A Semiautomatic Cell Counting Tool for Quantitative Imaging of Tissue Engineering Scaffolds

A Semiautomatic Cell Counting Tool for QuantitativeImaging of Tissue Engineering Scaffolds

Sebastian De Boodt, PhD,1,2 Ahmad Poursaberi, MSc,1 Jan Schrooten, PhD,2,3

Daniel Berckmans, PhD,1 and Jean-Marie Aerts, PhD1,2

Automatic image analysis algorithms are in general dedicated quantification tools used for very specific types ofmicroscopic cell images, but are not robust enough to accurately quantify the cell number and distribution in thewide variety for fluorescence images that exist in the field of tissue engineering (TE) today, where cell–material(scaffold) interactions are being evaluated more and more. In this study, a semiautomatic algorithm was de-veloped that allows the user to manually count a limited part of a TE scaffold image, and then automaticallycounts the cells of the full image based on that calibration dataset. The algorithm was validated on images ofcells on a two-dimensional (2D) titanium (Ti) substrate, in a three-dimensional (3D) Ti scaffold and in a fibrinhydrogel by comparison with manual cell counting and with an indirect cell counting using metabolic assay. Theaverage relative error between this semiautomatic and the manual approach was 3.4% for the 2D Ti substrates,5.9% for the 3D Ti scaffolds, and 14.1% for the fibrin hydrogels. Hereby a proof of concept was delivered thatcould lead to an increased use of automated cell imaging as a reliable 2D and 3D quantitative tool for both basicbiological research and process control of clinical TE products.

Introduction

As part of the tissue engineering (TE) cycle, cells aregenerally seeded into a three-dimensional (3D) open

porous scaffold material, and then cultured in a bioreactor tocontrol the proliferation and differentiation of the cells suchthat the construct will be able to integrate and eventuallyheal a patient’s tissue defect after implantation.1 In the entirecascade from obtaining sufficient biological knowledge at thelaboratory scale to up-scaled and automated clinical TEproducts, noninvasive and quantitative imaging of cells inboth space and time inside 3D scaffolds is regarded as animportant tool to assess cell behavior.2,3 Cell visualizationhas evolved to a widely spread and very powerful tool dueto the development of a large array of fluorescent labels tostudy cell proliferation, viability, and gene expression.4,5

Technological advances in optical equipment and micro-scope stage incubators make that fluorescence is no longerlimited to destructive imaging of histological sections,but can be used for repeated live cell imaging to follow uplarge amounts of samples at the same time.6–12 Despite theenormous amount of available images generated using thepresent advanced technology, cell imaging is not being usedat its full potential because it is often only used for qualita-tive inspection.13 Therefore, thorough image analysis can

easily increase the amount of data obtained from a singleexperiment.

In TE and in biological research in general, there is a largevariety in fluorescence images that is caused by differences inthe image quality, cell density, resolution, illumination,scaffold structure and material and also the 3D geometry,which makes that not all cells are in focus. To obtain quan-tities like cell number and local cell density, cells should bedistinguished from the background, a process called seg-mentation. The image processing steps needed for this seg-mentation are very much dependent on the image featuresmentioned above, which makes it impossible to have oneuniversal cell counting algorithm for all of them. For severalspecific applications, dedicated cell detection softwarehas been developed.14–25 In the absence of such software,commercial and open source software provide easy-to-usebuilding blocks that allow the user to make a cell countingalgorithm that performs best for his specific application. Thelimitation of many of these automatic cell counting algo-rithms is that they are very prone to systematic under or overestimation of the cell number26,27 and as an outsider it is verydifficult to know the reliability of the data. Polzer et al., forexample, thoroughly validated their automatic cell countingalgorithm with manual counting (R2 = 0.975), counting afternuclei staining (R2 = 0.997), and hemocytometer cell counting

1Division M3-BIORES: Measure, Model & Manage Bioresponses, KU Leuven, Heverlee, Belgium.2Prometheus, Division of Skeletal Tissue Engineering Leuven, KU Leuven, Leuven, Belgium.3Department of Metallurgy and Materials Engineering, KU Leuven, Heverlee, Belgium.

TISSUE ENGINEERING: Part CVolume 19, Number 9, 2013ª Mary Ann Liebert, Inc.DOI: 10.1089/ten.tec.2012.0486

697

(R2 = 0.629) and concluded their algorithm was robust,fast, and reproducible for their specific application.16 This is,however, a laborious approach that strongly reduces theadvantage of automatic cell counting and therefore limitsits use.

To use fluorescence cell imaging at its full potential, imageprocessing software needs to catch up with the evolution influorescence labeling and microscope technology.13 Existingalgorithms are often accurate, but not robust enough to beused for a wide variety of images. The objective of this studywas to develop and validate a semiautomatic cell countingalgorithm that

� allows the user to manually count a limited part of a TEscaffold image, and then automatically counts the cellsof the full image based on that calibration dataset,

� provides an objective measure for the accuracy ofthe semiautomatic counting compared to the manualcounting of the user,

� without any adaptations can be used for a variety ofcommonly used TE scaffold types.

Materials and Methods

Fluorescence images

Choice of image types. Three image types that are rep-resentative for TE research and which have an increasingcomplexity for automatic cell counting were selected: (1) atwo-dimensional (2D) titanium (Ti) substrate, (2) a 3D Tiscaffold, and (3) a fibrin hydrogel. The 2D Ti substrates werechosen to develop the algorithms. In a 3D scaffold geometry,it is not possible to have all the cells in a scaffold in focus inone image. Therefore, any image-based cell counting methodcan be used for counting all the cells in the image, but not allcells in the scaffold. The 2D Ti substrate was included in thisstudy as an intermediate between a 2D cell culture and a 3Dscaffold. Ti was chosen over a transparent microscope glassbecause of its use as TE scaffold material and because it is nottransparent for fluorescent light, making it a more realisticimage for TE scaffolds. Because of the flat surface, all cells on

the substrate were visible in the images, and therefore, thecounted cell number could be correlated to the total cellnumber that was obtained by a metabolic assay. In the 3D Tiscaffold images, the cells that are in focus are smaller andhave a sharper outline, while the cells out of focus appearbigger and less sharp in the image. These 3D features in-crease the complexity for automatic cell counting comparedto the 2D substrate. Cells that are encapsulated in a hydrogelgenerally appear rounder than cells that are attached to asurface. Depending on the optical properties of the material,hydrogels often generate an uneven background in fluores-cent images, which increases the complexity further for au-tomatic cell counting compared to the 2D substrates and 3DTi scaffolds.

2D Ti substrates. Six Ti substrates were produced withselective laser melting (SLM) and had a surface for the cellsto attach of 0.5 mm · 2 mm (Fig. 1, column 1).28 Humanperiosteum-derived stem cells (hPDCs) were stained withred fluorescent CellTracker� CMDiI (Invitrogen). After ex-pansion, cells were detached from the flasks, centrifuged,and suspended in 2 mM of the CMDiI solution. Cells wereincubated for 5 min at 37�C, and then for 15 min at 4�C tolimit endocytosis of the dye. Cells were centrifuged and re-suspended twice in phosphate-buffered saline (PBS) (LonzaGroup Ltd.) to remove residual dye. Six substrates were dropseeded with 200mL stained cell suspension at six differentdensities (0/2000/5000/8000/10,000/15,000 cells/cm2). Thecells were incubated overnight in a growth medium [theDulbecco’s modified Eagle’s medium (DMEM) with 10%fetal bovine serum (FBS)] at 37�C, 5% CO2, and 95% relativehumidity (RH) to allow cell attachment to the substrate.Next, the substrates were rinsed with PBS to remove non-adherent cells.

3D Ti scaffolds. The regular cylindrical 3D Ti scaffoldswere produced with SLM and measured 10 mm in heightand 6 mm in diameter (Fig. 1, column 2).28 Five scaffoldswere prepared and seeded with hPDCs according to themethod of Impens et al.29 After seeding and overnight

FIG. 1. Representativepictures of the three scaffoldtypes and the fluorescentimage of the cell-seededscaffolds (rows 1 and 2: scalebars = 5 mm; row 3: scalebars = 500 mm). Color imagesavailable online at www.liebertpub.com/tec

698 DE BOODT ET AL.

incubation, samples were rinsed with PBS to remove non-adherent cells. Live cells were then stained with 2mM CalceinAM (Invitrogen). After placing the samples at 37�C for20 min in the dark, the dye solution was discarded and theresidual stain was washed away with 2 mL of PBS.

Fibrin hydrogels. Cells were encapsulated in fibrin seal-ant (Tisseel VH S/D) to create cell-seeded carriers for 3Dculture. Five cylindrical hydrogels (8-mm diameter and 4-mmheight) were prepared with a final fibrinogen concentrationof 33 mg/mL and a thrombin concentration of 1 unit/mL(Fig. 1, column 3). The cell density was 1 · 106 cells/mL. Afterexpansion, cells were detached from the flasks, centrifuged,and resuspended in the thrombin component. The cell–thrombin solution was added to an equal volume of thefibrinogen component, vortexed briefly, and pipetted into acustom-made stainless steel mould. Subsequently, the mouldwas placed at 37�C for 1 h. After removing the hydrogelsfrom the mould, they were rinsed with 2 mL PBS and placedin 12-well plates in a 2 mL medium. Wells were coated withagarose to prevent cell attachment. After rinsing with 2 mL ofPBS, samples were first cut in half, and then stained with2mM Calcein AM. After placing the samples at 37�C for20 min in the dark, the dye solution was discarded and theresidual stain was washed away with 2 mL of PBS.

Microscopic imaging

Overview images of cells seeded on/in 2D/3D Ti sub-strates and encapsulated in fibrin hydrogels were taken witha stereomicroscope (SteREO Discovery.V8; Carl Zeiss Mi-croImaging, Inc.) equipped with a cooled charge-coupleddevice camera (SPOT Insight 2MP Firewire Colour Mosaiccamera; Diagnostic Instruments, Inc.). During imaging, all

samples were submerged in PBS to prevent drying. Four top-view images (1200 · 1600 pixels), which were partiallyoverlapping the longitudinal direction, were taken from each2D Ti substrate. These four images were manually combinedusing Matlab (The Mathworks, Inc.) to make one overviewimage of the entire sample substrate and cropped to a size of4000 · 1200 pixels. The 3D Ti scaffold images were taken inone top-view image (1200 · 1600 pixels). Two partiallyoverlapping images (1200 · 1600 pixels) were taken of eachfibrin hydrogel cross section. These two images were man-ually combined using Matlab to make one overview image ofthe entire cross section of each fibrin hydrogel (2000 · 1100pixels). All images were stored as 24-bit Red-Green-BlueTIFF files.

Semiautomatic cell counting algorithm

The semiautomatic cell counting algorithm, can be split upin four modules as depicted in Figure 2: (1) selection of arectangular patch of the image and manual counting of thecell in the patch, (2) automatic cell counting of the patch forall image processing parameter value combinations, (3) de-termining the best parameter value combination by means ofthe total mismatch, and (4) automatic cell counting of theoriginal image with the selected parameter values. Figure 2also illustrates that only module 1 requires a manual inter-action by the user and all the other modules are done auto-matically by the computer, hence, the term semiautomatic.The algorithm was made in Matlab. This paragraph explainsthe four modules subsequently.

Selection of a rectangular patch of the image and manualcounting of the cell in the patch. In the first module, theuser selects a rectangular image patch. The patches used here

FIG. 2. Schematic explanation of the semiautomatic cell counting algorithm, indicating the manual module on top and threeautomatic modules below. The images and numbers shown here are illustrative. The same algorithm was used to count thecells on the two-dimensional titanium (Ti) substrates, the three-dimensional (3D) Ti, and the fibrin hydrogel scaffold images.Color images available online at www.liebertpub.com/tec

SEMIAUTOMATIC CELL COUNTING FOR TISSUE ENGINEERING 699

were 200 · 200 pixels. In this image patch, the user clicks onall the centers of what he/she considers to be cells, thusstoring the manual cell coordinates.

Automatic cell counting of the patch for all image proces-sing parameter value combinations. To determine thenumber and coordinates of the cells in the fluorescent image,the image has to be converted to a binary image of whitecells (pixel values one) against a black background (pixelvalues zero). The image processing sequence that was usedfor this can be found in Figure 3. There are three parametersin this sequence that are sensitive to the differences in cellappearance in the different images: (1) size of the isotropicsecond-order Gaussian derivative filter, (2) gray valuethreshold, and (3) use of watershed. The second-orderGaussian filter enhanced the contrast of cell-shaped pixelareas, which had a higher gray value in the center and alower gray value at the edges. Preliminary tests showed thatthis was a good filter to increase the contrast between cellsand background and between adjacent cells in the image.This allowed for the gray value threshold to be used effec-tively for separating cells from the background. In general,when two cells are touching each other, the region wherethey touch is less wide than the diameter of the cell. There-fore, the watershed was a good algorithm to separate cellsthat were in contact with each other after the gray valuethreshold. The possible values of these parameters and theirmeaning with respect to the automatic cell counting aregiven in Table 1. The second-order Gaussian derivative filterincreased the contrast of the cells, and was implementedaccording to Geusebroek et al. and Freeman and Adelson.30,31

Watershed, which is available in the Image ProcessingToolbox in Matlab, is an algorithm that creates a distancemap (for each white pixel, the distance to the closest blackpixel is calculated) of a binarized image, and then splits thisdistance map where local minima are found. The algorithmperforms the image processing sequence (Fig. 3) on themanually selected image patch, for all 5610 combinations( = 11 · 255 · 2) of the values of these three parameters

(Table 1). The output of this module is a library of 5610 sets ofcell coordinates.

Determining the best parameter value combination bymeans of the total mismatch. The goal of this step is toselect the best parameter value combination (Table 1) forautomatic cell counting in the image patch. This is done bydetermining which set of automatically determined cell co-ordinates (output of module 2) are most similar to themanually determined cell coordinates (output of module 1).To quantify the similarity between two sets of cell coordi-nates, the total mismatch was defined as follows: if the dis-tance between an automatically and a manually detected cellis less than five pixels, than this is assumed to be the samecell and the manual–automatic cell duo is classified as amatched cell. The five-pixel distance is to take into accountthat the user will not click the cell center every time, but thisdistance will never be equal or more than five pixels. Everymanually and automatically detected cell can only be mat-ched once. After the matching step, the total mismatch(MMtotal) is calculated as follows:

MMtotal¼N�autoþN�man

N�autoþN�manþNmatch

� �· 100%, (1)

with N�auto the number of unmatched automatically detectedcells, N�man the number of unmatched manually detectedcells, and Nmatch the number of matched cells. The parametervalue combination that results in the lowest total mismatch isthen selected as optimal and is used to count the cells in theentire image. The total mismatch of the image patch is givenas an output of the algorithm and is an estimate of the errormade by the automatic cell counting.

Automatic cell counting of the original image with the se-lected parameter values. The selected parameter valuecombination (output of module 3) is used to automati-cally count the cells in the entire image. The output of thecell counting algorithm is the cell number and the cell

FIG. 3. Example images after subsequent image processing steps: (a) original image, (b) after conversion to gray value, (c)after morphological opening with a 15 pixel diameter disk-shaped structuring element to equalize the background intensitylevel, (d) after isotropic second-order Gaussian derivative filter increases the contrast of cell-shaped structures, (e) after grayvalue threshold, (f ) after watershed to separate touching cells, (g) each cell labeled with a different color, and (h) overlay ofthe cell center points on the original image. Color images available online at www.liebertpub.com/tec

700 DE BOODT ET AL.

coordinates in the entire image and the total mismatch of thecell count in the image patch used for manual counting.

Accuracy detection by manual cell counting

The accuracy of the semiautomatic cell counting algorithmwas quantified by comparison with manual cell counting.The standard deviation of manual cell counting was quan-tified. The same images used for semiautomatic cell countingwere also manually counted and the total mismatch wasused as an accuracy measure.

Intra- and interoperator variability of manual counting.Since the manual cell counting is also prone to errors, we firstaimed at quantifying the interoperator and intraoperatorvariability of the manual cell counting on the 2D substrateimages by three experienced operators. Ten image patches(100 · 100 pixels) were taken from the 2D substrate images. Aseries of 30 patches was created using all 10 images threetimes in a randomized order. This same series was thenmanually counted by three experienced operators. The in-traoperator variability of one operator for one patch wascalculated as follows:

Intraoperator variability¼ S:D:[Ncells]

Average[Ncells]· 100% (2)

where S.D.[Ncells] is the standard deviation of the three cellnumbers that were counted by the same operator on thesame patch, and Average[Ncells] is the average of these threecell numbers. The interoperator variability for one patch wascalculated as follows:

Interoperator variability¼ S:D:[Average[Ncells]]

Average[Average[Ncells]]· 100% (3)

where S.D.[Average[Ncells]] is the standard deviation of theaverage cell counts of each operator for this patch andAverage[Average[Ncells]] is the average of all the cell counts ofall operators for this patch.

Manual cell counting of the entire scaffolds. All imagesused in this study were also completely manually counted byone experienced operator. A user interface was made thatenabled the person to magnify the image and click the cells.

Accuracy measures. The total mismatch (Eq. 1) wascalculated for the manual and semiautomatic cell countingfor the entire scaffolds. The total mismatch consists ofan unmatched manually counted cell fraction and an

unmatched semiautomatically counted cell fraction calcu-lated as follows:

MMman¼N�man

N�autoþN�manþNmatch

� �· 100% (4)

MMauto¼N�auto

N�autoþN�manþNmatch

� �· 100% (5)

These are measures for the underestimation and theoverestimation of the semiautomatic cell counting algorithm,respectively.

Metabolic activity on the 2D Ti substrate

One reason to use the 2D Ti substrates (instead of the3D Ti or the fibrin hydrogel scaffolds) to develop thecell counting algorithm was that all the cells on the 2Dsubstrates were visible in one image. Consequently, thesemiautomatically counted cell number could be comparedwith a biological assay that is related to the cell number. AnalamarBlue� assay was performed on the 2D Ti substratesafter microscopic imaging to indirectly quantify the cellnumber based on the metabolic activity of the cells on thesubstrates. The scaffolds were put in a 500 mL culture me-dium (the DMEM with 10% FBS) with 10% alamarBlue andwere incubated for 4 h at 37�C, 5% CO2, and 95% RH. Ab-sorbance was measured using a 544 nm excitation wave-length and a 590 nm emission wavelength. A standard curvewith known cell numbers was used to indirectly estimate thecell number. The standard curve was created by seedingknown cell numbers (500, 1000, 2000, 5000, 10,000, 20,000cells) in 12-well plates (triplicate samples), which were in-cubated overnight in the growth medium at 37�C, 5% CO2,and 95% RH to allow cell attachment. Next, the metabolicactivity was measured as described above.

Statistics

Quantitative results are represented as mean – standarderror of the mean. Comparative studies of means were per-formed using a balanced one-way analysis of variance. Sig-nificant differences were determined with probability of 0.05and a significance level of 0.05.

Results

2D Ti substrates

Manual counting. The average intraoperator variability,which is a measure for the repeatability of manual cell

Table 1. Tuneable Parameters of Image Processing Sequence

Parameter Values Influence on the counted cell number

Filter size 2/2.2/2.4/ ./4 Determines the contrast between the cells and the background.Gray value

threshold1/2/3/ ./255 When set too high, parts of the background will be labeled as cells, leading to

an overestimation. When set too low, some cells will be labeled as backgroundand will not be counted, leading to an underestimation.

Watershed On/Off Used to separate connected cells ( = segmentation). Can lead to over-segmentationand consequently to overestimation.

Specifically at low cell density, a better result is often obtained without watershed.

SEMIAUTOMATIC CELL COUNTING FOR TISSUE ENGINEERING 701

counting on the 2D Ti substrate images by a single operator,was 5.5% (Fig. 4A). This is lower than the average inter-operator variability, which is a measure for the difference inmanual cell counting between different operators, which was12.9%. This latter value was used to estimate the standarddeviation of the manual counting of the entire 2D Ti sub-strate images shown in Figure 5A, because it reflects theuncertainty of the manually determined cell number.

Metabolic activity. The graph in Figure 4B shows therewas a linear relation (R2 = 0.99) between the metabolic ac-tivity measurement and the cell number of attached hPDCsas measured by manual cell counting. This relation was usedto calculate the cell number on the 2D Ti substrates shown inFigure 5A.

Semiautomatic cell counting. The average difference be-tween the manually and semiautomatically determined cellnumber was 3.4%, which was not significant (Fig. 5A). Theexample in Figure 6A and B shows, however, that althoughthe majority of the cells matched (green), there are also un-matched semiautomatic (red) and unmatched manual cells(blue). Figure 5B shows the ratios of these three cell fractionsfor all 2D Ti substrates. The total mismatch, which is the sum

of the two unmatched fractions, is 25.2% on average. It in-creases with cell density from 18.2% for 1940 cells/cm2 to35.0% for 12,680 cells/cm2. The cell number estimated fromthe metabolic activity measurement was significantly lowerthan the semiautomatically determined cell numbers, buthad a similar trend. This difference could be explained by asmall error in the cell number calculated from the metabolicactivity measurements, since this calculation was based on acalibration curve that was created for cells cultured on 2Dculture plastic instead of 2D Ti substrates.

3D Ti scaffolds

For the 3D Ti scaffolds, the average absolute differencebetween manually and semiautomatically determined cellnumbers was 5.9% (Fig. 7A). Figure 8B shows the presenceof both unmatched manually counted cells (blue dots) andunmatched semiautomatically counted cells (red dots),which were homogeneously distributed across the scaffold.The total mismatch, which is the summation of both un-matched fractions, was 28.4% on average (Fig. 7B).

Fibrin hydrogels

For the fibrin hydrogels, there was a net underestimationof the cell number by the semiautomatic counting of 14.1%compared to the manual counting (Fig. 9A). Figure 10Bshows that apart from the matched cells (green), there werealso unmatched manually counted (blue dots) and semiau-tomatically counted (red dots) cells. Unlike for the other twoscaffold types, there was a heterogeneous distribution ofboth unmatched cell fractions across the fibrin hydrogel. Atthe scaffold edges, where the background intensity wasdifferent, there were locally higher concentrations of un-matched semiautomatically counted cells (red dots), whereasthe unmatched manually counted cells were more concen-trated in the center of the scaffold (blue dots). Figure 9Bshows that there is on average a higher total mismatch 35.7%than for the other scaffold types, caused by the higher per-centage of both underestimation and overestimation.

Discussion

The semiautomatic approach could count cells on a 2D Tisubstrate, in a 3D Ti scaffold, and in a fibrin hydrogel withan average difference in the cell number compared to manualcounting of 3.4%, 5.9%, and 14.1%, respectively. The averagetotal mismatch, which takes into account the underestimatedand overestimated cell fraction, was respectively 25.2%,28.4%, and 35.7% on average for these three scaffold types.To put the performance of the algorithm in perspective,several things should be kept in mind. First of all, the aver-age difference in the cell number is an underestimation of theaccuracy, since the number of unmatched manual andsemiautomatic cells cancel out each other. The total mis-match does account for this effect and, consequently, resultsin higher error values than the absolute difference in the cellnumber. It is opportune to mention here that the majority ofthe unmatched fraction is caused by under- and over-seg-mentation of connected cells and not by objects in the imagesthat do not represent cells. Other studies that use dedicatedimage processing for one specific image type, report lowertotal mismatch values (7%–10%).17,19 However, because

FIG. 4. (A) Inter- and intraoperator variability of manual cellcounting on two-dimensional (2D) Ti substrate images. (B) Ca-libration of the alamarBlue assay for human periosteum-derivedstem cells attached to a 2D substrate, including the individualdata points and the linear trend line fitted by least squares.

702 DE BOODT ET AL.

these algorithms are developed for one specific application,their use for TE is limited.

There is a limit to the accuracy that can be obtained by anyautomatic or semiautomatic cell counting algorithm, which isdetermined by the repeatability of the manual cell counting.Numerous studies have depicted the interoperator variabil-ity in manual visual cell counting with reported values of 3%for well-separated leucocytes to 59% for tightly packedfluorescent cells.17,21 For the 2D Ti substrate, used tobenchmark the algorithm, the interoperator variability was12.9%, which was higher than the average difference in thecell number between the manual and the semiautomatic cell

counting on the 2D Ti substrates (3.4%). Keeping this inmind, the accuracy of this semiautomatic approach was ac-ceptable as proof of concept.

Purpose-specific cell counting algorithms have been de-veloped for various specific applications like white bloodcells20 or hepatocytes23 on 2D glass plates, fibroblasts in ahydrogel,14 or stem cells in embryoid bodies.17 The semiau-tomatic approach, however, is capable of counting cells invery different image types without any manual interventionsin the image processing sequence itself. It should be recog-nized that this approach will need further development fordifferent image types. For example, in the fibrin hydrogels,

FIG. 5. (A) Bar chart ofseeded cell number, cellnumber determined byalamarBlue, semiauto-matically, and manuallycounted cell number. (B)Ratio of match, underestima-tion, and overestimation ofall 2D Ti substrate images.

FIG. 6. (A) Example of anoriginal image of a 2D Tisubstrate with an inset ofthe manually counted patch.(B) Image with matched(green), overestimated (red),and underestimated cells(blue). (C) Contour plot ofthe local cell density. Colorimages available online atwww.liebertpub.com/tec

SEMIAUTOMATIC CELL COUNTING FOR TISSUE ENGINEERING 703

there was a net underestimation at locations of extreme highor low background intensity (mainly at the edges). Back-ground heterogeneity is a challenge that is not unique to thisalgorithm and existing algorithms can most likely be im-plemented to overcome this limitation.16

An important goal of this study was to find an objectiveway of showing the accuracy of the semiautomaticallycounted cells. The manually counted patch provided a wayto do this. The main advantage of the semiautomatic cellcounting is that the manual interaction is shifted from theimage processing to the manual cell counting. This way theexpertise of the researcher is used and the results are notinfluenced by manual changes in the image processing pa-rameters. Based on an image of the manually counted patchand on the numerical data from the error in cell number andtotal mismatch, any outsider can critically asses the reliabilityof the results. This is a crucial requirement for any automaticcell counting method to be accepted for research and to beused as a routine method for analysis or quality control.

This method can be categorized as computer learning,which has a broad toolbox of algorithms that can be usedalso for counting cells. Artificial neural networks, for exam-ple, is a learning method that is very usable for efficientlytraining the image processing parameters and has alreadybeen explored for cell counting.11 It enables the user to definethe desired total mismatch and can then iteratively belearned after every manually counted cell. This minimizedthe manual counting and thereby increased the efficiency ofthe training step.

The algorithm now trains for only three parameters, butcan be easily expanded with any number of extra modules.Most of the existing filters that were developed for dedicatedtrue automatic algorithms can be used to drastically increasethe robustness of the algorithm, without increasing the effortfor the user. Once this semiautomatic counting concept hasmatured, it could drastically increase the use of quantitativecell imaging. Less standardization would be needed for im-age acquisition, which provides biological researchers ahigher flexibility when designing experiments that includemicroscopic imaging or require information on cell numberand distribution when seeded in/on carrier materials.

Because this cell counting method is very generic, it couldpotentially be used in many other applications besides bio-reactor experiments and 3D cell culture for TE. One of thebiggest applications is probably the more standard cell cul-ture experiments on 2D plastic substrates (like flasks, Petridishes, or well plates), which in general are easily transfer-able to a microscope. In these experiments, live cell imagingin combination with semiautomatic cell counting could be apowerful tool to nondestructively measure changes in cell

FIG. 7. (A) Bar chart of semiautomatically and manually counted cell number of all five 3D Ti scaffold images (bars indicatesingle values). (B) Ratio of match, underestimation, and overestimation of all 3D Ti scaffold images.

FIG. 8. (A) An example of an original image of a 3D Ti scaffoldwith an inset of the manually counted patch. (B) Image withmatched (green), overestimated (red), and underestimated cells(blue). Color images available online at www.liebertpub.com/tec

704 DE BOODT ET AL.

number and cell distribution over time, thereby replacingdestructive DNA measurement techniques or indirectmethods like metabolic activity assays. As demonstrated inthis research, the metabolic activity (measured by an ala-marBlue assay) and cell number (measured by manualcounting) are highly correlated indicating that metabolicactivity assays might be used as an alternative method for

estimating cell numbers. Although this might be true forcertain applications, for TE applications, the use of (semi-)automatic cell counting tools has important advantagescompared to metabolic activity assays: (1) imaging tech-niques allow not only quantifying cell numbers, but also celldistribution; (2) metabolic activity assays need calibrationcurves for specific cell types and specific cell states, whereas

FIG. 9. (A) Bar chart ofsemiautomatically andmanually counted cellnumber for all five fibrinhydrogel images (barsindicate single values). (B)Ratio of match, underesti-mation, and overestimationof all five fibrin hydrogelimages.

FIG. 10. (A) Example ofan original image of a crosssection of a cell-seededfibrin hydrogel with aninset of the manuallycounted patch. (B) Imagewith matched (green), un-matched semiautomaticallycounted ( = overestimation,red), and unmatchedmanually counted cells( = underestimation, blue).Color images availableonline at www.liebertpub.com/tec

SEMIAUTOMATIC CELL COUNTING FOR TISSUE ENGINEERING 705

this is not necessary for imaging techniques; (3) imagingtechniques work also for low cell numbers, whereas formetabolic activity assays, a minimum number of cells isneeded; (4) imaging techniques allow very frequent estima-tion of cell numbers compared to metabolic activity assaysthat are much more time-consuming. However, both tech-niques are complementary and when applied together, theyallow generating more information about and gaining moreinsight in cell behavior, which can be used as a basis for amore optimal monitoring and control of cell behavior.

Acknowledgments

This work is part of Prometheus, the Leuven Research &Development Division of Skeletal Tissue Engineering of theKU Leuven. Thanks to Jan Demol and Yantian Chen for theimages of the fibrin hydrogels and 3D Ti scaffolds.

Disclosure Statement

No competing financial interests exist.

References

1. Langer, R., and Vacanti, J.P. Tissue engineering. Science 260,

920, 1993.2. Freed, L.E., Guilak, F., Guo, X.E., Gray, M.L., Tranquillo, R.,

Holmes, J.W., Radisic, M., Sefton, M.V., Kaplan, D., andVunjak-Novakovic, G. Advanced tools for tissue engineering:scaffolds, bioreactors, and signaling. Tissue Eng 12, 3285, 2006.

3. Eils, R., and Athale, C. Computational imaging in cell bi-ology. J Cell Biol 161, 477, 2003.

4. Lippincott-Schwartz, J., and Patterson, G.H. Developmentand use of fluorescent protein markers in living cells. Science300, 87, 2003.

5. Stephens, D.J., and Allan, V.J. Light microscopy techniquesfor live cell imaging. Science 300, 82, 2003.

6. Stephens, J.S., Cooper, J.A., Phelan, F.R.J., and Dunkers, J.P.Perfusion flow bioreactor for 3D in situ imaging: investigatingcell/biomaterials interactions. Biotechnol Bioeng 97, 952, 2007.

7. Tan, W., Sendemir-Urkmez, A., Fahrner, L.J., Jamison, R.,Leckband, D., and Boppart, S.A. Structural and functionaloptical imaging of three-dimensional engineered tissue de-velopment. Tissue Eng 10, 1747, 2004.

8. Moogk, D., Hanley, S., Ramunas, J., Blaylock, A., Skorepova,J., Rosenberg, L., and Jervis, E. Design and analysis of along-term live-cell imaging chamber for tracking cellulardynamics within cultured human islets of Langerhans. Bio-technol Bioeng 97, 1138, 2007.

9. Rice, W.L., Kaplan, D.L., and Georgakoudi, I. Quantitativebiomarkers of stem cell differentiation based on intrinsictwo-photon excited fluorescence. J Biomed Opt 12, 060504,2007.

10. Sawyer, N.B., Worrall, L.K., Crowe, J.A., Waters, S.L., Sha-kesheff, K.M., Rose, F.R., and Morgan, S.P. In situ monitor-ing of 3D in vitro cell aggregation using an optical imagingsystem. Biotechnol Bioeng 100, 159, 2008.

11. Rudolph, G., Lindner, P., Gierse, A., Bluma, A., Martinez, G.,Hitzmann, B., and Scheper, T. Online monitoring of micro-carrier based fibroblast cultivations with in situ microscopy.Biotechnol Bioeng 99, 136, 2008.

12. Wei, N., You, J., Friehs, K., Flaschel, E., and Nattkemper,T.W. An in situ probe for on-line monitoring of cell densityand viability on the basis of dark field microscopy in con-

junction with image processing and supervised machinelearning. Biotechnol Bioeng 97, 1489, 2007.

13. Meijering, E., Dzyubachyk, O., Smal, I., and van Cappellen,W.A. Tracking in cell and developmental biology. SeminCell Dev Biol 20, 894, 2009.

14. Thevenot, P., Nair, A., Dey, J., Yang, J., and Tang, L. Method toanalyze three-dimensional cell distribution and infiltration indegradable scaffolds. Tissue Eng Part C Methods 14, 319, 2008.

15. Dufour, A., Shinin, V., Tajbakhsh, S., Guillen-Aghion, N.,Olivo-Marin, J.C., and Zimmer, C. Segmenting and trackingfluorescent cells in dynamic 3-d microscopy with coupledactive surfaces. IEEE Trans Image Process 14, 1396, 2005.

16. Polzer, H., Haasters, F., Prall, W.C., Saller, M.M., Volkmer,E., Drosse, I., Mutschler, W., and Schieker, M. Quantificationof fluorescence intensity of labeled human mesenchymalstem cells and cell counting of unlabelled cells in phasecontrast imaging: an open source based algorithm. TissueEng Part C Methods 16, 1277, 2010.

17. Faustino, G.M., Gattass, M., Rehen, S., and de Lucena, C.J.P.Automatic embryonic stem cells detection and countingmethod in fluorescence microscopy images. Paper presentedat the IEEE International Symposium on Biomedical Ima-ging: From Nano to Macro, Boston, MA, 2009, 1, 799.

18. Ancin, H., Dufresne, T.E., Ridder, G.M., Turner, J.N., andRoysam, B. An improved watershed algorithm for countingobjects in noisy, anisotropic 3-D biological images. Paperpresented at the IEEE International Conference on ImageProcessing, Washington, DC, 1995, 3, 172.

19. Kothari, S., Chaudry, Q., and Wang, M.D. Automated cellcounting and cluster segmentation using concavity detectionand ellipse fitting techniques. Paper presented at the IEEEInternational Symposium on Biomedical Imaging: FromNano to Macro, Boston, MA, 2009, 1, 795.

20. Qingmin, L., and Yingying, D. An accurate segmentationmethod for white blood cell images. Paper presented at theIEEE International Symposium on Biomedical Imaging:From Nano to Macro, Washington, DC, 2002, 1, 245.

21. Bergen, T., Steckhan, D., Wittenberg, T., and Zerfass, T.Segmentation of leukocytes and erythrocytes in blood smearimages. Paper presented at the Annual International Con-ference of the IEEE Engineering in Medicine and BiologySociety, Vancouver, BC, 2008, 1, 3075.

22. Kenong, W., Gauthier, D., and Levine, M.D. Live cell imagesegmentation. IEEE Trans Biomed Eng 42, 1, 1995.

23. Refai, H., Li, L., Teague, T.K., and Naukam, R. Automaticcount of hepatocytes in microscopic images. Paper presentedat the IEEE International Conference on Image Processing,Barcelona, 2003, 2, 1101.

24. Rabut, G., and Ellenberg, J. Automatic real-time three-dimensional cell tracking by fluorescence microscopy. J Mi-crosc 216, 131, 2004.

25. Doukas, C.N., and Maglogiannis, I. Automated cell apoptosischaracterization using active contours. Paper presented at theAnnual International Conference of the IEEE Engineering inMedicine and Biology Society, Lyon, 2007, 1, 812.

26. Hand, A.J., Sun, T., Barber, D.C., Hose, D.R., and MacNeil, S.Automated tracking of migrating cells in phase-contrast vi-deo microscopy sequences using image registration. J Mi-crosc 234, 62, 2009.

27. Hodneland, E., Bukoreshtliev, N.V., Eichler, T.W., Xue-Cheng,T., Gurke, S., Lundervold, A., and Gerdes, H.H. A unifiedframework for automated 3-D segmentation of surface-stainedliving cells and a comprehensive segmentation evaluation.IEEE Trans Med Imaging 28, 720, 2009.

706 DE BOODT ET AL.

28. Van Bael, S., Vandenbroucke, B., Kerckhofs, G., Schrooten, J.,and Kruth, J.P. Design and production of bone scaffolds withselective laser melting. Paper presented at the TMS AnnualMeeting and Exhibition, San Francisco, CA, 2009, 1, 333.

29. Impens, S.L., Chen, Y., Mullens, S., Luyten, F.P., andSchrooten, J. Controlled cell seeding methodologies: a firststep towards clinically relevant bone tissue engineeringstrategies. Tissue Eng Part C Methods 16, 1575, 2010.

30. Geusebroek, J.M., Smeulders, A.W.M., and van de Weijer, J.Fast anisotropic Gauss filtering. IEEE Trans Image Process12, 938, 2003.

31. Freeman, W.T., and Adelson, E.H. The design and use ofsteerable filters. IEEE Trans Pattern Anal 13, 891, 1991.

Address correspondence to:Jean-Marie Aerts, PhD

Division M3-BIORES: Measure, Model & Manage BioresponsesKU Leuven

Kasteelpark Arenberg 303001 Heverlee

Belgium

E-mail: [email protected]

Received: August 12, 2012Accepted: January 08, 2013

Online Publication Date: February 26, 2013

SEMIAUTOMATIC CELL COUNTING FOR TISSUE ENGINEERING 707