbecause of (Algorithmstage resulHowever, aareas, residseparation between coground, incnote that their extenaddition, w

ting ece ailed tbstanuch that ntatind wd (Figntribe st

Dice obtas exion r206 so et a. Thng oith irec

ts be Thened by). Al

ompnnecppedon-detected sperm cells. (ac) Sperm heads are not detectedd is not detected because head overlaps incomplete tail of a

r for its simplicity for implementation, and also, adopted to solve illumination variation problem.bined with YCbCr and L*a*b* color spaces, provideeptional tool for illumination invariant segmenta-

roach, outperforming the state-of-the-art.tion of sperm head direction. One of the key con-

of our work is the proposal of a novel algorithm toe which direction the sperm head points. This has

considered before in any of the research works in Section 2. Properly identication of the headction could serve to accurately segment the spermd to discard mid-piece regions or residual cyto-eas that may have been included in the result of the. In related works [22,48,4,20], it has been observedors proposed ellipse tting for head morphologi-sis and experimental evaluation, but they have not

rene the segmentation itself. Our proposal gen-ght types of different growing masks, regarding allpositions in which a sperm head may appear. Toledge, none of the proposed approaches (includ-

illos method) have taken into account the headection, however, it is a very important issue thatlp in many other stages in the quest for an accurateogical analysis.

lts have shown that our approach, based on thosescribed above, outperforms the results achieveds method. In fact, we showed that our method

separamid-piand cois a sucards snoted segmecells amethothis cothan thof our cient

It idetectup to Carrill(Fig. 5)affecticells whead darea head).removtion 4.2an incare co(overla this article in press as: V. Chang, et al., Gold-standard and improved framiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018

higher Dice coefcient, lower Hausdorff distancespersion with respect to the results achieved byethod. This is clearly shown in gallery presentede believe that this outperforming occurs mainly

the results of the rst stage of our proposed method 1), combining RGB and L*a*b* color spaces. Thist is extremely accurate to segment sperm heads.s part of the heads, we also segmented mid-pieceual cytoplasm areas, and even tails. This pixelis probably due to the higher Euclidean distancelor pixels corresponding to the cells and back-orporating L*a*b* color space. It is important to

non-coiled tails are removed (though not all ofsion) by the procedure empheraseTails (Fig. 2). Inhen including a YCbCr chromatic component for

tail.There is

ber of falsewith detectrillo et al. p(Fig. 5). It istives numbvalidate pixtance to noof a sperm spaces (Algcolor, as Caremoves obapplying k-regards sizuse of excessive residual cytoplasm area. (d)er sperm cell.

sperm nucleus, our proposed approach removesnd residual cytoplasm areas, as well as rests of tailsails (Fig. 6, rst row of nucleus segmentation). Thistial difference against Carrillos method that dis-

cells because overcome size threshold. It should beindividualize each sperm head for a more accurateon, makes it possible to work with spatially closee obtain signicantly better results than Carrillos. 6, rst row of head segmentation). We believe thatutes to have an average Dice coefcient greater

ate-of-the-art method (Fig. 8), observing that mostcoefcients are greater than the average Dice coef-ined by the Carrillos method.pected that no detection method could achieveate of 100%. Our method is able to correctly detectperm cells (98%), while the method proposed byl. is able to correctly detect 208 sperm cells (99%)

ere are specic situations in test images that areur detection rate (Fig. 10). We observe few spermexcessive residual cytoplasm area, whose fronttion is erroneously detected (because cytoplasmtter with an ideal ellipse than frontal region of its, in size validation of candidate head, this head is

having a larger area than maxTs2 threshold (Sec-so, we observed a particular sperm cell overlappinglete tail of another sperm cell. Thus, as they bothted, the eraseTails procedure would remove both

head and incomplete tail) as if it were only oneework for sperm head segmentation, Comput. Methods

a tradeoff between correct detection rate and num- positives. Our method provides 25 false positivesion rate of 98%, while the method proposed by Car-rovides 49 false positives with detection rate of 99%

a very signicant difference, in terms of false posi-er that we believe is due to three factors. First, weels of the object/background using Euclidean dis-t include objects with a darker intensity than thathead, and combining RGB, L*a*b*, and YCbCr colororithms 1 and 2), better than using only RGB spacerrillos method does. Second, our proposed methodjects according to their size and roundness, aftermeans (Algorithm 2), while Carrillos method onlye validation. Therefore, Carrillos method is more



likely to report strangely shaped objects but with a size sim-ilar to a sperm head. Third, our method removes incompletesperm cellsCarrillos mheads of inIt is imporby our appnot drawn complete t

6. Co

We have prtion and se(including combinatioof a clustermination imethod prtify the heaincrease th

Our expframeworkDice coefcwith respecOur resultswith fewer nucleus seghand-segm

To tacklfor evaluatintroducedtion, built eld. This gpare our reused to comimprovemecontributio

In our onactive contwork, we win order to

Acknowl

The authoof gold-stawith Fig. 1FONDECYTcal Neurosca selected lence Initiaby CONICY24100118) aJose M. S78120425). and U-Rede

r e f e r e n c e s

n: Mo2) (20

Domiedmpres581orld aminealth

Nasentig. Co

Katzwis, sessm321

Krugenkvadingchnic

Freuorph180

Barretzeter-laorphaininalys

Auge Matrrnagira, P

humsessmum. R

Sole. Sanerm-erm

Cipaermiteriaoche

Walclipiakalysborat

Rive Juartimaalys964Augeodels010) 3

Schmermalua427Jagoeermathol.Moruanti this article in press as: V. Chang, et al., Gold-standard and improved

which touch the border of the image (Algorithm 1).ethod does not take this into account, and reportscomplete sperm cells as correct detected heads.tant to note that all the false positives reportedroach are actually sperm heads, however they arein the gold-standard because they do not present aail in the image.

nclusions

esented a two-stage improved framework for detec-gmentation of human sperm head characteristicsacrosome and nucleus). The usage of color spacens (RGB, L*a*b* and YCbCr), together with the usageing method, provides us an exceptional tool for illu-nvariant segmentation approach. In addition, ouroposed an ellipse tting based algorithm to iden-d front direction. This is a very relevant issue toe accuracy of the segmentation.erimental evaluation shows that our proposed

outperforms the state-of-the-art, with a higherient, lower Hausdorff distance and less dispersiont to the results achieved by Carrillos method [24].

achieve notable improvement in the detection ratefalse positives and an accurate head, acrosome andmentation achieving over 80% overlapping againstented mask.e the problem of lacking a public gold-standarding sperm head segmentation methods, we have

a gold-standard for head sperm parts segmenta-with the cooperation of a referent expert in theold-standard has been used to evaluate and com-sults with the state-of-the-art method, and can bepare not only known techniques but also future

nts to present approaches. This is a very signicantn to the scientic community.going work, we are analyzing the impact of adding

ours as a third segmentation stage. For the futureill focus on both mid-piece and tail segmentationprovide a complete sperm parts segmentation tool.

edgements

rs thank E. Labbe for support with the drawingsndard. The authors thank A. Garcia for support. Research in SCIAN-Lab (S. Hrtel) is funded by

(1120579), FONDEF (D11I1096), and the Biomedi-ience Institute (BNI, ICM P09-015-F). SCIAN-Lab ismember of the German-Chilean Center of Excel-tive (DAAD). Violeta Chang was partially fundedT (NAC-DoctoradoLatin 57090057/NAC-ApoyoTesisnd she is partially funded by FONDECYT (1020579).aavedra is partially funded by CONICYT (PAI-Victor Castaneda is funded by FONDECYT (3140444)s BioMedHPC.

[1] , i(1

[2] A.Frde11

[3] WExH

[4] V.idEn

[5] D.Leas20

[6] T.Mrete

[7] C.m97

[8] G.Coinmstan

[9] J. S.LoSpinasH

[10] C.Mspsp

[11] A.SpcrBi

[12] R.Fianla

[13] A.A.Esan63

[14] J. m(2

[15] A.Spev27

[16] J. spPa

[17] J. Quiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018


ther or nothing: the agony of infertility, WHO Bull. 8810) 881882.ar, A. Broome, P. Zuttermeister, M. Seibel, R.an, in: The prevalence and predictability ofsion in infertile women, Fertil. Steril. 58 (6) (1992)163.Health Organization, WHO Laboratory Manual for theation and Processing of Human Semen, World

Organization, Geneva, 2010.i, M. Moradi, M. Nasr-Esfahani, in: Spermcation using elliptic model and tail detection, Trans.mput. Technol. 7 (1) (2005) 419421., J. Overstreet, S. Samuels, P. Niswander, T. Bloom, E.in: Morphometric analysis of spermatozoa in the

ent of human male fertility, J. Androl. 7 (4) (1986)0.er, T. du Toit, D. Franken, A. Acosta, S. Oehninger, R.eld, C. Lombard, in: A new computerized method of

sperm morphology (strict criteria) is as efcient asian reading, Fertil. Steril. 59 (1) (1993) 202209.nd, in: Standards for the rating of human spermology. A cooperative study, Int. J. Fertil. 11 (1) (1966).oso, R. Mercan, K. Ozgur, M. Morshedi, P. Kolm, K.e, T. Kruger, S. Oehninger, in: Intra- andboratory variability in the assessment of spermology by strict criteria: impact of semen preparationg techniques and manual versus computerizedis, Hum. Reprod. 14 (8) (1999) 20362040.r, F. Eustache, B. Ducot, T. Blandin, M. Daudin, I. Diaz,ibi, B. Gony, L. Keskes, M. Kolbezen, A. Lamarte, J.e, N. Nomal, G. Pitaval, O. Simon, I. Virant-Klun, A.. Jouannet, in: Intra- and inter-individual variabilityan sperm concentration, motility and vitalityent during a workshop involving ten laboratories,

eprod. 15 (11) (2000) 23602368.r, J. de Monserrat, R. Gutirrez, J. Nu nez, M. Nu nez,cho, F. Prez-Snchez, T. Cooper, in: Use of theclass analyser for objective assessment of humanmorphology, Int. J. Androl. 26 (5) (2003) 262270.k1, P. Stanic, K. Duric, T. Serdar, E. Suchanek, in:

morphology assessment according to who and strict: method comparison and intra-laboratory variability,m. Med. 19 (1) (2009) 8794.zak-Jedrzejowska, K. Marchlewska, E. Oszukowska, E., L. Bergier, J. Slowikowska-Hilczer, in: Semenis standardization: is there any problem in polishories?, Asian J. Androl. 15 (2013) 616621.ra-Montes, A. Rivera-Gallegos, E. Rodriguez-Villasana,ez-Bengoa, M. Diaz-Perez, M. Hernandez-Valencia, in:te of the variability in the evaluation of semenis, Ginecol. Obstetr. Mexico 81 (11) (2013)4.r, in: Assessing human sperm morphology: top

underdogs or biometrics?, Asian J. Androl. 12 (1)646.assmann, G. Mikuz, B. Bartsch, H. Rohr, in:

iometrics: objective and reproducible methods forting sperm morphology, Eur. Urol. 8 (5) (1982)9., N. Washbrook, E. Hudson, in: Morphometry oftozoa using semiautomatic image analysis, J. Clin.

39 (12) (1986) 13471352.zzi, A. Wyrobek, B. Mayall, B. Gledhill, in:cation and classication of human sperm



morphology by computer-assisted image analysis, Fertil.Steril. 50 (1) (1988) 142152.

[18] F. Prez-Snchez, J. de Monserrat, C. Soler, in: Morphometricanalysis of human sperm morphology, Int. J. Androl. 17 (5)(1994) 248255.

[19] V. Gonzlez-Castro, E. Alegre, P. Morala-Argello, S. Surez,in: A combined and intelligent new segmentation methodfor boar semen based on thresholding and watershedtransform, Int. J. Imaging 2 (S09) (2009) 7080.

[20] L. Snchez, N. Petkov, Estimation of boar sperm status usingintracellular density distribution in grey level images, in:Similarity-Based Clustering, Springer-Verlag, 2009, pp.16918

[21] E. AlegGarca-of the aCompu

[22] K. Parkstrateg29430

[23] L. SncintraceLectureCluster16918

[24] H. CarrGutirrhumanInternaBioeng

[25] N. Otsuhistogr6266.

[26] A. BijardiscrimsemenSympo(ISPA 2

[27] A. BijarFully aparts insmear,

[28] V. Abbimorphteratoz

[29] R. Daviin: Accautoma(4) (199

[30] C. Wanin: CommorphSteril. 5

[31] R. Davipreparaautoma59 (2) (1

[32] F. LacqSarmie

and staining procedures for reliable results usingcomputerized morphology, Arch. Androl. 36 (2) (1996)133138.

[33] K. Coetzee, T. Kruger, C. Lombard, in: Repeatability andvariance analysis on multiple computer-assisted (IVOS)sperm morphology readings, Andrologia 31 (3) (1999)163168.

[34] J. Yao, N. Kharma, P. Grogono, in: A multi-population geneticalgorithm for robust and fast ellipse detection, Pattern Anal.Appl. 8 (12) (2005) 149162.

[35] L. Dice, in: Measures of the amount of ecologic associationbetween species, Ecology 26 (3) (1945) 297302.

Alt, terpompu

Chegmen2) (20. Pietlor dstribunfer

Littmmpauralmmdustr-I. Ohgion 980) 2-L. Hslor im002) 6

Garclor reans. . Celegmen990) 1

Hall,zdekusterages268

Juriotervausterternad 6thnfer

Isa, meanans. 452. Yi, Klipticaveleternaedici this article in press as: V. Chang, et al., Gold-standard and improved

4.re, V. Gonzlez-Castro, R. Alaiz-Rodrguez, M.Ords, in: Texture and moments-based classicationcrosome integrity of boar spermatozoa images,t. Methods Progr. Biomed. 108 (2) (2012) 873881., W. Yi, J. Paick, in: Segmentation of sperms using theic Hough transform, Ann. Biomed. Eng. 25 (1997)2.hez, N. Petkov, Estimation of boar sperm status usingllular density distribution in grey level images, in:

Notes in Computer Science, Similarity-Baseding, Springer-Verlag, 2009, pp.4.illo, J. Villarreal, M. Sotaquira, M. Goelkel, R.ez, A computer aided tool for the assessment of

sperm morphology, in: Proceedings of the 7th IEEEtional Conference on Bioinformatics andineering (BIBE 2007), 2007, pp. 11521157., in: A threshold selection method from gray-levelams, IEEE Trans. Syst. Man Cybern. 9 (1) (1979)

, M. Mikaeili, Sperms tail identication andination in microscopic images of stained human

smear, in: Proceedings of the 7th Internationalsium on Image and Signal Processing and Analysis011), 2011, pp. 709714., A. Pe nalver Benavent, M. Mikaeili, R. Khayati, in:utomatic identication and discrimination of sperms

microscopic images of stained human semen J. Biomed. Sci. Eng. 5 (7) (2012) 384395.ramy, V. Shanthi, in: Spermatozoa segmentation andological parameter analysis based detection ofoospermia, Int. J. Comput. Appl. 3 (7) (2010) 1923.s, D. Bain, R. Siemers, D. Thal, J. Andrew, C. Gravance,uracy and precision of the CellForm-Humanted sperm morphometry instrument, Fertil. Steril. 582) 763769.g, A. Leung, W. Tsoi, J. Leung, V. Ng, K. Lee, S. Chan,puter-assisted assessment of human sperm

ology: comparison with visual assessment, Fertil.5 (5) (1991) 983988.s, C. Gravance, in: Standardization of specimention, staining, and sampling methods improvested sperm-head morphometry analysis, Fertil. Steril.993) 412417.

uet, T. Kruger, T. du Toit, C. Lombard, C. Sancheznto, A. de Villiers, K. Coetzee, in: Slide preparation

[36] H.inCo

[37] H.se(1

[38] McodiCo

[39] E.coNe

[40] CoIn

[41] Y.re(1

[42] R.co(2

[43] C.coTr

[44] Mse(1

[45] L.Beclim67

[46] A.InclInanCo

[47] N.k-Tr21

[48] WelwInMiomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.06.018


L.J. Guibas, Discrete geometric shapes: matching,lation, and approximation, in: Handbook oftational Geometry, Elsevier, 1999, pp. 121153.ng, X. Jiang, Y. Sun, J. Wang, in: Color imagetation: advances and prospects, Pattern Recogn. 3401) 22592281.ikainen, S. Nieminen, E. Marszalec, T. Ojala, Accurateiscrimination with classication based on featuretions, in: Proceedings of the 13th International

ence on Pattern Recognition, vol. 3, 1996, pp. 833838.ann, H. Ritter, in: Adaptive color segmentationa

rison of neural and statistical methods, IEEE Trans. Netw. 8 (1) (1997) 175185.ission International de lclairage CIE, Improvement toial Color Difference Evaluation. Tech. Rep., 2001.ta, T. Kanade, T. Sakai, in: Color information forsegmentation, Comput. Graph. Image Process. 13 (3)22241.u, M. Abdel-Mottaleb, A. Jain, in: Face detection inages, IEEE Trans. Pattern Anal. Mach. Intell. 24 (5)96706.ia, G. Tziritas, in: Face detection using quantized skingions merging and wavelet packet analysis, IEEEMultimedia 1 (3) (1999) 264277.nk, in: A color clustering technique for imagetation, Comput. Vis. Graph. Image Process. 52 (2)45170.

A. Bensaid, L. Clarke, R. Velthuizen, M. Silbiger, J., in: A comparison of neural network and fuzzying techniques in segmenting magnetic resonance

of the brain, IEEE Trans. Neural Netw. 3 (5) (1992)2., M.D.P. Pagola, C. Lopez-Molina, P. Melo-Pinto,l-valued restricted equivalence functions applied oning techniques, in: Proceedings of the 13thtional Fuzzy Systems Association World Congress

European Society for Fuzzy Logic and Technologyence (IFSA-EUSFLAT), 2009, pp. 831836.S. Salamah, U. Ngah, in: Adaptive fuzzy movings clustering algorithm for image segmentation, IEEE

Consumer Electron. 55 (4) (2009)153.. Park, J. Paick, Parameterized characterization of

sperm heads using Fourier representation andt transform, in: Proceedings of the 20th Annualtional Conference of the IEEE Engineering inne and Biology Society, vol. 2, 1998, pp. 974977.

Gold-standard and improved framework for sperm head segmentation1 Introduction2 Related work3 Material and methods3.1 Gold-standard3.1.1 Sample preparation3.1.2 Image acquisition

3.2 Algorithm3.2.1 Detection of sperm heads3.2.2 Segmentation of sperm head3.2.3 Segmentation of sperm nucleus and acrosome

3.3 Evaluation metrics3.3.1 Evaluation metrics for detection3.3.2 Evaluation metrics for segmentation

4 Results4.1 Testing data set4.2 Parameter optimization4.3 Sperm detection4.4 Sperm head segmentation

5 Discussion6 ConclusionsAcknowledgementsReferences

Documents

1-s2.0-S0169260714002521-main