An Improved Segmentation Method for In Vivo ?CT Imaging
Jan H Waarsing, Judd S Day, and Harrie Weinans
ABSTRACT: Image segmentation methods for ?CT can influence the accuracy of bone morphometry
calculations. A new automated segmentation method is introduced, and its performance is compared with
standard segmentation methods. The new method can improve the results of in vivo ?CT, where the need to
keep radiation dose low limits scan quality.
Introduction: An important topic for ?CT analysis of bone samples is the segmentation of the original reconstructed
grayscale data sets to separate bone from non-bone. Problems like noise, resolution limitations, and beam-hardening
make this a nontrivial issue. Inappropriate segmentation methods will reduce the potential power of ?CT and may
introduce bias in the architectural measurements, in particular, when new in vivo ?CT with its inherent limitations
in scan quality is used. Here we introduce a new segmentation method using local thresholds and compare its
performance to standard global segmentation methods.
Material and Methods: The local threshold method was validated by comparing the result of the segmentation with
histology. Furthermore, the effect of choosing this new method versus standard segmentation methods using global
threshold values was investigated by studying the sensitivity of these methods to signal to noise ratio and resolution.
Results: Using the new method on high-quality scans yielded accurate results and virtually no differences between
histology and the segmented data sets could be observed. When prior knowledge about the volume fraction of the
bone was available the global threshold also resulted in appropriate results. Degrading the scan quality had only
minor effects on the performance of the new segmentation method. Although global segmentation methods were not
sensitive to noise, it was not possible to segment both lower mineralized thin trabeculae and the higher mineralized
cortex correctly with the same threshold value.
Conclusion: At high resolutions, both the new local and conventional global segmentation methods gave near exact
representations of the bone structure. When scanned samples are not homogenous (e.g., thick cortices and thin
trabeculae) and when resolution is relatively low, the local segmentation method outperforms global methods. It
maximizes the potential of in vivo ?CT by giving good structural representation without the need to use longer
scanning times that would increase absorption of harmful X-ray radiation by the living tissue.
J Bone Miner Res 2004;19:1640—1650. Published online on July 12, 2004; doi: 10.1359/JBMR.040705
Key words: QCT, histomorphometry
not sufficient to understand how bone loss affects fracture
risk. It is also important to study the actual structure of the
trabecular network, for instance by ?CT.(1)Since its intro-
duction by Feldkamp et al.,(2)?CT has become an important
tool to quantify the morphometry of the trabecular structure
of bone biopsy specimens of humans and of whole bones of
small animals like rats or mice. In addition, ?CT provides
the possibility to set up finite element (FE) models to
determine the strength and stiffness of the bone sample
based solely on the trabecular architecture.(3,4)Furthermore,
the 3D representations of bone can be used as the input for
computer simulations of bone remodeling.(5)Recent ad-
ISEASES LIKE OSTEOPOROSIS can seriously affect the me-
chanical integrity of bone. Measuring the BMD alone is
vances in technology have made in vivo ?CT possible;
increasing the value of ?CT as a research tool by enabling
the design of longitudinal studies to follow changes in
A number of studies have validated the accuracy of
?CT. Comparison with histology showed relative small
to rather high deviations in 2D morphometric parameters
depending on scan resolution and especially on which
segmentation method was used.(8–11)Other studies on the
effects of scanning resolution and segmentation methods
on morphometric parameters support and stress the im-
portance of choosing an appropriate segmentation tech-
nique to separate bone from non-bone.(12–15)The de-
mands on segmentation techniques become even more
stringent when the quality of the scans is limited, as is the
case for in vivo scanning, where scanning time should be
as short as possible to limit radiation dose absorbed by
the living tissue.
The authors have no conflict of interest.
Department of Orthopaedics, Erasmus Medical Centre, Rotterdam, The Netherlands.
JOURNAL OF BONE AND MINERAL RESEARCH
Volume 19, Number 10, 2004
Published online on July 12, 2004; doi: 10.1359/JBMR.040705
© 2004 American Society for Bone and Mineral Research
The most widely used segmentation techniques use global
thresholds; a single CT number is chosen, above which all
voxels (3D pixels) are marked as bone and below which all
remaining voxels are marked as non-bone. The value that is
used as a threshold is selected either visually, by analyzing
the histogram of CT numbers or by forcing the resulting
binary data set to have the same volume as the original bone
sample as determined by doing an Archimedes test.(15)
Despite the ease and speed of using a global threshold,
serious problems such as beam hardening, noise, and partial
volume effects can considerably reduce the quality of the
segmentation. The effect of beam hardening can be reduced
by using a physical filter during scanning and by using
corrective algorithms during reconstruction. Partial volume
effects (a voxel that contains both bone and non-bone has a
lower CT number than a bone voxel, but a higher CT
number than a non-bone voxel) will limit the frequency
content of the reconstructed images. This will “smear out”
the bone around its real edge in a reconstructed data set.
Trabeculae that are thin relative to the resolution can even
be smeared out such that the CT number of the trabeculae
does not reach the CT number that would represent the true
density of the material.(16)The combination of these effects
causes the optimal threshold value for a certain part of the
reconstruction to be different from the optimal value in
other parts. In general, using a single global threshold value
will result in the loss of thin trabeculae and oversizing of
Segmentation can be improved by using local threshold
values rather than a single global threshold value such that
each voxel can be thresholded optimally within its neigh-
borhood. Dufresne(17)has developed a local threshold algo-
rithm for CT scans to compensate for beam-hardening ef-
fects based on the analysis of the histogram of the local
neighborhood of a voxel. The efficiency of this method will
decrease when the resolution is limited and tissues are not
homogenous. Kuhn et al.(16)and Elmoutaouakkil et al.(18)
proposed more general local segmentation methods, both
based on the so-called one-half maximum height (HMH)
protocol. Summarized, voxels are considered bone if their
CT number is higher than one-half the difference between
local minima (background) and local maxima (bone). The
more rigorously validated method of Kuhn et al. gives very
accurate results for structures much thicker than the resolu-
tion of the ?CT system; however, structures that are thin
relative to the resolution are still oversized.
Segmentation techniques have been studied more exten-
sively for in vivo imaging of human bone by ?MRI(19–23)
and pQCT.(24–28)Because of their relatively low resolutions
(?150 ?m), these systems lack the ability to accurately
resolve individual trabeculae. Hence, most studies have
focused on parameter extraction techniques that indirectly
estimate properties of the trabecular bone. These methods
have only been validated by assessing the reproducibility
and by showing that they could discriminate between pop-
In this paper, we introduce an automated segmentation
method using local threshold values obtained by applying
standard edge detection algorithms, while compensating for
smeared out thin trabeculae. The method was validated by
comparing segmented data sets with traditional histological
sections. Furthermore, the consequences of choosing a local
versus a global threshold method were investigated by
studying the behavior of these methods at different scan
qualities and resolutions.
MATERIALS AND METHODS
Before applying the segmentation algorithm, noise was
reduced by filtering the data sets with a 3D Gaussian
smoothing function. Each voxel in the data set was replaced
by the weighted average of the voxels in its 5 ? 5 ? 5
neighborhood. The weights were functions of the distance
of each voxel to the central voxel in the 5 ? 5 ? 5
neighborhood, according to the Gaussian function of which
the SD or radius was a parameter (GaussRadius) that was
set by the user.
A standard edge detection algorithm, extended to 3D, was
used to find the surface of the bone in the smoothed data set
(Fig. 1A). The edges were detected by calculating the 3D
spatial gradient of the reconstructed data using three orthog-
onal 3 ? 3 Sobel operators(29)that favor localization of
edges above (noise sensitive) detection of edges(30)(Figs.
1B and 2B). The ridges in the gradient field are the edges,
indicating the transition from bone to non-bone and corre-
spond to the HMH points. The ridges were detected by
finding local maxima in the direction of the gradient. To
prevent detection of false (noisy) edges and streaking
(breaking up of an edge contour because of noise), only
local maxima that were above a high threshold were kept
(strong edges), together with smaller maxima that were
connected to these high maxima but were still above a low
threshold (weak edges; Figs. 1C and 2C).(30)The values of
the high- and low-threshold values that define the strong and
weak edges were set by the user (GradientCutOffHigh,
GradientCutOffLow). The possible values for these param-
eters ranged between 1 (strongest edges) and 0 (no edge).
These values are obtained by normalizing the cumulative
histogram of the gradient values in a data set. A parameter
value of 0.9 would refer to that gradient value where the
cumulative histogram is at 90% of its maximum value.
The set of edges that resulted from the edge detection step
was used to obtain local threshold values spanning the
surface of the bone. The CT number of a voxel that is part
of the set of edges served as a local threshold value for its
neighborhood. The total set of local thresholds was obtained
by dilating the local-threshold surface in 3D iteratively until
it filled a matrix of local threshold values, giving each voxel
its own local threshold. During a dilation step, a voxel was
included into the set of local thresholds if it had neighboring
voxels in a 3 ? 3 ? 3 neighborhood that were included in
previous steps. The local threshold value for the to-be-
included voxel was calculated by taking the Gaussian
weighted average of the threshold values of the voxels in the
3 ? 3 ? 3 neighborhood. The dilation process was contin-
ued until all voxels in the data set had been appointed a local
threshold value (Fig. 2D).
The last step of the algorithm consisted of comparing the
gray value of each voxel with its local threshold value and
1641IMPROVED SEGMENTATION FOR ?CT
marking it as bone when it was higher than this value and
marking it as non-bone otherwise, resulting in a binary data
set (Figs. 1D and 2E). An extra condition was included here
to decrease the effect of smearing out trabeculae that were
thin relative to the resolution. The maximal CT numbers of
such a thin trabecula will be lower than the CT number
corresponding to the average density of the material. For
rod-shaped trabeculae, this decrease in CT number is pro-
portional to the decrease in radius.(31)By assuming all
trabeculae have approximately the same BMD, thin trabec-
ulae could be identified by comparing a local threshold
value to the average threshold value; when it was lower, a
thin trabecula was identified. To reduce the risk of false
positives, that is, classifying a thick trabecula as being thin,
a minimal difference of 1 SD between the local and the
average threshold value was required before a trabecula was
identified as being thin. The local threshold value was
adjusted upward proportional to the difference between the
original local threshold value (TL) and the average threshold
value (TL,mean), scaled by the difference between the aver-
age threshold value and the average CT number of the
TL,adjusted? TL?1 ?
The average CT number of the background was a param-
eter that was set by the user (BackgroundValue). As a
consequence of the adjustment, the resulting trabecula was
slightly thinner than when the local threshold would not
have been adjusted.
If not stated differently, the value for the radius of the
Gauss filter (GaussRadius) was determined by trial and
error, as was the GradientCutOffLow parameter. The
GaussRadius had values ranging from 0.4 for scans with
low noise content to 1.0 for scans with high noise content.
GradientCutOffHigh was held constant for all segmenta-
tions and was set to 0.99. The value of BackGroundValue,
needed for the compensation of smeared out thin trabeculae
was set equal to the average CT value of the background
material, which is air in clean and dry samples and marrow
or soft tissue in whole bones and in vivo scanning.
Decreasing the value of GradientCutOffLow resulted in
the inclusion of thinner trabeculae and eventually in the
inclusion of noise, but did not influence the thickness of the
trabeculae. The value for this parameter was chosen just
above the level where noisy structures started to seem in the
data sets. Increasing GaussRadius decreased the noise in the
data sets, but also smeared out trabeculae, which could
result in thickening of trabeculae. This value was adjusted in
conjunction with GradientCutOffLow such that the lowest
values possible for both parameters could be used.
Validation: comparison with histological sections
Two rat tibias were scanned in a Skyscan-1072 MicroCT
scanner (Skyscan, Antwerp, Belgium), yielding recon-
structed data sets with a voxel size of 11 ?m. The system
had an actual resolution of 8 ?m specified by the manufac-
mentation method are shown. From the original
(A) grayscale images, (B) the 3D spatial gradient
is calculated (the highest gradients have the
darkest color). (C) The local maxima in the
gradient field are the edges of the bone (strong
edges are black, weak edges are white). (D)
From these edges, a set of local thresholds is
derived that is used to obtain the binary image.
The different steps of the local seg-
1642 WAARSING ET AL.
turer (10%MTF). The data sets were segmented both by
using the local threshold method and by applying a global
threshold such that the resulting data sets had the same
volume fraction as the data sets obtained from the local
After scanning, the bones were embedded in methyl
methacrylate (MMA). Serial sections were obtained by re-
peatedly slicing 30-?m-thick sections off the MMA block.
After each cut, the bone at the surface of the block was
stained with alizarine red, coloring the calcium red, and a
microphotograph with a pixel size of 2 ?m was taken of the
block surface. Processing the sample en block prevented
geometrical distortion that occurs when processing thin
Registration software from the University of Leuven(32)
was used to automatically reposition the segmented data
sets such that they optimally matched the set of histological
The cross-sections of the scan data were compared with
the histological data by overlaying the segmented serial
microphotographs with the registered scan data and calcu-
lating the difference in bone area.
Comparing segmentation methods at different
To obtain data sets with different scan resolutions, but
with equal noise content, scanning was simulated using a
real scan as input. To serve as input data sets, six pieces of
trabecular bone from core biopsy specimens of canine distal
femora, embedded in MMA, were scanned in a Skyscan-
1076 ?CT scanner. The system resolution, as specified by
the manufacturer, was 15 ?m (10% MTF). The resulting
data sets had a voxel size of 18 ?m. Before embedding, the
volumes of the core biopsy specimens were measured using
To simulate scanning at different resolutions, the radon
transform of the original scan was calculated resulting in
virtual shadow projections. These shadow projections were
resampled to pixel sizes of 35 and 53 ?m by selecting a
random voxel in a 2 ? 2 ? 2 neighborhood in case of the
53-?m samples. By resampling, the noise levels remained
constant. The inverse radon transform was calculated for the
original sets of virtual shadow projections and the two sets of
Detailed flowcharts of each subprocess in the general outline are presented for (B) the calculation of the 3D gradient, (C) the edge detection step,
(D) the dilation of the original set of local thresholds, and (E) the final thresholding step, including adjustment for thin trabeculae.
Flowchart diagram that gives a schematic representation of the local segmentation algorithm. (A) General outline of the entire algorithm.
1643 IMPROVED SEGMENTATION FOR ?CT
resampled shadow projections, giving reconstructed data sets
with voxel sizes of 18, 35, and 53 ?m. The calculations were
performed using Matlab (The Mathworks).
All reconstructed data sets were segmented in three dif-
ferent ways. One set of data were obtained by applying the
automatic local threshold algorithm with adjustment for
smeared out trabeculae (LocalAuto). A second data set was
obtained by applying a global threshold value that was
chosen such that the volume of the resulting data set was
equal to the volume of the original sample as determined
using Archimedes’ principle (GlobalArch). The third set
was obtained by setting the global threshold at a value
where the sensitivity of the volume to changes in threshold
value was smallest, as determined after analysis of the
density histogram(15)(GlobalHist). The study set-up is sum-
marized in Table 1.
For the LocalAuto method, the parameters GaussRadius and
GradientCutOffLow were determined by trial and error for the
data sets with 18-?m voxel size. The GaussRadius for the data
sets with voxel sizes of 35 and 53 ?m were, respectively,
scaled to one-half and one-third of the GaussRadius of the data
sets with the highest resolution. GradientCutOffLow was ad-
justed to get the lowest acceptable value.
All samples in the same group were segmented with the
same parameter values. To ascertain that only the segmen-
tation methods themselves were compared, the Gaussian
smoothing was also applied to all data sets before global
segmentation. The same value for GaussRadius was used as
for the local threshold algorithm.
For all resulting data sets, volume fraction and 3D-direct
thickness(33)were calculated, as well as connectivity(34)and
structure model index (SMI).(35)The first two parameters
are representations of the bone mass and its distribution, and
the last two parameters characterize geometry and topology
of the bone sample. The parameters were calculated using
the freely available software of the 3D-Calculator project
(www.eur.nl/fgg/orthopaedics/Downloads.html). The ef-
fects of scanning resolution and the addition of noise on
each segmentation method were assessed by analyzing the
relative change of the morphometric parameters.
A reliable segmentation method maintains the ability to
detect group differences at decreased scan qualities. The
method should be reliable in the sense that the way samples
relate to each other when scanning at different resolutions
should not change. In other words, if sample A has a volume
fraction that is slightly higher than the volume fraction of
sample B at high resolution, it should still have a higher
volume fraction when the samples are scanned at a lower
resolution. In this paper, we will refer to this quality of
segmentation algorithms with the term “reliability,” which
was assessed by using regression analysis. If the way in
which samples relate to each other is the same at different
resolutions, the variation between samples at a low resolu-
tion is explained completely by the variation at a high
resolution, which is expressed by the R2value. R2was
assessed for each parameter and each thresholding method
by regressing the outcomes for the data sets with 35 and 53
?m on the outcomes of the 18-?m data sets.
Finally, the different segmentation methods were ana-
lyzed visually by overlaying the cross-sections for the dif-
Comparing the influence of noise on the different
The effect of noise on the segmentation methods was
investigated by adding noise to the 35-?m data sets de-
scribed in the previous section. The original 35-?m data sets
served as the low-noise reference. Data sets with medium
and high noise levels were created by adding Gaussian-
distributed noise with an SD of 0.005 (medium) and 0.01
(high) to the virtual shadow projections of the 35-?m data
After reconstruction, all data sets were segmented using
the three different methods described above (GlobalArch,
GlobalHisto, and LocalAuto).
The LocalAuto parameter settings for the low noise data
sets as determined in the previous section were taken as a
reference for the data sets with higher noise levels. Gradi-
entCutOffLow was kept constant for these data sets, whereas
GaussRadius was increased until no more noisy structures
were present in the data sets.
The segmented data sets were analyzed in the same way
as in the study of the effects of decreasing resolution.
Comparing segmentation methods for whole bone
scans at different scan qualities
The tibia of a rat was scanned in vivo using the Skyscan
1076 in vivo system. Scanning time was kept short (20
minutes) to limit radiation dose absorbed by the living
tissue, resulting in a data set with high noise levels. The
resulting data set had a voxel size of 18 ?m. After death, the
tibia of the rat was dissected and scanned in vitro in a
Skyscan 1072 ?CT system. A long scan was made (3 h) to
obtain a high signal-to-noise ratio.
Both the high-quality (in vitro) and low-quality (in vivo)
scans were segmented using both the automated local
threshold algorithm and by using the most optimal global
TABLE 1. OVERVIEW OF THE STUDY SET-UP TO COMPARE
SEGMENTATION METHODS AT DIFFERENT
RESOLUTIONS AND NOISE LEVELS
18 ?m35 ?m53 ?m
One global threshold value determined
by volume measurements
One global threshold value determined
by histogram analysis
Local thresholds obtained through
automated segmentation algorithm
n ? 6. Each sample was reconstructed at three different resolutions with
low noise levels, and at 35 ?m also with medium and high noise levels.
Each reconstructed data set was segmented with all three segmentation
1644WAARSING ET AL.
threshold value. When the global threshold was chosen too
high, no trabecular bone was included in the segmentation.
On the other hand, when the global threshold was chosen
too low, all trabecular bone was included, but cortical bone
was oversized. For a range of threshold values, the differ-
ence in volume estimates between the globally and locally
segmented data sets for cortical and trabecular bone was
calculated as a percentage of the cortical and trabecular
bone volume, respectively, in the locally segmented data
set. The minimum in total difference was chosen as the most
optimal global threshold value.
Animal procedures formed part of a larger experiment for
which approval was obtained from the Animal Ethics Com-
Comparison with histological sections
Using the local threshold method as well as applying the
global threshold, set at a threshold value that resulted in the
good representations of the bone. For both thicker and thinner
structures, the histological images were nearly identical to the
segmented cross-sections obtained with the local threshold
method (Fig. 3). Quantitative analysis of bone area showed an
average difference of ?0.5% between the corresponding lo-
cally and globally thresholded cross-sections and the histolog-
ical sections. Where the locally thresholded cross-sections
gave differences in bone area compared with the histological
sections that varied between ?0.4% and ?0.6%, the globally
thresholded cross-sections gave differences that varied be-
tween ?1% and 1%. Detailed examination showed that the
most negative differences of the globally thresholded cross-
sections were measured in sections that contained thin struc-
tures that were less well detected by the global method than by
the local method (Fig. 3).
Influence of a reduction in resolution on segmentation
When resolution of the data sets was decreased from 18 to
35 ?m, the local segmentation method (LocalAuto) gave the
least change in parameter values. Parameters changed on
average by about 7% for the local method, whereas the
global methods resulted in changes of 12% and 13%. Reli-
ability was very high for all methods, with R2values rang-
TABLE 2. RELATIVE CHANGE IN PARAMETER VALUE (%) FOR DECREASING THE RESOLUTION
18–35 ?m 18–53 ?m
GlobalArch GlobalHistLocalAuto GlobalArchGlobalHist LocalAuto
voxel size of 11 ?m with (B) histological sec-
tions with a pixel size of 2 ?m. The ?CT image
was segmented using the global threshold
method (C), set at a level where the resulting
volume fraction was equal to the volume fraction
resulting from the local method, and was seg-
mented by the local threshold method (D). Light
gray arrows indicate thin structures that were not
detected by the global method in this section.
Also notice the hole that was almost filled up in
the globally thresholded image (dark gray ar-
Comparing (A) ?CT scans with a
1645 IMPROVED SEGMENTATION FOR ?CT
ing from 0.96 to 0.98. Changes in parameter values and the
results of the reliability test are summarized in Tables 2 and
3. When resolution was further decreased to 53 ?m, the
performance of all methods deteriorated strongly, resulting
in changes of ?30%.
At high resolution, the LocalAuto method resulted in a
slightly higher value for volume fraction than the two global
methods (Fig. 4). When resolution was decreased, the vol-
ume fraction for the GlobalHist method increased rapidly
and rose above values for the LocalAuto method. For all
resolutions, the LocalAuto segmentation method gave thin-
ner trabeculae and a more plate-like structure, indicated by
SMI values close to one. In absolute values, connectivity
density did not show much difference between the segmen-
Visual inspection of the segmented data sets showed
that, in general, both global methods resulted in more
bone on the outside of the core samples, whereas the
LocalAuto method resulted in more bone at the inside of
the core sample (Fig. 5). This difference became stronger
as resolution decreased.
Influence of noise on segmentation methods
Adding noise had the least effect on the parameter values
estimated with the GlobalArch method. On average, the
values changed 6% when medium noise levels were added
and 7% when high noise levels were added. The GlobalAuto
method gave changes of 8% and 10%, respectively, for the
addition of medium and high noise, and the LocalAuto
method gave changes of 10% and 17% (Table 4). Global-
Arch gave similar results for reliability as LocalAuto, with
R2values of about 0.97 for addition of medium noise levels
and 0.94 for addition of high noise levels. GlobalHist gave
the least reliable results, with R2values of 0.87 for addition
of medium noise levels and 0.84 for addition of high noise
levels (Table 5).
The effects on the various parameters showed strong
differences between the segmentation methods. The Global-
Arch method had problems with representing connectivity:
the reliability went down to 0.82 for the highest noise level.
GlobalHist had problems with estimating the volume (BV)
and especially with SMI (R2as low as 0.56 for the addition
samples for the different methods (GlobalArch,
GlobalHist, LocalAuto) as a function of the
different resolutions (18, 35, and 53 ?m). (A)
Volume fraction. (B) Mean trabecular thick-
ness. (C) SMI. (D) Connectivity density.
Parameter values averaged over all
TABLE 3. R2VALUES OBTAINED FROM REGRESSION ANALYSIS OF HIGH RESOLUTION TO LOW RESOLUTION
18–35 ?m 18–53 ?m
GlobalArch GlobalHistLocalAuto GlobalArch GlobalHistLocalAuto
1646 WAARSING ET AL.
of the highest noise level). The LocalAuto method gave
strong increases in the absolute value for SMI and had
problems with trabecular thickness, where reliability de-
creased to 0.86 when high noise levels were added.
Comparison of the segmentation methods for whole
Applying the local threshold algorithm to the high-quality
in vitro scan of a whole rat tibia resulted in a visually
convincing segmentation. The local segmentation of the
lower-quality in vivo scan showed only minor differences
compared with the high-quality data set; metaphysial tra-
beculae were slightly thicker (Figs. 6B2 and 6C2).
For high global threshold values, the cortical bone in the
globally segmented data set resembled the cortical bone in the
locally segmented data set. However, no trabecular bone was
included in the data set. Adjusting the global threshold value
downward resulted in the inclusion of more trabecular bone,
but this was accompanied by a strong increase in cortical bone
volume. Setting the global threshold at the value where there
was no difference in total bone volume with the locally seg-
mented data set would still result in a trabecular bone volume
that was ?40% less than the trabecular bone volume resulting
from the local segmentation algorithm (Fig. 6A).
At the optimal global threshold, where the summed dif-
ference in cortical and trabecular volume between globally
and locally segmented data sets was smallest, the globally
segmented data set showed a cortex that was thicker than
the cortex of the locally segmented data set while less
trabecular bone was detected. Compared with the globally
segmented data set of the high-quality scan, the globally
segmented data set of the in vivo scan showed few differ-
ences. The trabecular structure seemed slightly more frag-
mented (Figs. 6B3 and 6C3).
Segmentation of ?CT data sets is not a trivial issue, and
its complexity is often neglected or underestimated. In this
paper, we have tried to clarify this issue by comparing
different segmentation techniques and by introducing a new
segmentation algorithm using local threshold values.
Applying this local threshold method gave very accu-
rate results. For high-quality scans, the segmented data
set was a nearly exact representation of the real bone, as
could be seen by comparison with histological sections
(Fig. 3). All trabeculae, both thick and thin, were repre-
sented with the correct thickness. The structural integrity
of the trabecular network was represented correctly as
well, and no connections between trabeculae were
missed. Applying a global threshold performed almost
equally, but some of the thinner connections between
trabeculae were lost. These connections could be recov-
ered by adjusting the global threshold value, but this
would result in thickening of the trabeculae and thus in
an overestimation of the bone volume.
The global segmentation method could only give such a
good result given prior knowledge about the real volume of
the bone, a prerequisite that is not needed for the automatic
local threshold algorithm. In many situations, information
about the volume of a bone sample is not present or is
difficult to get. When scanning core biopsy specimens, the
bone volume can only be obtained after a cumbersome
measurement procedure based on Archimedes’ principle.
For entire bones, it is even practically impossible to get the
real volume of the bone, which excludes the global segmen-
tation method based on knowledge about bone volume.
Another situation in which it is impossible to obtain the real
volume of a bone is when animals are scanned in vivo. The
added value of scanning living animals lies in the fact that the
longitudinal study design allows comparison of scans of dif-
ferent time-points of the same animal, such that temporal
changes in bone architecture can be followed at the level of
single trabeculae.(6)Because radiation load on living tissues
should be as small as possible, the quality of the scans are
limited. The resulting images contain more noise and have a
are the original reconstructed image, together with the GlobalArch and
LocalAuto segmented images laid on top of each other (white is
GlobalArch only, black is LocalAuto only, gray is both). Notice that the
globally segmented images have more bone on the outside of the circle,
whereas the images segmented with the local threshold method have
more bone on the inside.
Same cross-section at three different resolutions. Shown here
1647 IMPROVED SEGMENTATION FOR ?CT
lower resolution than in conventional ?CT. This situation
places high demands on the method of thresholding. Although
the influence of noise seems rather small on both local and
global segmentation methods, global thresholding methods fail
to give a good representation of the bone. This is partly caused
by the relative low resolution of the system that makes thin
structures appear less dense. Besides, the result of global
thresholding will also be affected by differences in mineraliza-
tion between cortical and trabecular bone (Fig. 6). Thus, in
vivo scanning is more powerful when combined with local
This study indicates that when good-quality scans are
made at high resolution and the samples have a homogenous
structure, a global threshold performs just as well as the
local threshold method. A typical situation that satisfies
these conditions is when studies are undertaken in which
bone biopsy specimens are scanned at high resolution and
compared. The convenience and speed of applying a global
threshold makes using this method very tempting. However,
the implications of the choice of global threshold value
should not be underestimated. In most studies, bone biopsy
specimens of subjects with a certain pathological condition
or biopsy specimens resulting from some intervention study
are compared with controls. The possible changes in bone
morphology and bone mineralization caused by these pa-
thologies or interventions will affect the distribution of
densities of the scans and thus might interact with the choice
of threshold value. This problem can be reduced when the
real volume of the samples are known. However, often this
volume is not known. In general, using a global threshold
might result in uncertainties about which part of the mea-
sured difference between groups are caused by the choice of
threshold value. Because the result of the local threshold
method is not influenced by changes in mineralization and is
less sensitive to changes in architecture (especially to
changes in the amount of thick versus thin trabeculae), using
this method could reduce the uncertainty about measured
differences between groups. Although the results presented
in this study supports this assumption, more tests are needed
to see if the local segmentation method is truly better at
predicting group differences.
Local segmentation of high-resolution scans resulted in
accurate representations of the volume. However, when
resolution was decreased, the local segmentation method
started to overestimate the volume of the scan. This phe-
nomenon is related to the smearing out of thin structures.
The lower the resolution, the more trabeculae are smeared
out around their real edge, and compensation starts to be-
come problematic. Global methods were also affected by
this phenomenon, resulting in a strong increase in trabecular
thickness with decreasing resolution (Fig. 4B). The average
increase in trabecular thickness resulted in extra volume
being added to the thickened trabeculae. For the GlobalArch
method, in which the volume of the data sets was fixed, this
extra volume was balanced by an equal decrease in the
volume of thinner trabeculae and thus to changes in struc-
ture. This is clearly reflected in the change of structural
parameters like connectivity density and especially SMI for
this method (Fig. 4). As indicated by the less severe changes
in SMI and connectivity, the GlobalHist method favored
structural integrity above volume. As a consequence, there
was a strong rise in trabecular thickness and especially in
volume fraction. The local segmentation combined the
strong points of both global threshold methods, because it
combined preservation of structural integrity with good
representation of bone volume.
A further decrease in resolution to a voxel size of 53 ?m
resulted in unreliable results for all segmentation methods. At
this resolution, average trabeculae have a real thickness of less
than two or three voxels, which is not sufficient for good
representation of the bone. Still the results show that the
LocalAuto method gave better volume estimates than the
GlobalHist method and a better structural integrity than the
GlobalArch method. A study by Laib and Ruegsegger(27)tried
TABLE 5. R2VALUES OBTAINED FROM REGRESSION ANALYSIS OF LOW TO HIGH NOISE LEVELS
Adding medium levels Adding high levels
GlobalArch GlobalHistLocalAuto GlobalArch GlobalHistLocalAuto
TABLE 4. RELATIVE CHANGE IN PARAMETER VALUE (%) FOR ADDING NOISE
Adding medium levelsAdding high levels
GlobalArch GlobalHistLocalAutoGlobalArch GlobalHist LocalAuto
1648WAARSING ET AL.
at much lower resolutions than were used in this study (165
?m voxel size). Their methods gave surprisingly good reliabil-
ity values for the measured parameters, comparable with the
global methods in this paper. However, the low resolution
made it impossible to extract the exact structure.
The effects of noise on the different segmentation meth-
ods could mostly be explained by the effects of the smooth-
ing filter needed to remove this noise. Filtering the noisy
data sets blurred out the smallest details and thinnest tra-
beculae. Logically, this had the greatest influence on those
methods that were most sensitive to thin structures. Blurring
the thinnest structures did not influence the GlobalArch
method, which favored the detection of thick trabeculae
above the detection of thin trabeculae, because it already
failed to detect the thin trabeculae in the low noise situation.
The LocalAuto method on the other hand, detected both
thick and thin trabeculae and was therefore very sensitive to
blurring of the thin trabeculae. As a result, the thickness of
the blurred thin trabeculae was overestimated, and the av-
erage thickness increased. The thickness measure was af-
fected less for samples with mainly thick trabeculae than for
samples with both thin and thick trabeculae, explaining the
decrease in reliability of the trabecular thickness measure
for the LocalAuto method when high noise levels were
added to the samples (Table 5).
Overlaying the globally thresholded data sets with the
data sets obtained with the local threshold algorithm
showed an interesting phenomenon (Fig. 5). The local
segmentation method detected more bone in the center of
the bone sample, whereas the global method resulted in
thicker structures on the outside of the bone sample.
Imperfect beam hardening correction in a reconstructed
data set generally causes the outside of a cylindrical
object to appear denser than the inside. Using a global
threshold, therefore, resulted in overestimation of the
thickness of the trabeculae in the outer section of a bone
sample, whereas it underestimated the thickness on the
inside of the sample. Because this difference of apparent
density did not influence the local edges of the bone in a
reconstructed data set, beam hardening artifacts have
little influence on the local segmentation algorithm, as is
shown with the overlaid cross-sections (Fig. 5).
In conclusion, the local threshold method gives a nearly
exact representation of a scanned bone at high resolutions
without the need for a priori knowledge about the volume of
the bone. When analyzing high-resolution scans of homoge-
nous structures, for example, bone biopsy specimens, the per-
olds on in vitro and in vivo scans of whole
bones. (A) Absolute difference in the estimation
of cortical bone and trabecular bone with respect
to the locally segmented data set as a function of
the global threshold value. The minimum of the
summed difference indicates an optimum. Note
that at the optimum (gray arrow) there is a
difference between the volumes of the locally
and globally segmented data set. (B) Example of
in vitro scan, showing a gray-value cross-section
(B1), that is segmented with the local method
(B2) and the global method using the optimal
threshold (B3). (C) Example of in vivo scan,
showing a gray-value cross-section (C1), both
segmented with the local method (C2) and the
global method (C3). Notice the overestimation
of the cortical bone and the subchondral bone in
the globally segmented cross-sections.
The effect of global and local thresh-
1649 IMPROVED SEGMENTATION FOR ?CT
formance of global threshold methods perform similarly to the Download full-text
local threshold method. As soon as the scanned structures are
not homogenous and include both thick cortices and thin
trabeculae or when scan resolution is relatively low, the local
threshold method outperforms the global methods. This makes
the local threshold method ideally suited, and perhaps even
crucial, for use with in vivo ?CT.
This work was supported by European Union Grant
1. Niebur GL, Feldstein MJ, Keaveny TM 2002 Biaxial failure behavior
of bovine tibial trabecular bone. J Biomech Eng 124:699–705.
2. Feldkamp LA, Goldstein SA, Parfitt AM, Jesion G, Kleerekoper M
1989 The direct examination of three-dimensional bone architec-
ture in vitro by computed tomography. J Bone Miner Res 4:3–11.
3. Van Rietbergen B, Huiskes R, Eckstein F, Ruegsegger P 2003
Trabecular bone tissue strains in the healthy and osteoporotic
human femur. J Bone Miner Res 18:1781–1788.
4. Kinney JH, Haupt DL, Balooch M, Ladd AJ, Ryaby JT, Lane NE
2000 Three-dimensional morphometry of the L6 vertebra in the
ovariectomized rat model of osteoporosis: Biomechanical implica-
tions. J Bone Miner Res 15:1981–1991.
5. Van Der Linden JC, Verhaar JA, Weinans H 2001 A three-
dimensional simulation of age-related remodeling in trabecular
bone. J Bone Miner Res 16:688–696.
6. Waarsing JH, Day JS, van der Linden JC, Ederveen AG, Spanjers
C, De Clerck N, Sasov A, Verhaar JA, Weinans H 2004 Detecting
and tracking local changes in the tibiae of individual rats: A novel
method to analyse longitudinal in vivo micro-CT data. Bone 34:
7. David V, Laroche N, Boudignon B, Lafage-Proust MH, Alexandre
C, Ruegsegger P, Vico L 2003 Noninvasive in vivo monitoring of
bone architecture alterations in hindlimb-unloaded female rats
J Bone Miner Res 18:1622–1631.
8. Fajardo RJ, Ryan TM, Kappelman J 2002 Assessing the accuracy
of high-resolution X-ray computed tomography of primate trabec-
ular bone by comparisons with histological sections. Am J Phys
9. Muller R, Van Campenhout H, Van Damme B, Van Der Perre G,
Dequeker J, Hildebrand T, Ruegsegger P 1998 Morphometric
analysis of human bone biopsies: A quantitative structural com-
parison of histological sections and micro-computed tomography.
10. Mitton D, Cendre E, Roux JP, Arlot ME, Peix G, Rumelhart C,
Babot D, Meunier PJ 1998 Mechanical properties of ewe vertebral
cancellous bone compared with histomorphometry and high-
resolution computed tomography parameters. Bone 22:651–658.
11. Ito M, Nakamura T, Matsumoto T, Tsurusaki K, Hayashi K 1998
Analysis of trabecular microarchitecture of human iliac bone using
microcomputed tomography in patients with hip arthrosis with or
without vertebral fracture. Bone 23:163–169.
12. Hara T, Tanck E, Homminga J, Huiskes R 2002 The influence of
microcomputed tomography threshold variations on the assess-
ment of structural and mechanical trabecular bone properties. Bone
13. Muller R, Koller B, Hildebrand T, Laib A, Gianolini S, Ruegseg-
ger P 1996 Resolution dependency of microstructural properties of
cancellous bone based on three-dimensional mu-tomography.
Technol Health Care 4:113–119.
14. Cendre E, Kaftandjian V, Peix G, Jourlin M, Mitton D, Babot D
2000 An investigation of segmentation methods and texture anal-
ysis applied to tomographic images of human vertebral cancellous
bone. J Microsc 197:305–316.
15. Ding M, Odgaard A, Hvid I 1999 Accuracy of cancellous bone
volume fraction measured by micro-CT scanning. J Biomech 32:
16. Kuhn JL, Goldstein SA, Feldkamp LA, Goulet RW, Jesion G 1990
Evaluation of a microcomputed tomography system to study tra-
becular bone structure. J Orthop Res 8:833–842.
17. Dufresne T 1998 Segmentation techniques for analysis of bone by
Health Care 6:351–359.
18. Elmoutaouakkil A, Peyrin F, Elkafi J, Laval-Jeantet AM 2002
Segmentation of cancellous bone from high-resolution computed
tomography images: Influence on trabecular bone measurements.
IEEE Trans Med Imaging 21:354–362.
19. Gomberg BR, Saha PK, Wehrli FW 2003 Topology-based orientation
analysis of trabecular bone networks. Med Phys 30:158–168.
20. Laib A, Newitt DC, Lu Y, Majumdar S 2002 New model-
independent measures of trabecular bone structure applied to
in vivo high-resolution MR images. Osteoporos Int 13:130–136.
21. Majumdar S, Genant HK, Grampp S, Newitt DC, Truong VH, Lin
JC, Mathur A 1997 Correlation of trabecular bone structure with
age, bone mineral density, and osteoporotic status: In vivo studies
in the distal radius using high resolution magnetic resonance im-
aging. J Bone Miner Res 12:111–118.
22. Gordon CL, Webber CE, Christoforou N, Nahmias C 1997 In vivo
assessment of trabecular bone structure at the distal radius from
high-resolution magnetic resonance images. Med Phys 24:585–593.
23. Newitt DC, van Rietbergen B, Majumdar S 2002 Processing and
analysis of in vivo high-resolution MR images of trabecular bone
for longitudinal studies: Reproducibility of structural measures and
micro-finite element analysis derived mechanical properties. Os-
teoporos Int 13:278–287.
24. Muller R, Hildebrand T, Ruegsegger P 1994 Non-invasive bone
biopsy: A new method to analyse and display the three-dimensional
structure of trabecular bone. Phys Med Biol 39:145–164.
25. Muller R, Hildebrand T, Hauselmann HJ, Ruegsegger P 1996 In
vivo reproducibility of three-dimensional structural properties of
noninvasive bone biopsies using 3D-pQCT. J Bone Miner Res
26. Gordon CL, Webber CE, Adachi JD, Christoforou N 1996 In vivo
assessment of trabecular bone structure at the distal radius from
high-resolution computed tomography images. Phys Med Biol
27. Laib A, Ruegsegger P 1999 Comparison of structure extraction
methods for in vivo trabecular bone measurements. Comput Med
Imaging Graph 23:69–74.
28. Laib A, Hildebrand T, Hauselmann HJ, Ruegsegger P 1997 Ridge
number density: A new parameter for in vivo bone structure
analysis. Bone 21:541–546.
29. Russ J 1999 The Image Processing Handbook. CRC Press, Boca
Raton, FL, USA.
30. Canny J 1986 A computational approach to edge detection. IEEE
Trans Pattern Analysis Machine Intelligence 8:679–698.
31. American Society for Testing and Materials 1997 Standard Guide
for Computed Tomography (CT) Imaging: ASTM Standards, vol.
E1440–97. American Society for Testing and Materials, Philadel-
phia, PA, USA.
32. Maes F, Collignon A, Vandermeulen D, Marchal G, Suetens P
1997 Multimodality image registration by maximization of mutual
information. IEEE Trans Med Imaging 16:187–198.
33. Hildebrand T, Ruegsegger P 1997 A new method for the model-
independent assessment of thickness in three-dimensional images.
J Microsc 185:67–75.
34. Odgaard A, Gundersen HJ 1993 Quantification of connectivity in
cancellous bone, with special emphasis on 3-D reconstructions.
35. Hildebrand T, Ruegsegger P 1997 Quantification of bone micro-
architecture with the structure model index. Comput Methods
Biomech Biomed Engin 1:15–23.
tomographic imaging. Technol
Address reprint requests to:
Harrie Weinans, PhD
Erasmus Medical Center
Erasmus Orthopaedic Research Lab, EE1614
PO Box 1738
Rotterdam 3000 DR, The Netherlands
Received in original form November 19, 2003; in revised form
May 3, 2004; accepted May 21, 2004.
1650WAARSING ET AL.