Automatic morphometry of synaptic boutons of cultured cells using granulometric analysis of digital images

Journal of Neuroscience Methods xxx (2005) xxx–xxx

Automatic morphometry of synaptic boutons of culturedcells using granulometric analysis of digital images

Dimiter Prodanova,b,∗, Joost Heeromac, Enrico Marania,b

a Neuroregulation Group, Department of Neurosurgery, Leiden University Medical Center, University of Leiden, P.O. Box 9604,NL-2300 RC Leiden, The Netherlands

b Biomedical Signals and Systems Group, Institute for Biomedical Technology, Twente University, Enschede, The Netherlandsc Department of Neurogenomics, Center for Neurogenomics and Cognitive Research, Free University Medical Center, Amsterdam, The Netherlands

Received 11 April 2005; received in revised form 27 June 2005; accepted 10 July 2005

Abstract

Numbers, linear density, and surface area of synaptic boutons can be important parameters in studies on synaptic plasticity in culturedneurons. We present a method for automatic identification and morphometry of boutons based on filtering of digital images using granulometricanalysis. Cultures of cortical neurons (DIV8 and DIV21) were fixed and marked with fluorescently labeled antibodies for synapsin I (a markerf processed.G fluorescentg n integrali above thet d.C r at the lightm©

K

1

f1lssst2

pe

f theritesptionre,

ng inthod

ptictons

icro-

ue incles

0d

0

or synaptic boutons) and MAP-2 (a marker for dendrites). Images were acquired on a confocal microscope and automaticallyranulometry, a morphological operator sensitive to the geometry and size of objects, was used to construct a filter passing fuzzyrains of a certain size. Next, the filter was overlaid with the original image (masking) and the positive pixels were identified by a

ntensity threshold (thresholding). Disjoint grains, representing individual boutons, were reconstructed from the connected pixelshreshold, numbered and their area was measured. In total, 1498 boutons with a mean diameter of 1.63± 0.49m (S.D.) were measureomparisons with manual counts showed that the proposed method was capable of identifying boutons in a systematic manneicroscopic level and was a viable alternative to manual bouton counting.2005 Elsevier B.V. All rights reserved.

eywords: ImageJ; Mathematical morphology; Synapsin 1; MAP-2; Cell culture

. Introduction

Studies of synaptic morphology can sometimes be greatlyacilitated by the use of neuronal cell cultures (Boyer et al.,998). In cultures, synaptic boutons are arrayed in a thin

ayer at relatively low densities. This allows individual mea-urements to be performed more easily. In studies concerningynaptic plasticity in cell cultures, morphological parametersuch as linear density and number of synapses, and size ofhe boutons can be particularly important (Tarsa and Goda,002; Palizvan et al., 2004).

To establish the dose- or time-dependence of a specificharmacological effect often a large number of differentxperimental groups are needed. This rapidly increases the

∗ Corresponding author. Tel.: +31 71 527 6760; fax: +31 71 527 6782.E-mail address:[email protected] (D. Prodanov).

number of individual samples to be evaluated. Because ohigh variability of the localizations of synapses on dendand the clustering of boutons, the stereological assumof homogeneity of spatial distribution is violated. Therefocomplete dendritic trees are counted. To facilitate counticell cultures we developed a reproducible and robust mefor automatic identification and morphometry of synaboutons. The method was further applied to synaptic boumarked for synapsin 1 immuno-fluoresecence from misland cultures of neocortical neurons.

2. Methods

2.1. Granulometric filtering

Perhaps the oldest and most frequently used techniqthe empirical sciences to quantify the size of solid parti

165-0270/$ – see front matter © 2005 Elsevier B.V. All rights reserved.oi:10.1016/j.jneumeth.2005.07.011

NSM-4018; No. of Pages 1

2 D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx

Fig. 1. Granulometry of a binary image. An image containing three disks with diameters 10, 16 and 26 pixels is presented (noted “Original”). Duringgranulometry, the image is successively sieved with a family of homothetic disk SEs of increasing diametersd. Each sieving step that removes a disk from theimage results in a local maximum inG(d). Derived images corresponding to theG(d) maxima—I10, I16 andI26—are depicted together with the used SEs—SE10

(d= 10), SE16 (d= 16) and SE26 (d= 26). By subtracting derived images, it is possible to isolate individual disks.

is to use a series of sieves with increasing mesh openings.To quantify the properties of discrete sets of objects Math-eron theorized empirical sieving into the formal concept ofmathematical granulometry (Matheron, 1975) (seeAppendixA). Granulometry was later applied in image analysis toboth binary and continuous tone images (Serra, 1982). Ina way similar to sieving grains, pixels comprising an imageare “sieved” according to their connectivity to similar pix-els imposed by a certain primitive geometric body termedstructuring element (SE) (for illustration of the principle seeFig. 1). An integral characteristic of granulometry is the dis-tribution of the pixels with respect to the diameters of theused SEs (Fig. 1). A local maximum in its normalized firstderivative, the granulometric size densityG(d), indicates thepresence of a number of objects matching the particular SE.Moreover, granulometry can act as a band-pass filter capa-ble of discriminating grains of a certain size based on theirsimilarity to a SE (Dougherty and Chen, 2001). By exploit-ing this property, images can be simplified substantially tosuccessfully isolate various classes of objects (Fig. 1).

2.2. Outline of the algorithm

The algorithm employed in the study can be outlined inthe following main steps:

(

( ks ind-e

(3) Construct the binary “mask” using thek-means clusteringalgorithm (Jain and Dubes, 1988)—a method that canovercome background autofluorescence of the studiedcell structures more easily (see Section4).

(4) Delete irrelevant structures by superimposing the maskon the original image using bit-plane logical AND oper-ation (masking). Threshold the resulting image using theintegral thresholding (Appendix A, Eq.(A.9))–a methodthat can reduce the variation in the area measurements(see Section4).

(5) Construct disjoint grains from the pixels that are abovethe threshold based on their 8-connected neighborhoodsas implemented in ImageJ.

(6) Enumerate and measure the grains constructed in thisway.

2.3. Immuno-cytochemical procedures

Micro island cortical cultures were prepared from embry-onic day 18 murine embryos (Heeroma et al., 2004). Neu-ronal cells were harvested on the 8th day (DIV8) and on the21st day (DIV 21) of in vitro culturing (densities approxi-mately 6000 neurons/cm2). Cultures were stained with anti-bodies against a marker for dendrites, microtubule-associatedprotein 2 (MAP-2) (Six et al., 1992), and a marker forst twot ri-t mf cu-b ning0 00

1) Perform granulometry of the image and computeG(d)(Appendix A, Eq.(A.8)).

2) Identify the scale of interest by the pattern of the peaG(d); select low boundIlow = S Elow and high bounimagesIhigh = S Ehigh, where “” denotes the opening operation (Appendix AEq. (A.3)), and subtract thimages.

ynaptic vesicles, synapsin I (Huttner et al., 1983). Cul-ures were fixed by 4% paraformaldehyde, washedimes for 15 min with PBS, and incubated in 0.1% Ton X-100 for 5 min followed by 4% fetal calf seruor 20 min. After washing with PBS, cultures were inated for 1 h at room temperature in a mixture contai.1% Triton X-100, mouse monoclonal anti-MAP-2, 1/2

D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx 3

(Boehringer, Alkmaar, The Netherlands) and rabbit poly-clonal anti-synapsin I, 1/1000 (clone E028;Heeroma et al.,2004) antibodies diluted in PBS. After washing three timeswith PBS, the cells were incubated for 1 h at room temper-ature in secondary antibodies conjugated to anti-rabbit-Cy5or anti-mouse-Alexa546 (Molecular Probes, Oregon, USA)(Heeroma et al., 2004). Finally, the slides were washed in PBSand cover-slipped with Dabco-Mowiol (Heimer and Taylor,1974).

2.4. Imaging and Morphometric analysis

Cultures were analyzed on a Zeiss 510 Meta confocalmicroscope (Carl Zeiss, Heidelberg, Germany). A set of57 high-resolution digital images of different cultures wasrecorded at a resolution of 4.45 pixels perm. Images wereacquired on two channels—cyan, comprising anti-synapsinstaining and red, comprising anti-MAP-2 staining. In orderto reduce the amount of debris, all but the area situated within3m distance of any MAP-2 positive dendritic branches wasblanked in the images.

Image processing and measurement steps were performedin the public domain software for image analysisImageJ(NIH, Bethesda, Maryland, USA;http://rsb.info.nih.gov/ij/).The algorithm was run on the synapsin channel of everyimage. The identified boutons were numbered and their areas,e sureda s ofd

2

cedi nda heo theit ingf e top e thes ctedn mberc ent”d

ctst ter-om rences agea chan-nc rintsw iden-t mentm roce-d their

Table 1Parameters and comparisons used in the validation studies

Algorithm→ Na Observer 1→ N1 Observer 2→ N2

Algorithm – Mm1 Mm2

Observer 1 Dm1 = |Na−Mm1|Na

– Mo

Observer 2 Dm2 = |Na−Mm2|Na

Do = |Na−Mo|Na

–

N1—number of objects detected by observer 1;N2—number of objectsdetected by observer 2;Na—number of objects detected by the algorithm;Mm—number of matched objects between an observer and the method;Dm—method observer disagreement;Mo—number of matched objectsbetween observer 1 and 2,Do—inter-observer disagreement.

counts were recorded per quadrant (seeTable 1). To deter-mine inter-trial disagreements, manually counted printoutsin the trials were compared to each other by the same proce-dure.

Data were analyzed statistically by means of Student’st, χ2, and Kolmogorov–Smirnow’s tests, regression analy-sis, and analysis of variance (ANOVA) including post hoccomparisons by Scheffe, Neuman–Keuls, and L.S.D. tests.Probability levels of less than 0.05 were considered signifi-cant.

3. Results

3.1. Synthetic image example

The working of the algorithm is illustrated inFig. 2. A syn-thetic image (Fig. 2A) was produced by painting randomlylocated circles using different brush sizes and varying degreesof fuzziness. Granulometry was performed with flat disk-shaped SEs.G(d) showed pronounced maxima that matchedthe brush diameters used for the drawing atd= 4, 7, 10, and13 pixels (Fig. 2D). The filtering procedure is exemplifiedfurther for the marked peak (Fig. 2D, asterisk). A total of66 grains comprising 3.6% of the image area was identified(orange,Fig. 2C).

3c

wasp romd bedr n( on( ig-n eA sig-n nw ion( n thet Thet

quivalent diameters, and planar co-ordinates were meand recorded. The derived images containing markingetected objects were printed.

.5. Validation by manual counts

Two experimental trials involving observers experienn morphometry were performed: a “free” count trial a

“calibrated” count trial. During the “free” count trial, tbservers had to mark each bouton they could identify in

mages according to a size criterion (see Section3). Duringhe “calibrated” count trial, to reduce the variability comrom the “nature” of the observers (i.e. disagreement dureferences to lower or higher counts) and to increastatistical power, as an additional constraint, the expeumber of boutons per image was preset to the nuounted by the algorithm (i.e. “tea tasting lady experimesign;Fisher, 1935).

To quantify the differences in identification of objehe notions of “method-observer disagreement” and “inbserver disagreement” were used (seeTable 1for the for-ulae). This was necessary due to the absence of a “refe

tandard” method, a frequent situation in automated imnalysis studies in the biological sciences. The synapsinels of a representative set of images (n= 9) were printed inontinuous grey tone and divided in quadrants. The pere presented independently to two observers to

ify, mark, and count boutons. To evaluate the disagreeeasures, objects marked either by the automatic pure or by the human observers were matched and

.2. Determination of the filtering parameters for theultured cells

As the fluorescent patches were round, granulometryerformed with a family of flat disk-shaped SEs ranging f= 1 to d= 25 (0.2–5.6m). A consistent pattern couldiscerned in the shape of the averagedG(d) (Fig. 3)—a peakegion (P) in the range 0.7–1.6m, an intermediate regioshoulder, S) in the range 2.0–2.5m, and a decreasing regitail, T) from 2.9m on. The discerned regions were sificantly different from each other (p< 0.0001, univariatNOVA). Post hoc comparisons of the means showedificant differences by all tests (p< 0.0001). The P-regioas larger than the tail (Fig. 3; asterisks, P) and the S-reg

Fig. 3; double asterisks, P). The S-region was larger thaail (Fig. 3; asterisks, S), but not larger than the origin.ail was smaller than the origin.

4 D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx

Fig. 2. Granulometric filtering procedure of a synthetic image. (A) Synthetic image with dimensions 200× 200 pixels containing five types of bright grains ofvarying grades of fuzziness. (B) The mask derived from the marked peak inG(d) (asterisk, D). (C) Co-localization of the mask (red) and the original image(green). Co-localized grains appear in orange color. (D) SmoothedG(d) of the synthetic image (A).

Fig. 3. Averaged granulometric size density distribution of the entire image set. For descriptive purpose, we discerned peak (P), shoulder (S), and tail (T)regions (see text). Data are presented as means and confidence intervals.

D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx 5

Fig. 4. Illustration of the granulometric filtering technique on a real image. (A) High-resolution image of synapsin I staining; (B) MAP-2 staining; (C) co-localization of synapsin I (cyan) and MAP-2 (red); (D) inverted filtered image, step 2 of the algorithm. E—mask image, step 3; F—co-localization of thedetected synaptic boutons (red outline) with the original synapsin I image. Scale bar—5m.

The P-region corresponded to the presence of a large num-ber of synapsin positive grains of matching sizes. The diam-eter range of these conformed to the size range of synapticboutons described previously (Alsina et al., 2001). Accord-ingly, for the construction ofI low values of eitherdlow = 3or dlow = 5 were selected depending on the individualG(d)and the amount of debris in an image. To constructIhigh,the S-region was considered as well because in some imageslarger synaptic boutons were noticed. Accordingly,dhigh = 11was selected as a parameter forIhigh. An example is given inFig. 4D.

Based on the discrimination of dark background, auto-fluorescing cell mass, and synapsin positive grains threebrightness classes were used during thek-means clustering.The brightest class was selected for the construction of themask (Fig. 4E). During the integral thresholding, sufficientoverlap of the thresholded particles with the actual synapsingrains was typically achieved for AF = 0.8 (area factor, seeAppendix A). An example outcome of the detection proce-dure is presented inFig. 4F.

3.3. Morphometric analysis and comparisons withhuman observers

Measurements of the cultured cells are reported solelyto illustrate the suitability of the method (Table 2). In total,1498 boutons were automatically detected and measured in57 images; 90% of them had diameters between 0.95 and2.57m. Therefore, a simple size criterion for the manualvalidation could be formulated—boutons were defined assharp synapsin-positive grains having diameters between 1and 2.6m.

3.3.1. Free count trialThe free count trial was performed on 451 quadrants

(n= 9 images, 422m2 per quadrant on average). Sum-mary of the trial is given inTable 3. The method-observerand the inter-observer disagreements did not differ signifi-cantly (Table 3; pairedt-test, all comparisons gavep> 0.05).Observer 1 (O1), observer 2 (O2), and the algorithm (A) iden-tified the same mean number of synapses (pairedt-test, in all

Table 2Descriptive statistics of the measured boutons

Mean Median Mode P5% P95% S.D. Skewness Kurtosis

Area (m2) 2.28 1.92 0.91 0.71 5.19 1.41 1.29 1.47D 0.

P ion.

iameter (m) 1.63 1.56 1.075

5%—5th percentile, P95%—95th percentile, S.D.—standard deviat

95 2.57 0.49 0.66 −0.06

6 D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx

Table 3Matched synapses and disagreements—free trial

Algorithm Observer 1 Observer 2

Algorithm – Mm1=10.33± 11.19 Mm2=12.67± 13.06Observer 1 Dm1=47.21± 15.58% – Mo=9.50± 6.39Observer 2 Dm2=30.62± 17.28% Do=56.12± 31.60% –

Data are presented per image as mean± S.D. (seeTable 1for the notations).

comparisonsp> 0.30). The counts correlated strongly amongeach other (A and O1—r = 0.99, A and O2—r = 0.96, O1and O2—r = 0.95; r denotes Pearson’s correlation coeffici-ent).

An important finding was the correlation between thenumber of mismatching objects of each observer and thenumber of objects counted by the algorithm in the sameimage (A and O1—r = 0.69, A and O2—r = 0.71). By usingthe “automatic” number as denominator in the disagreementmeasure, we compensated for this effect in the statistical anal-ysis. Based on this correlation, only the images with similardensity of synapsin-positive grains were selected for the cal-ibrated count trial.

3.3.2. Calibrated count trialIn the calibrated count trial, 77 objects in 348 quad-

rants (n= 7 images; 380.12m2 per quadrant on average)were counted (summary inTable 4). In total, 44 quad-rants containing synaptic boutons were marked by either ofthe observers. The inter-observer and the method-observerdisagreements did not differ significantly (Table 4; pairedt-test, in all comparisonsp> 0.40). The mean numberof matched boutons per observer did not differ signifi-cantly from the expected number of boutons (Table 4; χ2

test).Further, we examined the origins of the encountered

m rants,t serve2 iden-t e. In1 menw

3as

3 -p trial,t d 2.4t ered2

3.4. Effects of the duration of culturing

Unexpectedly, the bouton counts per image at DIV8(21.92± 26.41) were not significantly different from thebouton counts at DIV21 (31.24± 27.52) (p> 0.05, t-test).Bouton diameters were distributed uni-modally at bothDIV8 and DIV21 (Fig. 5). The mean bouton area andthe mean bouton diameter were larger at DIV21 (area2.35± 1.47m2; diameter 1.65± 0.51m; n= 906) thanat DIV8 (area 2.18± 1.31m2; diameter 1.60± 0.47m;n= 592) (p< 0.05;t-test). However, the distributions did notdiffer in shape (p> 0.05; Kolmogorov–Smirnow test). Daysin vitro had a statistically significant influence onG(d)(ANOVA, p< 0.0001).

3.5. Sensitivity of the automatic bouton identificationand measurements to parameter variation

The variation in the numbers and the area of the automat-ically identified synaptic boutons, caused by the choice ofparameter settings, was studied in the set of images used forthe manual verification (n= 9).

To study the influence of the area factor parameter, theoriginal granulometric filter settings were retained while theAF value was varied and the numbers and areas of the boutonsw areasoma( ann heds

set( tedo nap-t -e um-b singsd

TM

bserve

A m1 = 8.1OO =15.48

D ).

ethod-observer disagreement. In 81.82% of the quadhere was a complete agreement between observer 1, oband the algorithm; in 4.55% of the quadrants, boutons

ified by the observers were not present in the mask imag3.64% of the quadrants, the method-observer disagreeas due to the setting of the AF value.

.3.3. Inter-trial comparisonsThe inter-trial disagreement for observer 1 w

1.90± 15.92% and for observer 2 was 13.66± 6.37% comared for the same images. During the calibrated count

he averaged inter-observer disagreement was lowereimes and the method-observer disagreement was low.2 times compared to the free count trial.

able 4atched synapses and disagreements—calibrated trial

Algorithm O

lgorithm – Mbserver 1 Dm1 = 21.73± 13.08% –bserver 2 Dm2=22.57± 16.45% Do

ata are presented per image as mean± S.D. (seeTable 1for the notations

r

t

ere measured. As expected, both the numbers and thef the synaptic boutons depended on AF (Fig. 6A and B). Theean bouton area (MA) varied linearly with AF (Fig. 6A)nd obeyed the regression equation MA = 3.09AF− 0.11R2 = 0.995; β = 0.998; p< 0.0001). In contrast, the meumber of boutons varied non-linearly with AF and reacaturation for values exceeding 0.4 (Fig. 6B).

The bandwidth settings used to count the imagedlow = 5, dhigh = 11), denoted as “original”, were selecn the basis of literature data for the sizes of the sy

ic boutons (Alsina et al., 2001). To study the influnce of the granulometric filter bandwidth on the ner of identified boutons, the algorithm was rerun ueveral different bandwidths: broader bandwidth (dlow = 3,high = 15), denoted “broad”; narrower bandwidth (dlow = 5,

r 1 Observer 2

4± 2.79,p= 0.42 (χ2) Mm2 = 8.00± 2.77,p= 0.29 (χ2)Mo = 8.86± 3.09

± 10.45% –

D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx 7

Fig. 5. Effect on DIV on the size distribution of the synaptic boutons.

Fig. 6. Sensitivity of the automatic area measurements and the counts to the variation of AF. The mean bouton area (A) varied linearly (dashed line) with AFwhile the mean number of boutons (B) rapidly reached saturation.

dhigh = 9), denoted “narrow”; and shifted bandwidth (dlow = 7,dhigh = 13), denoted “shifted” (Fig. 7). The AF was kept fixedto 0.8. InFig. 7, the relative count differences (RD, see leg-end) are depicted against the “original” bandwidth counts(baseline). Broad bandwidth counts were higher than narrowbandwidth counts (p< 0.05; pairedt-test;Fig. 7, asterisk) andshifted bandwidth counts (p< 0.001; pairedt-test;Fig. 7, dou-ble asterisks). There were no differences with the observercounts (all comparisonsp> 0.21; pairedt-test; Fig. 7). Asexpected, the narrow bandwidth counts were lower than theoriginal and the broad bandwidth counts were higher thanthe original (0 was not in the 95% confidence interval ofRD for either of them). The shifted bandwidth counts were

marginally different from the original (0 was not in the 90%confidence interval of RD).

4. Discussion

Until now, analysis of images by granulometry wasapplied to the detection of fluorescent signals in “in situ”hybridization (Grigoryan et al., 2002), DNA micro-arrays(Angulo and Serra, 2003), and to the counting of bloodcells (Di Ruberto et al., 2002), and parasites (Theera-Umponet al., 2001). By means of similar granulometric analysisand filtering, we could successfully identify and perform

8 D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx

Fig. 7. Sensitivity of the automatic counts to the variation of filteringbandwidth. Relative differences—RDi = 1−Ni /Noriginal; i denotes either of“broad”, “narrow”, or “shifted” bandwidth counts, or the averaged observercounts in the free count trial (“manual”);Noriginal denotes the “original”bandwidth (dlow = 5, dhigh = 11) counts. Observer counts have the largestvariation. Data are presented as means and confidence intervals.

measurements on synaptic boutons using specific immuno-fluorescence (Fig. 3).

The most common approach in automatic image anal-ysis for identification of objects in a gray level image isto threshold the image and then to identify the objectsbased on a priori information of their sizes and shapes. Thisapproach has potentially two shortcomings. First, in fluo-rescence microscopy, the uneven illumination of the totalfield can result in a global intensity gradient across theacquired digital image caused by the emission of the mount-ing medium. If such an intensity gradient possesses intensitycomparable to the objects of interest, a global threshold willintroduce a bias towards the lighter intensities. Eventually,fusing of objects and overestimation of their area can occur.In contrast, an algorithm based on granulometry will have anadvantage because intensity gradients may affectG(d) onlyat very large scales and therefore will not interfere with theextraction of small objects. Second, if digital particles areclassified by the circularity of their shapes, values exceed-ing the theoretical bonds for perfect continuous circles (i.e.larger than one) can be calculated for the small grains con-sisting of several pixels. This is caused by the digitizationsampling resulting in underestimation of the particle perime-ter, i.e. the pixel is an elementary measure of area and not oflength. Therefore, the small digital particles will be misclas-sified resulting in bias in the identification procedure.

4

oseda al

image, in step 3, thek-means clustering was used. This algo-rithm divides the image histogram in multiple classes ofpixels that potentially discriminate between different objects(Jain and Dubes, 1988). The choice of multilevel threshold-ing was supported by the presence of cell structures havingvarying degrees of autofluorescence intermingled with theactual synapsin 1 signal. Such structures were not com-pletely removed by the granulometric filter (Fig. 4A andD). The automatic thresholding, based on iterative sub-histogram average intensity decomposition as implementedin ImageJ (http://rsb.info.nih.gov/ij/docs/faqs.html), did notidentify correctly the synaptic boutons in some of the fil-tered images. This was due to the overlapping distributionof the foreground pixels and the background gray levels asdiscussed in literature (Sezgin and Sankur, 2004).

In step 4 of the algorithm, an integral thresholdingapproach was used. By fixing AF, the estimate of the synap-tic boutonarea can be controlled (Fig. 6A) assuming thesame conditions for image acquisition (i.e. same microscopeoptics, magnification, and camera). Moreover, values of AFgreater than 0.4 did not affect the meannumberof identifiedbuttons (Fig. 6B).

4.2. Analysis of the manual counts

In both trials, human observers identified the same objectsa erd uallyt ree”t ent.P mbero inter-o l.

iousa time( dyb ncedo jectsi ast, ad bero eforep

4

d inte l-l ly toe dB et y, an cro-s icro-s 98;H by

.1. Choice of thresholding procedures

We used thresholding steps in two stages of the proplgorithm. After the granulometric filtering of the origin

s the algorithm did (Tables 3 and 4). The method-observisagreements indicated that the algorithm performed eq

o a human observer. An unexpected finding in the “frial was the magnitude of the inter-observer disagreemresetting the expected outcome of identification (i.e. nuf boutons to be detected by observers) reduced thebserver disagreement 2.4 times in the “calibrated” tria

Manual counts of anatomical structures can be tednd error-prone, and result in a very long processingBenali et al., 2003). As demonstrated also in our stuy the observed inter-trial disagreement, even experiebservers do not count exactly the same number of ob

f the same picture is presented several times. In contreterministic algorithm will always count the same numf objects with the same parameter settings and is therreferable (Geuna et al., 2001; Benali et al., 2003).

.3. Synaptic morphology

Within presynaptic terminals, synapsin I is concentratehe compartment occupied by the synaptic vesicles (Sudhoft al., 1989; Fletcher et al., 1991). It is considered an exce

ent marker for the synaptic terminal and is used routinestimate synaptic density (De Camilli et al., 1983; Moore anernstein, 1989; Fletcher et al., 1991). Due to the prohibitiv

ime expenditure of electron microscopic morphometrumber of studies were performed to correlate light micopic measurements of synaptic structures to electron mcopic measurements (Papa et al., 1995; Boyer et al., 19arata et al., 2001). It was found that the area occupied

D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx 9

synaptic vesicles loaded with fluorescent dye corresponded tothe dimensions of the bouton on electron microscopic level(Harata et al., 2001). Therefore, we can correlate the areaof the synapsin positive grain (i.e. the total projection areaof the synaptic vesicles) to the area of the synaptic bouton.Reported subtle differences in the mean bouton area (i.e. fourtimes smaller than the resolution of the image) and in theshape ofG(d) at DIV8 and DIV28 that do not reflect in cor-responding differences in the bouton area histogram shapesprobably correspond to slight differences either in the imageacquisition conditions or in the performance of the immuno-cytochemical procedures.

An advantage of the proposed automated approach is that,apart from the number of synaptic boutons in a light micro-scopic image, it also allows to be measured morphologicalcharacteristics such as area, circularity, and maximal diame-ter. The proposed approach is not restricted solely to boutondetection. At different scales and using suitable labeling othercell structures such as nuclei or axonal profiles can also beidentified and measured.

Acknowledgements

D. Prodanov was funded by a European Commission granta PRO( uralP W.S.R sda,M porto

A

A

andp perat ,1 y 3Df ,1 s bell onsi ningaS

S

S

O

S

andclosingof SbyE as

S • E = (S E) ⊕ E. (A.4)

Mathematically, granulometry is defined as a parameterisedfamily of openings or closings characterized by the scaleparameterd (Matheron, 1975). In the case of a homotheticfamily of SEs, granulometry can be represented by:

Ψd(S) =⋃

d

S dE. (A.5)

In the formula,d is the proportionality parameter of the homo-thety. By convention0(S) =Sand for negatived opening isreplaced by closing. For greyscale images, the measure of theinteraction with SE is the volumeV[ ] removed by sieving.Its mapping onto the scale axisd is called granulometric sizedistribution:

Ω(d) = V [S] − V [Ψd(S)]. (A.6)

The granulometric size density is the normalized generalizedderivative ofΩ(d):

G(d) = 1

V [S]

∂Ω(d)

∂d. (A.7)

The volume of image can be estimated from its histogramH∑255 n-

en-

rizednl

iond the

the5theesh-avingdard

.

s part of the Research and Training Network NeuralFramework 5, contract No HPRN-CT-2000-00030—Nerostheses). The authors would like to acknowledgeasband from the National Institutes of Health, Bethearyland, USA for his continuous development and supf ImageJ.

ppendix A

.1. Mathematical morphology

Mathematical morphology offers tools for analysisrocessing of images based on the key set theory o

ions inclusion (⊂, ⊆), union (∪) and intersection (∩) (Serra982). Grey level images and SEs can be represented b

unctions with their elevation indicating intensity (Sternberg986). The associated sets are then the geometric bodie

ow the functions—so called “umbras”. Principal operatin mathematical morphology are erosion, dilation, opend closing. IfS(x) is the umbra of the imageSandE(x) is aE set,erosionanddilation are defined by

⊕ E = maxb ∈ E

S(x − b) + E(b) (A.1)

E = minb ∈ E

S(x + b) − E(b) . (A.2)

peningof SbyE is defined as

E = (S E) ⊕ E (A.3)

-

-

by V [S] = g=0H(g) × g and the granulometric size desity by

G(d) = V [S dE] − V [S (d − 1)E]

V [S]. (A.8)

Gray-level morphological operations erosion, dilation, oping and closing were implemented as a sub-routine inImageJ(http://rsb.info.nih.gov/ij/plugins/gray-morphology.html).

A.2. Integral thresholding

To produce binary particles we introduce a parameteintegral threshold operator TrAF (Tmin, Tmax) depending othe area factor parameter AF. TrAF (Tmin,Tmax) labels a pixein the image if its intensityg falls between lower (Tmin) andupper threshold (Tmax) bounds constrained by the conditthat the ratio between the area of the labeled pixels anarea of the non-black pixels equals AF:

A[g ∈ [TminTmax]]

A[g > 0]= AF. (A.9)

In the formula,A[ ] denotes the area operator acting onwhole image. For the aims of the study,Tmax was set to 25while Tmin was determined by varying its value untilcondition was globally satisfied. The behavior of the throld operator was studied in several simulated images hvarying amounts of gaussian noise specified by its standeviation (50, 100, 150, and 200;Appendix B Supp. Fig1). The functional relationship between AF andTmin in thesynthetic images is depicted inSupp. Fig. 1.

10 D. Prodanov et al. / Journal of Neuroscience Methods xxx (2005) xxx–xxx

Appendix B. Supplementary data

Supplementary data associated with this article can befound, in the online version, atdoi:10.1016/j.jneumeth.2005.07.011.

References

Alsina B, Vu T, Cohen-Cory S. Visualizing synapse formation in arboriz-ing optic axons in vivo: dynamics and modulation by BDNF. NatNeurosci 2001;4:1093–101.

Angulo J, Serra J. Automatic analysis of DNA microarray images usingmathematical morphology. Bioinformatics 2003;19:553–62.

Benali A, Leefken I, Eysel UT, Weiler E. A computerized image analysissystem for quantitative analysis of cells in histological brain sections.J Neurosci Methods 2003;125:33–43.

Boyer C, Schikorski T, Stevens CF. Comparison of hippocampal dendriticspines in culture and in brain. J Neurosci 1998;18:5294–300.

De Camilli P, Cameron R, Greengard P. Synapsin I (protein I), anerve terminal-specific phosphoprotein. I. Its general distribution insynapses of the central and peripheral nervous system demonstratedby immunofluorescence in frozen and plastic sections. J Cell Biol1983;96:1337–54.

Di Ruberto C, Dempster A, Khan S, Jarra B. Analysis of infectedblood cell images using morphological operators. Image Vision Comp2002;20:133–46.

Dougherty ER, Chen Y. Robust optimal granulometric bandpass filters.Signal Proc 2001;81:1357–72.

F 935.

F n ofing in

G MG.bers

G r G,for

n situ

H zingpho-

Heeroma JH, Roelandse M, Wierda K, van Aerde KI, Toonen RF,Hensbroek RA, et al. Trophic support delays but does not preventcell-intrinsic degeneration of neurons deficient for munc18-1. Eur JNeurosci 2004;20:623–34.

Heimer GV, Taylor CE. Improved mountant for immunofluorescencepreparations. J Clin Pathol 1974;27:254–6.

Huttner WB, Schiebler W, Greengard P, De Camilli P. Synapsin I (pro-tein I), a nerve terminal-specific phosphoprotein. III. Its associationwith synaptic vesicles studied in a highly purified synaptic vesiclepreparation. J Cell Biol 1983;96:1374–88.

Jain AK, Dubes RC. Algorithms for clustering data. Prentice Hall: Engle-wood Cliffs, New Jersey; 1988. p. 1–320.

Matheron G. Random sets and integral geometry. 2nd ed. New York:Wiley; 1975. p. 1–288.

Moore RY, Bernstein ME. Synaptogenesis in the rat suprachias-matic nucleus demonstrated by electron microscopy and synapsin Iimmunoreactivity. J Neurosci 1989;9:2151–62.

Palizvan MR, Sohya K, Kohara K, Maruyama A, Yasuda H, Kimura F,et al. Brain-derived neurotrophic factor increases inhibitory synapses,revealed in solitary neurons cultured from rat visual cortex. Neuro-science 2004;126:955–66.

Papa M, Bundman MC, Greenberger V, Segal M. Morphological analysisof dendritic spine development in primary cultures of hippocampalneurons. J Neurosci 1995;15:1–11.

Serra J. Image analysis and mathematical morphology. London: AcademicPress Ltd; 1982. p. 318–72.

Sezgin S, Sankur B. Survey over image thresholding techniques andquantitative performance evaluation. J Electron Imaging 2004;13:146–65.

Six J, Lubke U, Mercken M, Vandermeeren M, Ceuterick C, Van d V,et al. Specific monoclonal antibodies against normal microtubule-

atho-Acta

S Proc

S A,in a474–

T se for-USA

T nulo-att

isher RA. The design of experiments. Edinburgh: Oliver and Boyd; 1p. 1–320.

letcher TL, Cameron P, De Camilli P, Banker G. The distributiosynapsin I and synaptophysin in hippocampal neurons developculture. J Neurosci 1991;11:1617–26.

euna S, Tos P, Guglielmone R, Battiston B, Giacobini-RobecchiMethodological issues in size estimation of myelinated nerve fiin peripheral nerves. Anat Embryol (Berl) 2001;204:1–10.

rigoryan AM, Dougherty ER, Kononen J, Bubendorf L, HostetteKallioniemi O. Morphological spot counting from stacked imagesautomated analysis of gene copy numbers by fluorescence ihybridization. J Biomed Opt 2002;7:109–22.

arata N, Ryan TA, Smith SJ, Buchanan JA, Tsien RW. Visualirecycling synaptic vesicles in hippocampal neurons by FM 1-43toconversion. Proc Natl Acad Sci USA 2001;98:12748–53.

associated protein-2 (MAP2) epitopes present in Alzheimer plogical structures do not recognize paired helical filaments.Neuropathol (Berl) 1992;83:179–89.

ternberg SR. Grayscale morphology. Comp Vision Graphic Inf1986;29:377–93.

udhof TC, Czernik AJ, Kao HT, Takei K, Johnston PA, Horiuchiet al. Synapsins: mosaics of shared and individual domainsfamily of synaptic vesicle phosphoproteins. Science 1989;245:180.

arsa L, Goda Y. Synaptophysin regulates activity-dependent synapmation in cultured hippocampal neurons. Proc Natl Acad Sci2002;99:1012–6.

heera-Umpon N, Dougherty ER, Gader PD. Non-homothetic grametric mixing theory with application to blood cell counting. PRecogn 2001;34:2547–60.