SS

Ga

b

c

a

ARRA

K2AOTB

1

hmaaLrmag

pobpddptspl

0d

Micron 42 (2011) 911–920

Contents lists available at ScienceDirect

Micron

journa l homepage: www.e lsev ier .com/ locate /micron

emi-automated Acanthamoeba polyphaga detection and computation ofalmonella typhimurium concentration in spatio-temporal images

eorge D. Tsibidisa,∗, Nigel J. Burroughsb, William Gazec, Elizabeth M.H. Wellingtonc

Institute of Electronic Structure and Laser, Foundation for Research and Technology, P.O. Box 1527, Vassilika Vouton, 71110 Heraklion, GreeceSystems Biology Centre, University of Warwick, Coventry, CV4 7AL, UKBiological Sciences, University of Warwick, Coventry, CV4 7AL, UK

r t i c l e i n f o

rticle history:eceived 22 September 2010eceived in revised form 17 June 2011ccepted 25 June 2011

a b s t r a c t

Interaction between bacteria and protozoa is an increasing area of interest, however there are a fewsystems that allow extensive observation of the interactions. A semi-automated approach is proposed toanalyse a large amount of experimental data and avoid a time demanding manual object classification. Weexamined a surface system consisting of non nutrient agar with a uniform bacterial lawn that extended

eywords:D Imagingutomated image analysisrganism detectionime-lapse microscopy

over the agar surface, and a spatially localised central population of amoebae. Location and identificationof protozoa and quantification of bacteria population are performed by the employment of image analysistechniques in a series of spatial images. The quantitative tools are based on intensity thresholding, oron probabilistic models. To accelerate organism identification, correct classification errors and attainquantitative details of all objects a custom written Graphical User Interfaces has also been developed.

acteria counting

. Introduction

The importance of interactions between bacteria and protozoaas gained significance since the discovery that Legionella pneu-ophila could replicate within Acanthamoeba polyphaga (Edelstein

nd Meyer, 1984; Stevens and O’Dell, 1973). L. pneumophila isn organism which is strongly linked to the development of theegionnaires’ disease in humans (Edelstein and Meyer, 1984). Theecognition of a connection between A. polyphaga and L. pneu-ophila (Stevens and O’Dell, 1973) has attracted much attention

nd possible association to other infections has also been investi-ated (Huws et al., 2008).

A powerful approach to the elucidation of therotozoan–bacteria interactions is to analyse the movementf protozoa in a bacterial lawn. Principally, the movement andehaviour is pertinent to the environment through two mainrocesses, encystment under poor resources and bacteria depen-ent motility. A statistical analysis of the movement allowsetermination of the factors that characterise the underlyingrocesses but usually a large number of organisms are necessaryo infer useful information and draw a convincing picture. Video

equences that contain such organisms should be analysed torobe the protozoa behaviour (Gaze et al., 2003), however, the

arge number of the organisms complicates the process towards

∗ Corresponding author.E-mail address: tsibidis@iesl.forth.gr (G.D. Tsibidis).

968-4328/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.oi:10.1016/j.micron.2011.06.010

© 2011 Elsevier Ltd. All rights reserved.

a quantitative investigation. A variety of sophisticated methodsbased on neural networks (Casasent and Smokelin, 1994), wavelettransforms (Casasent et al., 1992) and genetic algorithms (Kim andJung, 2004), have also been widely employed and those state of theart techniques are designed to detect and analyse more compleximages. The implementation of algorithms which are based on theabove techniques are usually computationally more demanding.Moreover, subdivision of the original images into its constituentparts requires the application of filters (Gerlich et al., 2003), theemployment of confinement trees and connected operators (Soille,1999) or edge-based segmentation (Tvarusko et al., 1998, 1999). Bycontrast, image analysis techniques based on image thresholdingoffer an adequate object detection method in pictures or videosequences characterised by a small number of objects with welldefined boundaries. Several computer-assisted systems havebeen developed to track objects of biological importance in videosequences based on intensity thresholding (Degerman et al., 2009;Rabut and Ellenberg, 2004; Sbalzarini and Koumoutsakos, 2005;Tsibidis and Tavernarakis, 2007). With respect to the detection ofprotozoans in lawns of bacteria, computational tools are requiredto be capable to facilitate object extraction, provide efficient clas-sification (i.e., cysts or amoebae), determine quantitative detailsabout the organism motility, count the bacterial concentration andoffer a correlation of the bacterial distribution to the presence of

organisms in the neighbourhood. Although most quantitative toolsmanage to extract organisms adequately and they are capable toidentify protozoa in a series of images, they do not provide a satis-factory answer to how bacteria concentrations could be counted.

9 icron

Miictcifasbe1mTtt

aeitbccidIwepismblmea

2

2

e1bc

2

(c1ataiaw(ei

12 G.D. Tsibidis et al. / M

oreover, existing algorithms fail to characterise organism motil-ty in two, temporally separated, images by a small time differencef the object boundaries are nearly immobile. Furthermore, spe-ialised image processing methods are not generally applicable tohe bacterial counting due to the difficulty to detect bacteria. Mostommonly, the intensity of pixels in an image occupied by bacterias very close to that associated with background which complicatesurther the bacteria identification. As a result, bacteria enumer-tion would be a rather laborious process that requires a highlykilled expert. Previous works considered fluorescently taggedacteria, a method that simplified the identification (Drozdovt al., 2006; Pernthaler et al., 2003; Singleton et al., 2001; Soll et al.,988). Quantitative tools that address the above issues entail theodification of the existing algorithms with additional routines.

he improved algorithms need firstly to manage the detection ofhe organisms and secondly allow an accurate classification of thewo types of organisms in an efficient and robust way.

In the present paper, to overcome existing problems, a semi-utomated system has been developed aimed to assist thextraction of quantitative information of A. polyphaga and cystsn a lawn of Salmonella typhimurium bacteria. Image processingechniques are employed to identify objects in a series of imagesased on thresholding. An automated classification of amoeba andysts is firstly performed by considering a circularity criterion (i.e.,ysts are almost circular) and image segmentation is based onntensity thresholding. A semi-automated correction of the objectistinction is subsequently performed by means of a Graphical User

nterface (GUI) designed in Matlab (The Math Works, Natick, MA)hich provides a synergy of clustering and viewing tools to accel-

rate organism identification. The method aims to integrate imagerocessing and statistical classification technologies. A synergy of

mage processing techniques and probabilistic models for noise areubsequently used to assist in the bacterial counting. The proposedethod is used as an alternative to the employment of fluorescent

acteria as a technique to facilitate bacterial detection. A simi-ar approach is ensued towards the characterisation the organism

otility. The algorithms developed in this paper intend to be anssential component of a completely automated system that willllow an adequate quantification of protozoa–bacteria interactions.

. Materials and methods

.1. Strains

The virulent strain of S. typhimurium was used in co-culturexperiments, and was grown in Luria Broth (LB). A. polyphaga CCAP501/18 was used as a model organism since L. pneumophila haseen shown to grow in this strain. A. polyphaga was grown axeni-ally in proteose peptone glucose medium (PPG).

.2. Co-culture experiments

Bacterial cultures were washed 3 times in Page’s Amoeba SalinePAS) and diluted to approximately 108 cells/ml (high bacteriaoncentration) and 107 cells/ml (medium bacteria concentration).00 ml were spread onto non-nutrient agar (NNA) plates. Differentpproaches were tested to ensure uniform distribution of bac-eria over the plates. It was found that by diluting the bacterialiquot with 500 ml of distilled water before spreading and allow-ng the liquid to dry on the plate a uniform distribution could bechieved. One week old cultures of A. polyphaga grown at 25 ◦C were

ashed 3 times in PAS, and 25 ml of approximately 6 × 104 cells/ml

1500 cells) was spotted in the centre of agar plates. Drop diam-ter and centre were recorded for each plate. Co-cultures werencubated at 35 ◦C.

42 (2011) 911–920

2.3. Imaging

Destructive sampling was carried out daily, with plates exam-ined using a Zeiss Axioskop 2 microscope and 10× objective.Individual plates were used for measurements on days 0, 1–3, 5and 7. Later time points were not required in bacteria free systemsas they rapidly encyst. At each time point a single plate was used foreach bacterial density (high, medium, none). A rectangular lattice ofimages was generated at a periodic spacing of 1 mm covering partof the fourth quadrant measured from the centre of the amoebasolution drop (centre corresponds to lower right corner of firstimage). Lattice dimensions were adjusted depending on amoebamigration. Images were 1022 pixels × 1024 pixels (8bit), capturedusing a Hamamatsu black and white digital camera at magnifica-tion 10× and Improvision Openlab software. Each image coveredapproximately 0.46 mm2. No coverslip was used to preserve spatialintegrity of the bacteria and amoebae populations.

2.4. Image segmentation

The refractive halo is used to locate amoeba/cysts by intensitythresholding; in principle, a threshold value equal to 0.85 (by con-verting gray scale image intensities in the range [0, 1]) was chosenbased on the fact that the great majority of the objects could beextracted adequately. A satisfactory object extraction entails onacquisition of the halo pixels of the organism without manual inter-vention. The boundaries of the organisms is attainable by means ofa varied threshold, however, it is evident that the selected value(∼0.85) provides an accurate segmentation for images that arecharacterised by a low and medium bacteria density. A test withrelatively higher (∼0.90) or lower values (∼0.80) results into anoverestimation or underestimation of the perimeter contour line,respectively. By contrast, images with high concentration of bacte-ria are harder to analyse and establishment of an optimal thresholdlevel is not feasible; as a result alternative techniques are requiredto facilitate segmentation including edge-based methods (Gerlichet al., 2003), however, in this study focus will be restricted onthe segmentation process in the two former cases. A small num-ber of objects were not detectable by their halo and thereforethey were added manually by means of a Graphical User interfacedesigned in Matlab (Burroughs et al., 2003). Image thresholdingallows transformation of intensity images into black and whiteimages (BWI). Object recognition and separation is based on theapplication of morphological operations on BWI (Sonka et al., 1998)using connected-component labelling for identifying each objectin binary images. Furthermore, dilation adds pixels to the bound-aries of objects in image, while erosion removes pixels on objectboundaries to improve perimeter extraction (Sonka et al., 1998).Matlab built-in functions are employed to facilitate labelling andperform morphological operations. Image segmentation processyields quantitative information associated to the centre of masslocation, perimeter, size, area, and shape statistics. More specifi-cally, the location of centre of the object is estimated by consideringall mask pixels and computing the centre of mass (Fig. 1a).

2.5. Amoeba-cyst classification

Cysts have a nearly spherical shape and the projection of theorganisms on the plane of focus yields an almost circular shape.By contrast, amoebae are usually characterised by an oval shape.Therefore, the ‘circularity index’ (CI) of the object perimeter can be

used as a criterion for the distinction between amoebae and cysts(Soll et al., 1988). It is defined by

CI = ır

r(1)

G.D. Tsibidis et al. / Micron 42 (2011) 911–920 913

F ter ofp ed iml

wdmciaibdto

2

f

ig. 1. (a) Image that contains amoeba and cysts. Blue lines represent the perimeerimeter in case manual correction was performed, (b) intensity image of delay

egend, the reader is referred to the web version of the article.)

here r is the average radius of the object and ır is the standardeviation of the radius values. The average radius of the object iseasured by summing over all distances between the estimated

entre of mass of the organism and the perimeter pixels and divid-ng by the number of boundary pixels. Therefore cysts should bessociated with a small CI. In principle, a circularity index values determined by the user in a way that allows an efficient amoe-ae/cyst distinction. A library of images with the circularity indexisplayed (Fig. 2) and a histogram graph (Fig. 3a) allows an estima-ion of the threshold value of CI. It appears that a circularity indexf 0.035 offers an efficient preliminary amoebae-cyst separation.

.6. Mobile–immobile analysis

The distinction between mobile and immobile organisms wasacilitated by the employment of a delayed image comparison.

Fig. 2. A sample of library of objects in which an object n is ex

each object, red dots indicate their centre of mass and light blue lines stands forage Id = I(t + �t) − I(t). (For interpretation of the references to colour in this figure

More specifically, for two observations of the same region (�t = 4 sapart), we define

Id = I(t + �t) − I(t)⟨I⟩

= 12

[I(t + �t) + I(t)](2)

where I(t) and I(t + �t) are the intensities of images at time t andt + �t, respectively. The delayed image Id is illustrated in Fig. 1b.Due to fluctuations in photon counts, noise can be modelled as aPoisson variate, i.e., the pixels in images I(t), I(t + �t) should be Pois-son distributed with mean 〈I〉 under the noise model. This suggests

that the variable f = Id/√⟨

I⟩

is approximately Gaussian with zero

mean, standard deviation � and it follows a distribution of the form

A√2��

exp

(− f 2

2�2

)(3)

tracted from an image s and the ır/r value is provided.

914 G.D. Tsibidis et al. / Micron 42 (2011) 911–920

F ct: per( age ir f the

wfmpgPis(tnif

p

siidptootlsa

R

ig. 3. (a) Histogram that displays frequency of ır/r values, (b) details for each objeupper right), quantitative information (lower left), intensity image of delayed imeferences to colour in this figure legend, the reader is referred to the web version o

here A is a rescaling parameter. Thus, an intensity histogram offor two consecutive images should be Gaussian when there is noovement. By contrast, if the organism moves significantly, some

ixels inside the mask will be characterized by intensity changesreater than the values that can be described by the Poisson model.arameters � and A of the distribution are estimated from the imagentensity profile by a recursive algorithm that runs over the wholeet of images. The above model yields an excellent fit to real dataFig. 4a) despite discretisation of image intensities requires a carefulreatment when generating histograms. After subtracting out theoise from the observed distribution profile (i.e., h(f)), the probabil-

ty that a pixel with observation f is unrelated to noise is computedrom the expression

(f ) = max

(0, 1 − A

h(f )√

2��exp

(− f 2

2�2

))(4)

Fig. 4b illustrates the probability p(f) as a function of the inten-ity difference and it is remarkable that for pixels with largentensity changes the probability is very close to 1. The probabilitys restricted to be positive; moreover, noise subtraction is not exactue to random statistical fluctuations. Using a confidence level on(f) (i.e., p(f) = 20%), the number of ‘mobile’ pixels is counted insidehe amoeba/cyst perimeter to remove fluctuations. The selectionf the confidence level was based on an analysis of a large numberf images (using viewing facilities of the GUI) which demonstratedhat a threshold value equal to 20% provides an intensity change,arge enough to be associated with ‘mobile’ pixels. By contrast,maller values of p(f) corresponded mostly to ‘immobile’ pixels. We

lso define the ratio R,

= maximum pixel value�

(5)

Fig. 4. (a) Gaussian fitting of variable Id , (b) distribution of probability value

iphery and centre of mass for the organism (upper left), extraction of the organismn a small region that includes the object (lower right). (For interpretation of thearticle.)

where maximum pixel value is the maximum intensity expressed bypixels inside the mask of the object and R characterises the mobil-ity of the organism. On average, mobile objects display greaterintensity changes and contain a higher percentage of ‘mobile’ pix-els. The aforementioned technique is particularly valuable whenthere is a movement in the interior of the organism and maskoverlapping between two different times is significant. In thesecases, previous approaches based on determining possible motil-ity on tracking the new position (Degerman et al., 2009; Rabut andEllenberg, 2004; Sbalzarini and Koumoutsakos, 2005; Tsibidis andTavernarakis, 2007) are not capable to estimate accurately internalmotility due to the fact the centre of mass of the organisms doesnot move.

2.7. Bacterial coverage

Bacteria analysis methodology was similar to the approachwhich was ensued to achieve mobile–immobile object distinction.The method was motivated by the observation that transmissionvariation through agar produced an approximate Gaussian agarintensity profile and darker pixels corresponded to bacteria. Fromthe original intensity image I (Fig. 5a), an image Ir was constructedwhere all objects were removed, i.e., all pixels inside the mask ofthe organism are ignored. Provided bacterial coverage was less than30% of the plate, the intensity profile was dominated by the agarsurface with an approximate Gaussian peak with mode near inten-sity levels 0.4–0.6, which varied from image to image because ofrefocusing and intensity adjustment during data collection. Theagar intensity profile was estimated by modelling it as a Gaussian

distribution and fitting parameters such as the mean �, the max-imum A and the standard deviation � of the distribution from aregression analysis on the derivative of the intensity profile over anappropriate range. In practice, the fit was weighted (biased) to the

s that a pixel with temporal intensity change Id is unrelated to noise.

G.D. Tsibidis et al. / Micron 42 (2011) 911–920 915

F age irt

loittt

B

ig. 5. (a) Original image of bacteria (dark spots) in the presence of protozoa, (b) imhis figure legend, the reader is referred to the web version of the article.)

eft of the mode of the distribution to optimise the fit in the regionf agar–bacteria intensity separation. A probability of a pixel withntensity Ir being bacteria is subsequently computed by subtractinghe Gaussian agar intensity profile from the curve that correspondso the observed data. Thus, results can be attained by computinghe percentage of the area covered by bacteria, BC,

C = B

S − M× 100 (6)

Fig. 6. Graphical U

covered with bacteria (red spots). (For interpretation of the references to colour in

where B stands for the number of pixels which are identified as bac-teria (probability weighted provided p(f) > 20%, threshold removesstochastic fluctuations), and S − M is the number of pixels whichare not covered by objects (S: size of image in pixels, M: number ofpixels occupied by objects).

2.8. Graphical user interface features

A preliminary automated analysis performed for objectclassification and quantification was followed by user interven-

ser Interface.

916 G.D. Tsibidis et al. / Micron 42 (2011) 911–920

Fig. 7. (a) Clustering procedure to simplify amoeba-cyst classification. The manually drawn region denoted by the blue polygonal line indicates objects initially identifiedas immobile cysts. MA, IA, MC and IC stand for mobile amoebae, immobile amoebae, mobile cysts and immobile cysts, respectively, (b) Clustering procedure to simplifymobile–immobile amoebae classification. The manually drawn region denoted by the blue polygonal line indicates objects initially identified as immobile amoebae. (Forinterpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

F l lawnd inal drf ur in t

tilmgvcmaoctw

Fbl

ig. 8. (a) Distribution of amoebae and cysts on day 2 on medium density bacteriaensity bacterial lawn (drop size is 15 mm). The dotted blue line indicates the origact that they contain multiple objects. (For interpretation of the references to colo

ion/reclassification using a Graphical User Interface (GUI) designedn Matlab. The GUI (Fig. 6) was developed to overcome prob-ems related to classification errors, incorrect perimeter estimation,

issing objects due to small intensity contrast with respect to back-round and inadequate separation of multiple objects situated in aery small region. The interface allows the user to intervene andorrect interactively classification and morphological errors. Allanual changes are recorded and overall analysis is adapted tomodification which makes the GUI a dynamic interface. More-

ver, the GUI contains viewing tools to obtain data, measurements,

onfidence statistics and images for individual objects that allowhe acquisition of summary information for individual objects orhole categories of objects. The analysis tools are employed to

ig. 9. (a) Amoeba radius size on day 3, (b) fitting of the peak of the intensity values (racterial curve (green dotted line). Gaussian distribution parameters were calculated (� =

egend, the reader is referred to the web version of the article.)

(drop size is 11 mm), (b) distribution of amoebae and cysts on day 3 on mediumop size. ‘MO’ corresponds to objects that are unclassifiable by observer due to thehis figure legend, the reader is referred to the web version of the article.)

facilitate a fast quantitative and qualitative investigation for objectsby enabling biologists to view plots of results (Figs. 1–3 and 7–9)(Burroughs et al., 2003). The GUI comprises six parts: (i) in the firstsection (inside green box A), quantitative details of all objects forradius, ır/r, perimeter and library of images of objects classifiedaccording to the values of the aforementioned quantities area canbe displayed by viewing facilities. Cyst/amoeba classification andreclassification are achievable by modifying the threshold of CI andmobile/immobile organism distinction is performed by evaluatingthe ration R (Eq. (5)). Number of objects according to their clas-

sification is conducted and enumeration results are displayed, (ii)in the second section (inside magenta box B), a detailed analysisof individual images is performed. All objects in the image are dis-

ed line) with a Gaussian distribution (blue line) to compute noise and derive the0.3931 and � = 0.019). (For interpretation of the references to colour in this figure

icron

pmatihiaibtcabEoafRgcba

3

ilwdpfrottaibceycfdobCratwtasttuuGafso

G.D. Tsibidis et al. / M

layed according to previous classification, perimeter and centre ofass of the organisms and delayed images are illustrated. The prob-

bility p(f) (Eq. (4)) is also shown for the whole image, (iii) in thehird section (inside red box C), a detailed analysis is presented forndividual objects, including radius, perimeter, area, ır/r, whether itas been reclassified manually, position in the lattice and a delayed

mage to allow mobility speculation. A summary of information islso illustrated in figures created by means of the GUI (Fig. 3b), (iv)n the fourth section (inside yellow box D), a redrawing tool haseen incorporated that aims to facilitate manual perimeter correc-ion for organisms with a refractive halo that has not been obtainedorrectly. Furthermore, objects that have been attained during theutomated procedure and they are not of biological importance cane removed from consideration. (v) the fifth section (inside blue box) contains reclassification facilities and any object in the datasetf images can be reclassified and change is incorporated into thenalysis. (vi) the sixth section (inside green box F) contains viewingacilities of statistics of all objects according to their classification.adius, perimeter, area, ır/r and R distribution can be illustratedraphically and histograms can be displayed. The most importantomponent of the section, though, is the clustering tools that wille described in Section 3 which aims to correct the preliminarymeobae/cyst and mobile/immobile classification.

. Results and discussion

Basic image analysis techniques based on intensity threshold-ng were firstly used to segment all set of images and subsequentlyocate and count amoebae and cysts. Image processing algorithms

ere applied to all datasets aimed to identify all objects andetermine a set of morphological details such as position, shape,erimeter and area. Small objects of no importance were removedrom consideration by adapting the algorithm to ignore tiny, lightegions (i.e., comprising fewer than 20 pixels). Based on objectutlines, a quantification model was developed. Results revealedhat a small number of organisms were not detected correctly byhe algorithm or equivalently, the automated system resulted in

considerably overestimated number of objects. Time consum-ng manual classification for every individual object was avoidedy running an automated classification procedure based on theircularity index criterion to distinguish amoebae from cysts. Asxplained in Section 2, the preliminary circularity index of 0.035ields an efficient organism amoebae/cyst classification in mostases. Although small CI appears to provide a satisfactory criterionor cyst identification, special attention is required for small objectsue to the fact that decreasing radius results into bigger fluctuationf ır/r (i.e., bigger CI). Hence, a secondary classification of amoe-ae/cyst classification based on the dependence of object radius onI is required. The employment of the technique aims to improveesults from automated classification in some cases. A clusteringlgorithm which is a component of the GUI is applied to data inhe scatter plot ır/r vs. r (Fig. 7a); due to small values of ır/r, cystsere clustered in the lower section of the graph and objects were

hereby classified based on approximate clustering. ‘Clustering’ isssociated to ‘manual selection’ of cysts that are characterised bymall ır/r at large distances and large CI at small distances. The clus-ering process is performed by drawing a perimeter around a regionhat includes all objects that are assumed to be cysts (Fig. 7a). Thepper borderline of the region is specified by examining the objectssing the viewing facilities of the GUI. The viewing facilities of theUI provide a rapid scan of a library of objects, allow a comparison

nd offer the capability of organism reclassification. Hence, the GUIacilitates the creation of clustering patterns that allow a faster clas-ification and decision making process. Fig. 7a illustrates a handfulf objects that can be identified as amoebae despite they reside

42 (2011) 911–920 917

inside the cluster and have a nearly circular shape. In this case,a manual classification is required because amoebae are immobileand highly spherical when they are in the division phase. As a result,amoebae and cysts are indistinguishable and they are identifiableonly through visual inspection of the image. The advantage of themethod, however, is pertinent to the minimisation of the need fora manual object classification of the entire data set. It is impor-tant to emphasise that the above object classification criteria havebeen applied to objects with a well defined boundary. By contrast,amoebae and cysts with a part outside the image (edge objects)are either excluded from consideration or indentified manually byobservation despite a possible automated extraction and classifi-cation. An example is illustrated in Fig. 1a, where an edge objecthas been attained by the algorithm, however, the perimeter of theorganism is coloured in cyan for distinction from the rest of the pro-tozoans. Similarly, analysis was not performed on a small numberof unidentified fragments (e.g., agar crystals) that were removedfrom consideration.

To distinguish mobile from immobile objects and facilitatecytoplasmic movement, delayed images Id = I(t + �t) − I(t) weregenerated (Fig. 1b) and Id values were fitted with a Gaussian dis-tribution (Fig. 4a). A probabilistic computation of the percentagemobile pixels in every object (Fig. 4b) was subsequently usedto identify mobile objects following the procedure described inSection 2. Furthermore, a clustering process was performed toenhance, if required, the efficiency of mobile/immobile classifi-cation. More specifically, a region in the scatter plot percentageof mobile pixels vs. R is defined to characterize immobile objects(Fig. 7b). Clustering in conjunction with examination of objects inthe bordering areas was performed to define a region. It is evidentthat small percentage of mobile pixels for small R (which indicatesthe maximum intensity change as explained in the previous sec-tion) is expected to be linked to immobile organisms. Thus, theemployment of the GUI’s clustering and viewing tool allowed arapid and efficient determination and characterization of all mov-ing organisms. Separate mobile–immobile classification regionswere explored for amoebae and cysts since cytoplasmic move-ment was large in amoebae, in contrast to movement inside cysts.Therefore mobile-immobile classification for cysts was less reliable,however, the ecosystem dynamics is not influenced due to the factthat the role of cysts is of minor importance.

A powerful approach to the elucidation of theprotozoan–bacteria interactions is associated to a thoroughinvestigation of the organism number dependence on bacteriadistribution. A consistent elucidation entails bacteria counting thatare derived by analysing sets of images based on experiments withorganisms in lawns of bacteria of various concentrations (low,medium, high). Without loss of generality, for the inoculation pro-cedure (Fig. 8), a drop of amoebae on a lawn of bacteria (mediumconcentration) was used and the distribution of amoebae and cystswas investigated during several days. Fig. 8a and b illustrate theposition of amoebae-cysts on days 2 and 3, respectively, in thefirst quadrant. On day 2, there exist 56 amoebae (50 mobile and 6immobile) and 50 cysts (22 mobile and 28 immobile) and most ofthem are situated inside the drop radius (i.e., 11 mm). By contrast,on day 3, the number of amoebae has increased to 160 (156mobile and 4 immobile) while the number of cysts has essentiallyremained unchanged (49). There are also 11 organisms that cannotbe classified to any of the above categories even by the expert’seye. The unclassifiable objects are usually amoebae in the processof division or single organisms made up of two or three amoebae(‘multiple objects’). ‘Multiple objects’ constitute a set in the ‘All

others’ category in Table 1. Other types that are classified in the‘all others’ category include bacterial colonies, carcasses, remnants(i.e., unidentifiable objects), other unclassifiable organisms. It isevident that most of the mobile objects have moved out of the

918 G.D. Tsibidis et al. / Micron 42 (2011) 911–920

Table 1Objects obtained through the synergy of the semi-automated procedure and the clustering algorithms on various days. Comparison of results from computer assisted methodsand manual intervention is also presented by reference to the number of organisms that needed manual reclassification.

Type of objects Data

Day 1 Day 2 Day 3 Day 5 Day 7

Mobile amoebae Semi-automated 50 156 180 152 160Manual 1 8 12 10 9

Immobile amoebae Semi-automated 6 4 22 15 22Manual 0 0 2 4 6

Mobile cysts Semi-automated 22 29 55 47 38Manual 1 1 5 3 4

Immobile cysts Semi-automated 22 20 42 35 39Manual 3 0 4 4 8

All others Semi-Automated 47 60 86 51 61Manual 3 13 15 5 10

dpbed(

pIicfaitIhftpitoT

i

fir�ibdpGwrptemcba1m

Redrawn objects – 13Manual classification vs. semi-automated classification percentage (%) 4.8

rop radius on day 3 (i.e., 15 mm) which is an indication that inresence of bacteria, very few amoebae encysted; by contrast, inacteria-free plates (data not shown), the majority of amoebaencysted within 3–5 days with little migration out of the originalrop. Moreover, amoebae demonstrate a wide variation in sizeFig. 9a), which does not correlate with distance of migration.

Spatial bacterial coverage was also calculated by means of arobabilistic approach described in Section 2. It is assumed thatis a matrix that represents the original image (Fig. 5a) before an

ntermediate modification is introduced by masking out the pixelsovered by all objects in the image. Special attention is requiredor dark pixels inside the organism due to the fact that the associ-ted intensity is close to the value that corresponds to the bacterialntensity. Therefore, to avoid overestimation of bacteria counting,he mask of the objects is removed. A new matrix ir is derived andmin and Imax are the smallest and largest values in the intensityistogram ir and i∈[Imin, Imax]. It is assumed that noise in the image

ollows a Gaussian distribution G (Fig. 9b) and a quantity pi, equalo i − G/i, is assigned to every intensity value i. The physical inter-retation of pi is that it represents the probability that a pixel in the

mage with intensity i corresponds to a bacterium, for pixels darkerhan the agar. There is a lower bound i0 which is a cut-off valuever which image pixels fail to contribute to the bacteria density.he number of bacteria pixels in the image is provided by

Imax∑=Imin

pi(i − G) (7)

The peak of the intensity values (red line) in every image wastted with a Gaussian distribution (blue line) that yields the bacte-ial curve (green dotted line). The Gaussian distribution parameters

(mean) and � (standard deviation) were calculated for all imagesn the datasets through an automated procedure and subsequently,acteria coverage was calculated along with the spatial bacterialistributions. All dark pixels that correspond to bacteria have beenainted in red (Fig. 5b). Occasionally, when the automated fit with aaussian fails, a fit on a different range of intensities (i.e., not in thehole range) is performed and if the procedure still leads to poor

esults, either a manual threshold is selected or the image is sim-ly removed from consideration. Then, pixels with intensity lesshan the threshold were regarded to be bacteria. In principle, thefficiency of our approach was very high for bacterial lawn withedium density and a very small number of images required spe-

ial treatment. The latter often occurred for images characterised

y uneven illumination; regions in the centre of the image system-tically exhibited a higher degree of brightness, by approximately0% relative to the edges caused by the microscope system. A nor-alization procedure was followed to remove the discrepancy;

30 32 24 286.9 8.4 7.4 9.6

more specifically, every image with intensity I was replaced bya normalised image with new intensity I/D, where D representsthe average of all images in every set. If this manipulation, though,failed to provide a satisfactory bacteria counting either the entireimages or problematic regions of pictures were removed from theanalysis. Findings attained by the above method demonstrated thatthe percentage of images that turned to be problematic and failedto lead to acceptable results was extremely small (about 2%).

A spatio-temporal analysis was also conducted for bacteriacoverage on several days (plots are illustrated for days 2 and 3,respectively) and results demonstrated decrease of the bacterialdensity due to amoebal grazing (Fig. 10). This behaviour is indica-tive of a correlation of the amoeba density and reduced bacterialcoverage. Each image in the lattice was represented by a colouredbox where the colour corresponded to the bacterial coverage. Theuniform colouring of each box does not suggest that the bacte-rial coverage is constant in the whole image but the particularrepresentation of the spatial bacterial distribution is attributed tosimplicity reasons. The white lines in Fig. 10a and b represent theoriginal drop radius.

In the evaluation procedure, the main objective was to assessa set of requirements including rapid and accurate object recog-nition and reliability against the decision of experts. The accuracyof the tool comprising the semi-automated object/mobility classi-fication procedure and GUI was tested. It is evident that for anyautomated or semi-automated system to be of use and merit, apartfrom its ability to be fast in the extraction and classification ofobjects, it must be trustworthy. A convincing approach to esti-mate the reliability of the technique necessitates an examination invarious imaging conditions and a quantitative measurement of itsperformance in comparison with a human expert decision. Objectextraction and classification was performed in all image datasetscontaining organisms in bacterial lawns of various densities (low,medium, high). Depending on the number of objects in every image,computation time to extract all objects varied. Nevertheless, theprocedure to obtain morphological, mobility and type details didnot require longer than 10 s for every image (fifteen organismswas the maximum number of objects in every image). To illustratethe validity of the proposed method, Table 1 illustrates the num-ber of organisms in a lawn of bacteria of medium density during aperiod of five days. Furthermore, reliability of results was evaluatedagainst the decision of experts and results due to visual reclas-sification and manual morphological intervention (i.e., perimeterredrawing) are also presented in the same table. Table 1 shows

excellent correspondence between the two sets of results. Only afew objects required manual reclassification or perimeter redraw-ing by the expert (i.e., morphological operation entailed less than0.5 s/object). In principle, conflicting results were only noted when

G.D. Tsibidis et al. / Micron 42 (2011) 911–920 919

F entagi

otmtttaoETtmgtmplmalpciitla

tiaSoctaeanemepbmntcm

ig. 10. Bacterial coverage for days 2 (a) and 3 (b), respectively. Colour shows percn this figure legend, the reader is referred to the web version of the article.)

rganisms entailed on features that enabled a consistent and objec-ive computer-aided or expert judgement. The percentage of the

anually reclassified objects (Table 1) emphasises the efficiency ofhe machine performance and it indicates that the combination ofhe semi-automated system and the clustering algorithms consti-ute a trustworthy technique that provide precise measurementsnd minimises manual intervention. Unlike other intensity thresh-lding based methodologies (Degerman et al., 2009; Rabut andllenberg, 2004; Sbalzarini and Koumoutsakos, 2005; Tsibidis andavernarakis, 2007), our technique utilises a very robust probabilis-ic method of assessing mobility of objects. The aforementioned

ethodologies are capable to quantify object motility (includingrowth, division, etc.) accurately by tracking the centre of mass ofhe organisms, however, they are not efficient in tracking move-

ent inside the object boundary. Hence, the key impact of theroposed approach is significantly useful in cases where object out-

ine is not substantially mobile while interior parts of the organismove. Furthermore, viewing facilities of the developed GUI offered

ssistance to acquire a fast overview of the results while morpho-ogical correction functions facilitated manual intervention. Theroposed system is characterised by an increased computationalapability for precise extraction and classification of amoebae/cystsn lawns of bacteria of low and medium density. By contrast, signif-cant improvements are required for the analysis of sets of imageshat contain organisms in lawns of bacteria with high density. Evo-ution of bacteria colonies produces regions of high intensity whichre mistakenly regarded as parts of organisms.

In regard to the enumeration of bacteria, the applicability ofhe probabilistic algorithm was tested against visual estimationn bacterial lawns of various densities. Unlike previous reportsnd automated systems (Grivet et al., 2001; Jansen et al., 1999;ingleton et al., 2001), our method avoids the need for trainingf the system and it offers a precise computation of the bacterialoverage. Furthermore, most systems attempted to enumerate bac-eria coverage through image segmentation in order to attain shapend size characteristics of bacteria (Drozdov et al., 2006; Pernthalert al., 2003; Singleton et al., 2001). While such methodologiesppear to provide accurate results for large sized bacteria, they areot capable to assist in determining bacteria coverage in films cov-red with a small-sized bacteria population (Fig. 5b). For low andedium bacterial coverage about 95% of the images were analysed

fficiently and realistic computations of the bacterial coverage waserformed. By contrast, bacterial analysis in areas characterisedy a higher concentration is problematic and either a tediousanual or an alternative methodology is required. Therefore, sig-

ificant modifications and multiple selective decision rules needo be incorporated into the system to improve the efficiency of theomputer-aided bacterial computation. Nevertheless, the proposedethod for bacterial spatial characterisation appears to be highly

e of image area covered by bacteria. (For interpretation of the references to colour

suitable for a quantitative and qualitative investigation of systemscharacterised by a non high bacterial density. Furthermore, thetechnique represents an early initiative for a systematic approach ofmodelling the ecology of S. typhimurium in various environments.Although analysis demonstrates that stochastic behaviour at thelevel of individual has significant impact on spatial heterogeneityat the population level, some clear dynamic trends are apparent.At present, the mechanism involved in the bacterial decrease dueto grazing is not well understood and requires further investiga-tion. Protozoan population regulation is the major factor governingencystment and interestingly, the rate of encystment is controlledby a bacterial density dependent mechanism. In principle, determi-nation of the protozoan–bacteria dynamics can be modelled withreaction diffusion equations (Ekelund et al., 2002) which confirmsthe significance of the knowledge of the bacterial identification.The results obtained through the proposed probabilistic approachdemonstrate that the system can sample interactions of protozoanswith bacteria with high accuracy and it can constitute a major com-ponent of any improved analysis tool.

4. Conclusions

We have developed a widely applicable technique for accu-rate analysis and elucidation of amoebae–bacteria interactions.In the first application, amoebae and cysts were localised onlawns of a varied bacterial density and morphological quanti-ties were obtained. The method proved to be particularly usefulbecause it allowed a semi-automated analysis with a remarkablyhigh efficiency and reliability. A second technique was addition-ally presented that aimed to compute the bacterial coverage inlattice experiments. The method is also characterised by highefficiency and it demonstrates the decrease in the bacterial den-sity due to amoebal grazing. This finding indicated a correlationbetween amoeba density and reduced bacterial coverage. The eval-uation procedure demonstrated that the efficiency of the systemwas remarkably high and it constituted a competent method toobtain information that can elucidate protozoa–bacteria interac-tions, migration and growth events in an ecosystem at both theindividual and population levels. Moreover, our methodology couldbe also applied to other biological systems that exhibit a similarbehaviour.

References

Burroughs, N.J., Tsibidis, G.D., Gaze, W., Wellington, L., 2003. Study of spatial biolog-

ical systems using a graphical user interface. In: Computer–Human InteractionConference 2003, Conference Proceedings, vol. 1 , Heraklion, Crete, Greece, pp.48–52.

Casasent, D.P., Smokelin, J.S., 1994. Neural-net design of macro gabor wavelet filtersfor distortion-invariant object detection in clutter. Opt. Eng. 33, 2264–2271.

9 icron

C

D

D

EE

G

G

G

H

J

K

20 G.D. Tsibidis et al. / M

asasent, D.P., Smokelin, J.S., Ye, A.Q., 1992. Wavelet and gabor transforms for detec-tion. Opt. Eng. 31, 1893–1898.

egerman, J., Thorlin, T., Faijerson, J., Althoff, K., Eriksson, P.S., Put, R.V., Gustavsson,T., 2009. An automatic system for in vitro cell migration studies. J. Microsc. 233,178–191.

rozdov, V.N., Sergeeva, V.N., Maksimenko, S., Zemskaia, T.I., 2006. Computer systemfor image analysis of fluorescently stained bacteria. Mikrobiologiia 75, 861–864.

delstein, P.H., Meyer, R.D., 1984. Legionnaires’ disease. A review. Chest 85, 114–120.kelund, F., Frederiksen, H.B., Ronn, R., 2002. Population dynamics of active and total

ciliate populations in arable soil amended with wheat. Appl. Environ. Microbiol.68, 1096–1101.

aze, W.H., Burroughs, N., Gallagher, M.P., Wellington, E.M., 2003. Interactionsbetween Salmonella typhimurium and Acanthamoeba polyphaga, and observa-tion of a new mode of intracellular growth within contractile vacuoles. Microb.Ecol. 46, 358–369.

erlich, D., Mattes, J., Eils, R., 2003. Quantitative motion analysis and visualizationof cellular structures. Methods 29, 3–13.

rivet, M., Morrier, J.J., Souchier, C., Barsotti, O., 1999. Automatic enumeration ofadherent streptococci and actinomycetes on dental alloy by image analysis. J.Dent. Res. 38 (1–2), 33–42.

uws, S.A., Morley, R.J., Jones, M.V., Brown, M.R., Smith, A.W., 2008. Interactions ofsome common pathogenic bacteria with Acanthamoeba polyphaga. FEMS Micro-biol. Lett. 282, 258–265.

ansen, G.J., Wildeboer-Veloo, A.C.M., Tonk, R.H.J., Franks, A.H., Welling, G., 1999.

Development and validation of an automated, microscopy-based methodfor enumeration of groups of intestinal bacteria. J. Microbiol. Methods 37,215–221.

im, E.Y., Jung, K.C., 2004. Object detection and removal using genetic algorithms.In: Pricai 2004: Trends in Artificial Intelligence, Proceedings 3157 , pp. 411–421.

42 (2011) 911–920

Pernthaler, J., Pernthaler, A., Amann, R., 2003. Automated enumeration of groupsof marine picoplankton after fluorescence in situ hybridization. Appl. Environ.Microbiol. 69, 2631–2637.

Rabut, G., Ellenberg, J., 2004. Automatic real-time three-dimensional cell trackingby fluorescence microscopy. J. Microsc. 216, 131–137.

Sbalzarini, I.F., Koumoutsakos, P., 2005. Feature point tracking and trajectory anal-ysis for video imaging in cell biology. J. Struct. Biol. 151, 182–195.

Singleton, S., Cahill, J.G., Watson, G.K., Allison, C., Cummins, D., Thurnheer, T.,Guggenheim, B., Gmur, R., 2001. A fully automated microscope bacterialenumeration system for studies of oral microbial ecology. J. ImmunoassayImmunochem. 22, 253–274.

Soille, P., 1999. Morphological Image Analysis – Principles and Applications.Springer-Verlag, Berlin.

Soll, D.R., Voss, E., Varnum-Finney, B., Wessels, D., 1988. Dynamic morphology sys-tem: a method for quantitating changes in shape, pseudopod formation, andmotion in normal and mutant amoebae of Dictyostelium discoideum. J. Cell.Biochem. 37, 177–192.

Sonka, M., Hlavac, V., Boyle, R., 1998. Image Processing Analysis and Machine Vision.Stevens, A.R., O’Dell, W.D., 1973. The influence of growth medium on axenic culti-

vation of virulent and avirulent Acanthamoeba. Proc. Soc. Exp. Biol. Med. 143,474–478.

Tsibidis, G.D., Tavernarakis, N., 2007. Nemo: a computational tool for analyzingnematode locomotion. BMC Neurosci. 8, 86.

Tvarusko, W., Bentele, M., Misteli, T., Rudolf, R., Kaether, C., Spector, D.L., Gerdes, H.H.,

Eils, R., 1999. Time-resolved analysis and visualization of dynamic processes inliving cells. Proc. Natl. Acad. Sci. U.S.A. 96, 7950–7955.

Tvarusko, W., Bentele, M., Satzler, K., Spector, D.L., Eils, R., Misteli, T., 1998. A compu-tational approach to quantitative analysis of functional processes in living cells.Mol. Biol. Cell 9, 444a–444a, Meeting Abstract: 2576.