J Mater Sci: Mater Med (2010) 21:3109–3118
DOI 10.1007/s10856-010-4163-9
Image processing and fractal box counting: user-assisted method
for multi-scale porous scaffold characterization
Vincenzo Guarino • Angela Guaccio
Paolo A. Netti • Luigi Ambrosio
•
Received: 18 March 2010 / Accepted: 21 September 2010 / Published online: 5 October 2010
Ó Springer Science+Business Media, LLC 2010
Abstract Image analysis has gained new effort in the
scientific community due to the chance of investigating
morphological properties of three dimensional structures
starting from their bi-dimensional gray-scale representation. Such ability makes it particularly interesting for tissue
engineering (TE) purposes. Indeed, the capability of
obtaining and interpreting images of tissue scaffolds,
extracting morphological and structural information, is
essential to the characterization and design of engineered
porous systems. In this work, the traditional image analysis
approach has been coupled with a probabilistic based
percolation method to outline a general procedure for
analysing tissue scaffold SEM micrographs. To this aim a
case study constituted by PCL multi-scaled porous scaffolds was adopted. Moreover, the resulting data were
compared with the outputs of conventionally used techniques, such as mercury intrusion porosimetry. Results
indicate that image processing methods well fit the porosity
features of PCL scaffolds, overcoming the limits of the
more invasive porosimetry techniques. Also the cut off
resolution of such IP methods was discussed. Moreover,
the fractal dimension of percolating clusters, within the
pore populations, was addressed as a good indication of the
interconnection degree of PCL bi-modal scaffolds. Such
findings represent (i) the bases for a novel approach complementary to the conventional experimental procedure
V. Guarino (&) L. Ambrosio
Institute for Composite and Biomedical Materials (IMCB-CNR),
P.le Tecchio 80, Naples, Italy
e-mail: vguarino@unina.it
A. Guaccio (&) P. A. Netti
Interdisciplinary Research Centre on Biomaterials CRIB and
Italian Institute of Technology IIT, P.le Tecchio 80, Naples, Italy
e-mail: angela.guaccio@unina.it
used for the morphological analysis of TE scaffolds, in
particular offering a valid method for the analysis of soft
materials (i.e., gels); also (ii) providing a new perspective
for further studies integrating to the structural and morphological data, fluid-dynamics and transport properties
modelling.
1 Introduction
Image Processing (IP) has been developed in the past
decades as a challenging approach to a number of applications in a wide variety of fields [1]. The main reason is
that, today, visual information instruments such as high
computational power microcomputer, digital cameras,
image digitizer and image processing tools are often low
cost and available consumer goods. Typically, IP deals
with two dimensional discrete mapping used as representation of three-dimensional real objects. Each point holds a
number of information in terms of intensity, amplitude, and
grey level colour, carrying, in turn, morphological and
structural knowledge. Therefore, the IP process takes
image or image sequence as input and produces as output a
set of parameters and function related to the content of
images. The central step to get from the input to the output
is constituted by a series of operators or processing techniques. These concern the modulation of brightness and
contrast, binary operations involving 8-bit images, complex processes in the Fourier space and methods for
quantitatively measuring image contents. For example,
bulk properties of three dimensional samples, were determined by extracting correlation function from the images
and applying them to statistical analyses of the sample [2].
Thus far, it was possible to calculate porosity, chord length
distribution, pore size distribution and coarseness, as well
123
3110
as pore connectivity degree, and to correlate these to
physical properties of fluid flow dynamics [2–4]. Recently,
IP methods have been adopted to characterize the structure
and the bulk properties of scaffolds for Tissue Engineering
(TE) applications [5].
The more common approaches are based on the use of
X-ray based microtomography, which allows the reproduction of the complex three dimensional trabecular
architecture of TE scaffolds [6–15], scanning electron
microscopy (SEM) [16–19] and confocal laser scanning
microscopy [18]. In particular, scanning electron microscopy (SEM) analysis offers a qualitative evaluation of
hierarchically complex porous scaffolds providing a direct
measurements of strut/wall thickness and a visual estimation of pore size, interconnectivity, cross-section area and
anisotropy [20]. Once collected the micrographs, image
analysis efficiency is highly dependent on the image contrast, image resolution and analysis methods used to segment the image [15, 21]. In particular, image segmentation
results a crucial point in the outcome of structural analysis.
Indeed, the main problem is the partitioning of an image
into meaningful pieces or, alternatively, the confinement of
regions of interest (i.e. pores and/or trabeculae). The ideal
case consists of two distinct intensity distributions, the one
for the pores and the other for the polymer continuous
phase. However, the more complex is the structure the less
defined are the two distributions [22], and the more challenging is the application of robust and efficient algorithm
for image segmentation. Thus far, up to date, it has not yet
been developed any image processing technique to provide
accurate image segmentation not affected by image variability. To help with this, several researchers have proposed mechanisms for validating the effectiveness of
various segmentation algorithms [23, 24]. One strategy
consists in quantifying the general behaviour of the algorithm relative to an ideal reference model, which is, usually, hand contouring one slice at time by expert observers
(manual segmentation) [25]. Thus a study may use handcontouring results to asses the performance of a segmentation technique. This paper takes such approach, and
provides general shape metrics to compare results of an
edge-based algorithm for segmentation, implemented on
porous Poly e-caprolactone scaffolds by a self-implemented Matlab code (Matlab v.7.0), with the computeraided manual segmentation (NIH ImageJ v.3.7). Once
segmented, scaffold images may be broken up in their
partition of voids and strut, and scaffold characteristic
features may be extracted. Among the others, the interconnection degree (ID) of scaffolds represents a key
parameter in the definition of scaffold architecture functionality for tissue engineering. Indeed, the spatial and
topological arrangement of pores within the structures is
fundamental to provide fluid flow, i.e. nutrient supply and
123
J Mater Sci: Mater Med (2010) 21:3109–3118
waste removal, and tissue in-growth [26]. In particular, an
open interconnected porous network enhances the
exchange of nutrients and waste materials with the environment [26–28], while a lack of pore interconnection may
hinder cell proliferation and even promote cell death [29].
From the image analysis point of view, the existence of
interconnections implies the presence of percolative pathways made of black-void pixels within the white polymer
matrix. Here, it has been proposed the fractal dimension of
such percolating patterns as an indication for scaffold ID.
The aim of the work is, therefore, to provide a schematic
representation of an IP algorithm for the complete characterization of polymeric structures as scaffold for TE
applications. To this aim, in the following sections it has
been described the methods adopted for scaffold preparation, and data collection and analysis; further, the results of
these studies have been critically discussed in comparison
with conventionally used experimental techniques, and the
implication for future tests have been summarized.
2 Materials and methods
2.1 Scaffold fabrication
PCL scaffolds have been prepared by thermally induced
phase separation and salt leaching technique. Briefly, Poly
e-caprolactone pellets (Sigma-Aldrich MW 65 kDa) were
dissolved in absolute Dioxane (Sigma-Aldrich, purity
99.5%) by stirring for about 3 h at room temperature up to
form a homogeneous solution. Three different polymer/
solvent weight ratio were selected, 05/95, 15/85 and 30/70
(wt/wt), respectively. NaCl particles, sieved into a size
ranges of 300–500 lm, were mixed to the polymer solution
by using volume ratio of PCL/NaCl, 09/91 (v/v). The
mixture was placed into Teflon mould and cylindrical disks
(14 mm in diameter, 3 mm in thickness) were carried out
by fast quenching down to (-18 ± 1)°C. Solvent removal
was allowed in pure ethanol (C2H5OH) bath at atmospheric
pressure and -18°C for 48 h. Ethanol solution was changed daily for a more efficacious dioxane extraction. The
complete elimination of the volatile components (ethanol
and dioxane residuals), by evaporation, was let under
chemical hood for 3 h. Finally, all specimens were placed
in bi-distilled water for 7 days to leach out sodium chloride
crystals and any other contaminants. The complete evaporation of the adsorbed water was assured by keeping
samples under hood for about 3 h.
2.2 Scaffold imaging
Scanning Electron Microscopy (SEM, Stereoscan 440,
Leica Oxford, UK) was used to capture image of scaffolds
J Mater Sci: Mater Med (2010) 21:3109–3118
along their longitudinal and cross sections. A scalpel blade
was used to cut the scaffolds along the vertical axis of
cylinders. Each sample was mounted on an aluminium
stub, by using an adhesive carbon tab, and sputter coated
with gold, under vacuum, before imaging. Images were
digitized on a matrix of 1,024 * 1,024 pixels with 256 grey
levels in the tagged image file format (TIFF).
2.3 Image processing
Image analysis procedure has been reported step by step in
the schematic representation of Fig. 1, in agreement with the
ASTM F2603-06 (Standard Guide 2007). Primarily, image
selection was performed based on the criteria of image
similarity in terms of magnification, brightness and contrast,
information reported on the micrograph raw data. Three
images of the same polymer scaffold were chosen for each
magnification. Once captured an image can be manipulated
as it follows. First scaffold images were converted to
monochrome that spans a grey scale ranging from near black
to near white (pre-elaboration). The conversion of images
into binary form (8-bit conversion) reduces the calculation
complexity. In detail, the ‘‘image binarization’’ will output a
monochromatic image with 256 levels composed of pure
black pixels (level 0) and pure white pixels (level 1) which
cover the 256 grey level of the original image. With this
transformation the void and the solid component of the
samples, respectively level 0 and level 1 of the binary
images, result represented by the values at the opposite ends
of the greyscale. Intermediate shades of grey represent the
transition in brightness that occurs as the solid component
terminates into pores. The next step consists of the establishment of a criterion for segmentation of images, by distinguishing between solid (wall) and voids (pore) phases,
respectively. In this phase, filtering procedures by image
smoothing and sharpening, the adjustment of contrast/
brightness and the adequate definition of the grey threshold
level was needful to evidence the pores respect to the
polymer matrix and eliminate artefacts which could
3111
negatively affect next calculations. At this time the procedure split into two direction.
On one side, the individuation of voids was performed
manually (hand contouring), even if the procedure presented some difficulties due to the complexity of structures
that hindered the proper definition of pore boundaries.
Afterwards, the images were analysed by the use of ImageJ
software and porosity degree, pore size and spatial distribution were derived by using the ‘‘particle analysis’’ tools
contained into ImageJ software pack: this provides the
tracing of the entire inter-phase boundaries related to the
pore walls, and then, the counting of all pore object, calculating quantitatively their size and their surface area. The
porosity degree was evaluated from the total surface area of
counted pores whereas the pore sizes were derived.
In the Matlab routine the threshold level was set, by
using the criteria of edge detection (ED), on a single
greyscale value, after stepping it to investigate upon the
threshold value effect on porosity calculation [22, 26]. This
approach was used for separating touching objects without
a subjective threshold, i.e. depending upon a visual
inspection of the investigator [8]. For the evaluation of
micro and macropores, the analyses were performed by
selecting an adequate size range of the object to count
(region-based approach), according to the characteristic
pore sizes reported in other works [30, 31]. Specifically, the
analysis was based on the assumption that a surface of
1,000 lm2 defines the boundary line between micropores
and macropores. In both cases, it might be uncertainty in
the measured values due to ambiguities in the edge object
recognition, homogeneity in the size and shape of pores,
and their orientation with respect of the images. The values
calculated on three different samples have been reported as
mean ± standard deviations (SD).
2.4 Mercury intrusion porosimetry
Pore features were also determined by mercury intrusion
porosimetry (Thermo Electron Pascal 140–240) according
Fig. 1 Schematic flow chart of the image processing algorithm
123
3112
to the Washburn equation [19] which allows to correlate:
the applied pressure on the intrusion fluid, the pore radius,
the surface tension of the mercury and the contact angle
between mercury and polymer. In particular, a mercury
surface tension of 480 mN/m and a contact angle of 141.3°
were imposed, while pmax equal to 400 Pa and 200 kPa,
respectively, was applied in conjunction to pore size
analysis. This method yields the total porosity, expressed
as a percentage value, simultaneously to the average pore
diameter. As a consequence, the elaboration of data in the
light of the Washburn equation allows to build the pore size
distribution.
2.5 Estimation of pore interconnection degree
Interconnection degree of scaffolds was evaluated by calculating the fractal dimension, D, of percolating clusters
through the use of freeware ImageJ plugin based on the box
counting method. In the simplest terms, the routine counts
the number of boxes of a given size needed to cover a one
pixel wide, binary (black on white) border. The procedure
is repeated for boxes that are 2–64 pixels wide. The output
consists of two columns representing the size and the boxes
number (count) needful to cover the borders. A plot is
generated with the log of size on the x axis and the log of
count on the y axis, while the data are fitted with a straight
line. Line slope (S) represents the fractal dimension (D)
[32].
J Mater Sci: Mater Med (2010) 21:3109–3118
3 Result and discussions
SEM images of PCL scaffolds with different polymer/
solvent ratio have been analysed by IP techniques, to
derive morphological and structural information (Fig. 2a,
c). Visual inspections of images allow the identification of
the porous and polymeric phases, as well as, of a bimodal
porous structure where macropores, with irregular shape,
coexist with micropores, located along the trabeculae. A
remarkable reduction of micropores occurs with increasing
the polymer concentration, because of a progressive
reduction of capability of the polymer system to de-mixing
during the phase separation [33]. Image segmentation is
essential to discriminate between the void and the polymeric phase. The complexity of structures is well represented by the intensity histograms of the original SEM
micrographs relative to 5/95 and 30/70 (wt/wt) polymer
concentrations, respectively (Fig. 2b, d). As expected,
histograms of real images are characterized by two distribution, of voids and strut, merging in an intermediate gray
shade, between the two peaks of black (voids) and white
(strut). The more effective is the segmentation algorithm
adopted, the more distinct would be the two intensity distributions [22]. Noteworthy, SEM images show highly
asymmetric intensity distribution with an evident peak at
zero, corresponding to the black pixels, and the second
peak corresponding to the trabecular structure, more evident as the volume of solvent increases. Such findings may
Fig. 2 Segmentation based on threshold selection: selected images of PCL scaffolds with PCL/dioxane weight ratio at 05/95 (wt/wt) (a) and
30/70 (wt/wt) (c) and relative histograms of intensity (b, d)
123
J Mater Sci: Mater Med (2010) 21:3109–3118
be addressed to the presence of visible micropores affecting the strut wall, in the 05/95 (wt/wt) scaffolds, which
results in more frequent boundary regions, associated to
gray scale pixels, respect to the white peak of continuous
polymer matrix. Conversely, in the 70/30 (wt/wt) scaffolds,
white regions, related to scaffold struts in the images, are
well identified and the separation between the two phases is
more evident.
Before segmentation, all datasets have been processed to
reduce noise and smooth homogeneous regions. The
smoothing pre-processing step reduces the time necessary
to perform the segmentation by decreasing sensitivity to
small scale features and texture, also removing the noise
and enhancing edges. Binary images derived by this preelaboration step have been reported in Fig. 3a–d. Finally,
to assist in the identification of pore boundaries, an
assigned limiting dimension for macro- and micropores has
been defined for each sample, addressing to macropores as
pores of about 30 lm average diameter, while to micropores as pores of few microns in size.
Figure 4 is a comparison of the results of a typical
ED-based segmentation (Matlab algorithm) with a typical
manual segmentation, for each porous population. The
surface of the respective experimental and manual masks
are rendered in figure a–d and e–f. Images show that the
elaboration failed in areas where there was not enough
colour contrast with surrounding structure for the algorithm
3113
to detect an edge (Fig. 4a, e). Such weak edges, however,
were also difficult to visualize for manual segmentation.
This yields thinner trabecular segments and less defined
pore shapes within the macropore population (Fig. 4a, b
and e, f). In Table 1, the average pore dimensions and the
porosity degree, carried out by the using of the proposed
segmentation methods, have been reported. Results show
that there is not any significant difference about macropore
characteristics, strictly related only to the porogen used for
both scaffold composition. Furthermore, the two methods
demonstrated a good agreement with respect to the calculated values (SD = ±0.1–1.3). In the micro-pores population, as expected, the porosity percentage is varying with
polymer concentration, with a reduction of about 65% from
the 05/95 (wt/wt) to the 30/70 (wt/wt) PCL scaffolds.
Besides, a multi-population of pores is strongly desired
in order to assure nutrient transport and waste removal
within the scaffold, but even cell growth and migration,
processes favoured by a higher surface/volume ratio [34].
Concerning the effect of polymer/solvent ratio on the
average pore size, data indicate that the macro-pores population is constituted by greater pores for the 05/95 (wt/wt)
respect to the 30/70 (wt/wt) scaffolds—an increment of
approximately 17% of the mean value—with the good
agreement of the manual segmentation. Such results may
be explained with the effect of segregation between salt
crystals allowed by the less viscous solution [35]. Finally,
Fig. 3 Image segmentation based on threshold selection of PCL scaffolds with PCL/dioxane weight ratio at 05/95 (wt/wt) (a, c) and 30/70
(wt/wt) (b, d): 8-bit images by Matlab code (a, b) and Image J (c, d) manipulation
123
3114
J Mater Sci: Mater Med (2010) 21:3109–3118
Fig. 4 Pore analysis on PCL scaffolds with PCL/dioxane weight ratio at 05/95 (wt/wt) and 30/70 (wt/wt): boundary definition and pore counting
of macropores and micropores by Matlab (a–d) and Image J (e–h) procedure
Table 1 Summary of porosity and pore size of micropores
(\1,000 lm2) and macropores ([1,000 lm2) obtained by user-assisted image processing
PCL/dioxane
weight ratio
Porosity (%)
Macro
05/95
30/70
71.2 ± 0.5
71.5 ± 1.3
Average pore
size (lm)
Micro
6.46 ± 0.8
1.91 ± 0.1
Macro
190 ± 8
157 ± 17
Micro
12 ± 2
11 ± 4
the high standard deviation for micropores probably
depends on the image resolution, for that micro-pores have
sizes comparable with the pixels of images and this may
determine a source of error. It is worth to note that in many
cases the manual segmentation extends beyond object
boundaries indicated by sharp changes in colour or contrast. This trend explains much of the discrepancy between
the manual and Matlab-ED segmentation, but it also brings
into question the usefulness of these manual segmentation
as ground truth. Thus far, two main points need to be
discussed, if IP processing might represent a method for
overcoming the limits of traditionally well known techniques, and if these resulting data may be extended to the
3D pore volume properties.
Traditionally, scaffold properties have been measured
essentially by indirect methods, based on the evaluation of
volume capacity of pores (mercury intrusion porosimetry,
MI) or weight loss from the bulk to the porous constructs
(i.e., gravimetric methods) [36]. However the most diffused
method of fluid intrusion porosimetry for studying porous
materials [37], fails for soft scaffolds such as flexible foams
(with porosities higher than 90%) due to the application of
high pressures (up to 400 MPa). This may provoke the pore
structure collapse which compromises the entire scaffold
architecture and therefore the measurements [36]. Moreover, this technique underestimates porosity by neglecting
123
Table 2 Comparison of porosity and pore size data calculated by IP
and MI techniques
Measure
Method
PCL/dioxane
weight ratio
Porosity (micro
and macro) (%)
Average pore size
(micro/macro)
(lm)
IP
15/85
76.65 (71.86 ? 4.79)
(10.6/180.9)
MI
15/85
77.58
121.63
closed pores, since the fluid does not intrude into them,
preventing a quantitative estimation of further basic
structural parameters, such as pore interconnectivity, strut/
wall thickness and pore anisotropy. Therefore, the adoption
of optical techniques able to compute digitally images
integrated with analytical procedures for the definition of
morphological cues is a promising alternative approach. In
Table 2 it has been reported the comparison between IP
and MI results, for calculation of porosity features of PCL
scaffolds. While IP methods allow the definition of a
macro- and a micro-pore separated distributions (Fig. 5a,
b), the MI porosimetry measurements individuate a wider
porosity distribution characterized by the presence of several peaks (Fig. 5c) associated to the intruded macroporosities, with an average value intermediate between the
media of micro and of macropores, derived by IP measurements, and a lower average pore size (Table 2). This is
ascribable to the partial collapse of pore architecture during
the mercury intrusion under vacuum as just suggested in
other works [16]. Conversely, no relevant mismatch has
been evidenced in terms of total porosity degree. Noteworthy, the analysis has been conducted only in the case
of scaffolds with intermediate microporous values,
corresponding to 15/85 (wt/wt) PCL/dioxane ratio, as
compromise solution between scaffold softness and pore
interconnectivity.
On the other hand, while MI represents a method of
investigation of a 3D structure, image analysis techniques
J Mater Sci: Mater Med (2010) 21:3109–3118
3115
Fig. 5 IP analysis of PCL scaffolds with 15/85 PCL/dioxane ratio (wt/wt): micropore (a) and macropore (b) surface distribution curves.
c Overlapping of pore surface and pore volume cumulative distribution by IP and MI techniques, respectively
are based on two dimensional cross section analysis. Such
limit may explain the reason why IP processing technique
may be considered not convincing as instruments for
quantitative evaluation of scaffold features, among the
scientific community. To overcome such limits, up to date,
the three dimensional structures, have been investigated
only by a computer aided reconstruction mainly based on
l-CT experiments [38]. Nonetheless, predicting 3D pore
features from 2D data has been addressed from many
researcher for sand stones [39]. Two basic underlying
assumption in these studies are (i) the structure of the sand
stone is homogeneous and isotropic, and (ii) the 2D image
covers an area large enough to avoid that the microscopic
variability dominate the prediction. In other words, for
regions larger enough than the ‘‘representative elementary
region’’ or region of interest (ROI), average properties are
independent of the area—properties of self-similarity—and
reflects the values for the bulk volume [40]. For sand stone,
it has been observed that the length of four grain diameters
is the length scale required to obtain a stable asymptotic
value for porosity properties and is the same minimum
length scale for porosity determined for glass beads [41].
The image size required to reach a stable, asymptotic
value, however may not be the same for porosity, permeability and tortuosity factor, i.e. the region whose properties may be extended to the bulk volume depends on the
property measured [42]. Based upon these findings, in this
study, a ROI with size of 300 9 300 lm was selected as
representative of the morphological properties of PCL
scaffolds. A proper choice of the ROI of 2D images allows
the extension to the 3D structure of the features calculated
herein. In particular it is interesting to evaluate the interconnection degree of scaffolds by starting from their
bi-dimensional representation.
In order to overcome the relevant limitation of traditional
approaches on the quantification of the interconnectivity
123
3116
J Mater Sci: Mater Med (2010) 21:3109–3118
degree, recently, it has been demonstrated that percolation
model may be successfully used to extract additive information about pore interconnectivity from 2D images of 3D
scaffolds [43]. Some applications of percolation model to
image processing have been attempted previously [44, 45].
However, they were limited to determine the parameters of
an existing image processing method, or to detect global
fluid flow in an image [46]. The percolation is a physical
model based on the natural phenomenon of the permeation
of liquid into porous networks [46]. The basic concept is
that, once identified a starting point for the penetration of
fluid within a porous matrix, it is possible to estimate the
most probable pathway for fluid flow. The process starts by
assigning a probability value P to each point close to the
first, based on a specific characteristic (pore dimension,
morphology of pore boundaries), and identifying as second
the point with the highest probability among the neighbourhood. The assigned probability P is a measure of the
‘‘ease of percolation’’ where points which exhibit the
largest value of P describe the percolation path. Noteworthy, the percolation pathway is always constituted by
interconnected pores. To implement the percolation theory
in image processing, the probability P is assigned as a
function of the brightness range in scaffold images, so that
only the nearest pixels with brightness ranging in a defined
interval have the higher probability to be percolated and to
form clusters. Thus far, it is possible to associate the fractal
dimension of percolating cluster to the interconnection
degree of a structure [47]. Indeed, the closer is the fractal
dimension to the geometric dimension of the object under
Fig. 6 Evaluation of the pore structure interconnectivity of PCL
scaffolds by the calculation of cluster fractal dimension: scheme of
box-counting method on representative image (a), estimation of
percolation cluster size of b micro-pore network, c macropore
network and d bi-modal pore network
123
J Mater Sci: Mater Med (2010) 21:3109–3118
investigation (i.e., one for a line, two for a plane, etc.), the
higher is the interconnection degree of the structure [48,
49]. The procedure adopted herein has been schematically
reported (Fig. 6a). IP for fractal dimension calculation was
performed on our PCL scaffolds by using ImageJ freeware
plug-in, adopting the box counting method, as previously
described. The slope of curve depicted in Figs. 6b–d is the
fractal dimension of percolating clusters. It is worth to note
that it is expected a fractal ranging from 1 to 2, for that the
original object is a planar image. Moreover, since two
porous populations have been depicted, their fractal
dimensions were calculated separately. Curves reported in
Fig. 6b represent the application of box counting method to
the population of macropores in PCL 5/95 and 30/70
(wt/wt) respectively. As expected, the fractal dimensions of
the two percolative clusters is almost the same (slope
1.27–1.32) and the curve are nearly parallel. Indeed, since
the macropore percentage is about the same in both scaffolds, and their distribution is absolutely random, there is
the same probability that a pore belongs to the percolating
cluster.
Instead, the application of box counting method to
micropore population (Fig. 6c) shows a decrease in cluster
fractal dimension as a function of the curve slope passing
from 1.39 to 1.07, in line with micropore fraction variation
highlighted by previous quantitative image calculations.
Finally, similar elaborations on more complex structures
obtained by the merging of isolated micro- and macro-pore
images which better represent the structural complexity of
the effective three-dimensional scaffold, shows a fractal
cluster with dimensions varying from 1.76 to 1.33 as
micropore volume fraction decreases (Fig. 6d). All these
data concur to indicate that the micro-pores population
might control the ID of scaffolds, with the consequent
implications in design and preparation of polymeric
constructs.
4 Conclusions
In this paper it has been proposed a schematic procedure
for the analysis of morphological features of porous scaffolds by the application of image processing and fractal
box counting to 2D micrographs extracted from the 3D
structures. The major advantages of such technique and the
eventual drawbacks due to the eventual limits in image
resolution have been explored. Moreover, IP approach was
implemented on PCL scaffold by using a self developed
Matlab routine, for image segmentation, based on edge
detection method. Results have been compared to the
manual threshold to validate the algorithm used. Two pore
populations with proper characteristics, in terms of pore
size and porosity degree, have been individuated, also
3117
overcoming the problem of measurement artefacts in softmatter based scaffolds, of traditionally adopted mercury
intrusion techniques. Finally, it has been verified that the
estimation of fractal dimension of percolating clusters may
result a promising method for evaluating the interconnection degree of complex multi-scaled porous networks. In
perspective, these data pave the way for a probabilistic
estimation of pore morphological features, which will
constitute a powerful tool for the definition of design criteria of scaffolds for tissue engineering.
Acknowledgments This study was financially supported by the IP
STEPS EC project, FP6-500465. The authors wish to acknowledge
the support obtained from Italian Ministry of University and Research
(TISSUENET) for this research. Moreover, they would also like to
thank Mr. Maurizio Cotugno for its collaboration during its bachelor
master thesis spent in the Institute of Composite and Biomedical
Materials and Mr. Paolo Carboni for its support in the image
reconstruction.
References
1. Costa MFM. Application of image processing to the characterization of nanostructures. Rev Adv Mater Sci. 2004;6:12–20.
2. Coker DA, Torquato S, Dunsmuir JH. Morphology and physical
rpoperties of Fontainebleu sand stone via a tomographic analysis.
J Geophys Res. 1996;101:17497–506.
3. Lindquist WB, Lee SM, Coker DA, Jones KW, Spanne P. Medial
axis analysis of void structure in three dimensional tomographic
images of porous media. J Geophys Res. 1996;101(B4):8297–310.
4. Proussevitch AA, Sahagien DL. Recognition and separation of
discrete objects within complex 3D voxalized structures. Comput
Geosci. 2001;27(4):441–54.
5. Jones AC, Milthorpe B, Averdunk H, Limaye A, Senden TJ,
Sakellariou A, Sheppard AP, Sok RM, Knackstedt MA, Brandwood A, Rohner D, Hutmacher DW. Analysis of D bone
ingrowth into polymer scaffolds via micro-computed tomography
imaging. Biomaterials. 2004;25(20):4947–54.
6. Moore MJ, Jabbari E, Ritman EL, Lu L, Cumer BL, Widebank
AJ, Yoszemki MJ. Quantitative analysis of interconnectivity of
porous biodegradable scaffolds with microcomputed tomography.
J Biomed Mater Res. 2004;71A:258–67.
7. Cartmel S, Huynh K, Lin A, Nagaraja S, Guldberg R. Quantitative
microcomputed tomography of mineralization within threedimensional scaffolds in vitro. J Biomed Mater Res. 2004;69A(1):
97–104.
8. Brumke O, Odenbach S, Beckman F. Quantitative methods for
the analysis of syncroton- CT data sets of metallic foams. Eur
Phys J Appl Phys. 2005;29(1):73–82.
9. Cooper DML, Turinky AL, Sensen CW, Hellgrimsson B.
Quantitative 3D analysis of the canal network in cortical bone by
micro-computed tomography. Anat Rec. 2003;274B:169–79.
10. Darling AL, Sun W. 3D microtomographic characterization of
precision extruded polycaprolactone scaffolds. J Biomed Mater
Res. 2004;70B:311–7.
11. Otsuki B, Takemoto M, Fujibayashi S, Neo M, Kokubo T, Nakamura T. Pore throat size and connectivity determine bone and
tissue ingrowth into porous implants: three-dimensional microCT based structural analysis of porous bioactive titanium
implants. Biomaterials. 2006;27:5892–900.
123
3118
12. Rüegsegger P, Koller B, Muller R. A microtomographic system
for the nanostructure evaluation of bone architecture. Calcif
Tissue Int. 1996;58(1):24–9.
13. Hildebrandt T, Rüegsegger P. Quantification of bone microarchitecture with the Structure Model Index. Comput Methods
Biomech Biomed Eng. 1997;1:15–23.
14. Van Lenthe GH, Hagenmüller H, Bohner M, Hollister SJ, Meinel
L, Müller R. Nondestructive micro-computed tomography for
biological imaging and quantification of scaffold–bone interaction in vivo. Biomaterials. 2007;28(15):2479–90.
15. Rajagopalan S, Lu L, Yaszemski MJ, Robb RA. Optimal segmentation of microcomputed tomographic images of porous tissue-engineering scaffolds. J Biomed Mater Res A. 2005;75:
877–87.
16. Oh SH, Park IK, Kim JM, Lee JH. In vitro and in vivo characteristics of PCL scaffolds with pore size gradient fabrication by a
centrifugation method. Biomaterials. 2007;28:1404–13.
17. Moquet V, Bòacher S, Pirard R, Pirard JP, Jerome R. Characterization of porous polylactide foams by image analysis and
impedance spectroscopy. Langmuir. 2000;16:10463–70.
18. Grant PV, Voz CM, Tomlins PE, Mikvalovska L, Jones S,
Mikalovsky S, Vadgama P. Surface chemistry in biomedical and
environmental science part 2. The Netherlands: Springer; 2006.
pp. 215–28.
19. Winslow DN. Advances in experimental techniques for mercury
intrusion porosimetry. In: Matjevic E, Good RJ, editors. Surface
and colloid science. New York: Plenum Press; 1984. p. 259–82.
20. Zein I, Hutmacher DW, Tan KC, Teoh SH. Fused deposition
modeling of novel scaffold architectures for tissue engineering
applications. Biomaterials. 2002;23(4):169–85.
21. Li CT. Multiresolution image segmentation integrating Gibbs
sampler and region merging algorithm. Signal Process. 2003;
83(1):67–78.
22. Mather ML, Morgan SP, White LJ, Tai H, Kockenberg W, Crowe
JA. Image-based characterization of foamed polymeric tissue
scaffolds. Biomed Mater. 2008;3:1–11.
23. Malpica N, Solorzano C, Vaquero J, Santos A, Vallcorba I,
Garcia-Sagredo J, F Pozo. A multichannel watershed-based
algorithm for supervise texture segmentation. Cytometry. 1997;
28:289–97.
24. Sijbers J, Verhoye M, Scheunders A, Van der Linden A, Van Dyck
D, Raman E. Watershed-based segmentation of 3D MR data for
volume quantification. Magn Reson Imaging. 1997;15(6):679–88.
25. Cates JE, Whitaker RT, Jones GM. Case study: an evaluation of
user-assisted hierarchical watershed segmentation. Med Image
Anal. 2005;9(6):566–78.
26. Sezgin M, Sankur B. Survey over image thresholding techniques
and quantitative performance evaluation. J Electron Imaging.
2004;13(1):146–65.
27. Whang K, Thomas CK, Nuber G, Heoly KE. A novel method to
fabricate bioresorbable scaffolds. Polymers. 1995;36:837.
28. Heoly KE, Tsai D, Kim JE. Osteogenic cell attachment to
degradable polymers. Mater Res Soc Symp Proc. 1992;252:
109–14.
29. Murphy WL, Dennis RG, Kileny JL, Mooney DJ. Salt fusion: an
approach to improve pore interconnectivity within tissue engineering scaffolds. Tissue Eng. 2002;8(1):43–52.
30. Mathieu LM, Mueller TL, Bourban PE, Pioletti DP, Müller R,
Månson JAE. Architecture and properties of anisotropic polymer
composite scaffolds for bone tissue engineering. Biomaterials.
2006;27(6):905–16.
123
J Mater Sci: Mater Med (2010) 21:3109–3118
31. Guarino V, Taddei P, Di Foggia M, Fagnano C, Ciapetti G,
Ambrosio L. Influence of hydroxyapatite particles on in vitro
degradation behaviour of poly e-caprolactone-based composite
scaffolds. Tissue Eng. 2009;15(11):3655–68.
32. Smith TG Jr, Lange GD, Marks WB. Fractal methods and results
in cellular morphology. J Neurosci Methods. 1996;69:1123–26.
33. Schugens C, Maquet V, Grandfils C, Jerome R, Teyssie P. Biodegradable macroporous polylactide implants for cell transplantation: I. Preparation of polylactide foams by solid-liquid phase
separation. Polymer. 1996;37:1027–38.
34. Salerno A, Guarnieri D, Iannone M, Zeppetelli S, Di Maio E,
Iannace S, Netti PA. Engineered l-bimodal poly(e-caprolactone)
porous scaffold for enhanced hMSC colonization and proliferation. Acta Biomater. 2009;5(4):1082–93.
35. Guarino V, Causa F, Ambrosio L. Porosity and mechanical
properties relationship in PCL based scaffolds. J Appl Biomater
Biomech. 2007;5(3):149–57.
36. Ho ST, Hutmacher DW. A comparison of micro-CT with other
techniques used in the characterization of scaffolds. Biomaterials.
2006;27:1362–76.
37. Guarino V, Causa F, Netti PA, Ciapetti G, Pagani S, Martini D,
Baldini N, Ambrosio L. The role of hydroxyapatite as solid signals on performance of PCL porous scaffolds for bone tissue re
generation. J Biomed Mat Res Appl Biomater. 2008;86B:548–57.
38. Srinivasan R, Lu L, Yaszemski MJ, Robb RA. Optimal segmentation of microcomputed tomographic images of porous
tissue-engineering scaffolds. J Biomed Mater Res A. 2005;
75(4):877–87.
39. Reed AH, Briggs KB, Lavoie DL. Porometric properties of a
siliciclastic marine sand: a comparison of traditional laboratory
measurements with image analysis and effective medium theory
modelling. IEEE J Ocean Eng. 2002;27:581–92.
40. Ehrlich R, Kennedy SK, Crabtree SJ, Cannon RL. Petrographic
image analysis I. Analysis of reservoir pore complexes. J Sediment Pet. 1984;54:1365–78.
41. White JV, Kirkland BL, Gournay JP. Quantitative porosity
determination of thin sections using digitized images. J Sediment
Res. 1998;68:220–2.
42. Reed AH, Pandey RB, Lavoie DL. Fractal dimensionality of pore
and grain volume of a siliciclastic marine sand. Int J Mod Phys C.
2000;11:1555–9.
43. Yamaguchi T, Hashimoto S. Image processing based on percolation model. IEICE Trans Inf Syst. 2006;E89-D(7):2044–52.
44. Hussain I, Reed TR. A bond percolation-based model for image
segmentation. IEEE Trans Image Process. 1996;6(12):1698–704.
45. Craciunescu OI, Das SK, Clegg ST. Dynamic contrast-enhanced
MRI and fractal characteristics of percolation clusters in twodimensional tumor blood perfusion. Trans ASME J Biomech
Eng. 1999;121:480–6.
46. Stauffer D, Aharony A. Introduction to percolation theory. Philadelphia: Taylor and Francis; 1994.
47. Mandelbrot BB. The fractal geometry of nature. San Francisco:
CA Freeman; 1982. pp. 468–69.
48. Orford JD, Walley WB. The use of fractal dimension to quantify
the morphology of irregular shaped particles. Sedimentology.
1983;30:655–68.
49. Oh W, Lindquist WB. Image thresholding by indicator kriging.
IEEE Trans Pattern Anal Mach Intell. 1999;21(7):590–602.