Osteoporos Int (2005) 16: S25–S35

DOI 10.1007/s00198-004-1701-7

O R I GI N A L A R T IC L E

Long-term prediction of three-dimensional bone architecture

in simulations of pre-, peri- and post-menopausal
microstructural bone remodeling
Ralph Müller

Received: 4 June 2004 / Accepted: 9 June 2004 / Published online: 31 August 2004

International Osteoporosis Foundation and National Osteoporosis Foundation 2004

Abstract The mechanical behavior of trabecular bone the onset of menopause an imbalance in the cell activi-
depends on the internal bone structure. It is generally ties was modeled resulting in a net bone loss. The dou-
accepted now that the trabecular bone structure is a bling of the activation frequency in the peri-menopausal
result of a load adaptive bone remodeling. The mathe- phase caused a pronounced loss. Using advanced ani-
matical laws that relate bone remodeling to the local mation tools and quantitative bone morphometry, the
state of stress and strain, however, are still under changes in bone architecture associated with the bone
investigation. The aim of this project was to investigate loss were monitored over an average observation time of
if changes in the trabecular architecture as observed with 43 years until the age of 80 years. In that time, bone
age-related bone loss and osteoporosis can be predicted volume density decreased monotonously with the pro-
from a computer model that simulates bone resorption gression of the simulation for all specimens. Right after
after hormone depletion based on realistic models of the onset of menopause, bone was lost fast, where with
trabecular microstructure using micro-computed the progression of age losses slowed down. The struc-
tomography (lCT). A compact desktop lCT providing tures at the end-point of the simulations were then
a nominal isotropic resolution of 14 lm was used to compared qualitatively and quantitatively to the struc-
measure two groups of seven trabecular bone specimens tures of the post-menopausal group with all morpho-
from pre-menopausal and post-menopausal women metric indices being within a narrow margin of error.
respectively. A novel algorithm was developed to simu- These results suggest the feasibility of transforming
late age-related bone loss for the specimens in the first ‘‘normal’’ to ‘‘osteopenic’’ bone on a microstructural
group. The algorithm, also referred to as simulated bone level yielding in realistic bone models similar in
atrophy (SIBA), describes a truly three-dimensional appearance as well as in structural behavior if compared
approach and is based directly on cellular bone remod- to a post-menopausal group of women.
eling with an underlying realistic time frame. Bone
resorption is controlled by osteoclastic penetration Keywords Age-related bone loss Æ Bone remodeling Æ
depth and bone formation is governed by the efficiency Micro-computed tomography (lCT) Æ Osteoporosis Æ
level of the osteoblasts. The simulation itself describes Trabecular bone architecture
an iterative process with a cellular remodeling cycle of
197 days. Activation frequency is controllable and can
be adjusted for the different phases of pre-, peri- and
post-menopause. For our simulations, osteoblastic and Introduction
osteoclastic activities were in balance until the onset of
menopause, set to be at the age of 50 years. In that Osteoporosis occurs most frequently in post-meno-
period, the structure remained almost constant. After pausal women and the aged. It is defined as a systemic
skeletal disease characterized by low bone mass and
micro-architectural deterioration of bone tissue. Cur-
R. Müller rently osteoporosis is considered an epidemic in the
Institute for Biomedical Engineering, United States with approximately one-third of women
Swiss Federal Institute of Technology (ETH) and over the age of 65 afflicted with the disease [1]. The re-
University of Zürich, Moussonstrasse 18, duced quality of osteoporotic bone greatly increases the
8044 Zürich, Switzerland
E-mail: [email protected] risk of fractures [2,3]. As a result, health care costs re-
Tel.: +41-1-6324592 lated to osteoporotic fractures neared $14 billion in 1995
Fax: +41-1-6321214 [4]. Although many older persons may lose bone, not all
S26

develop fractures. The mechanical behavior of bone istic three-dimensional microstructural models of tra-
depends on the internal bone structure as well as the becular bone. Today, such models can be generated
loads applied. It is also widely assumed that mechanical directly using either the destructive method of serial
stresses and strains influence the remodeling process and sectioning [17] or by employing non-destructive three-
subsequently the structure and strength of the bone. dimensional imaging techniques like micro-computed
This adaptive response is generally referred to as Wolff’s tomography (lCT) [18,19,20]. In this study, we used
law after the German anatomist who first called atten- models of trabecular bone imaged with a compact
tion to it. In 1892, Wolff [5] postulated that the trajec- desktop lCT system [21]. For the purpose of the study,
tories of the trabecular bone align with the principal two groups of seven specimens were defined for a pre-
stress trajectories. Although the basic concepts of menopausal and a post-menopausal group of women. A
Wolff’s law are generally accepted, the mathematical novel algorithm was developed to simulate age-related
laws relating bone remodeling to the stress/strain rela- bone loss for the specimens in the first group. The
tions are still under investigation. algorithm, also referred to as simulated bone atrophy
Traditionally, animal experiments have been em- (SIBA), describes an iterative, truly three-dimensional
ployed to investigate bone response from load altera- approach based on cellular bone remodeling and, for the
tions or during the progress of age-related bone loss. first time, an underlying realistic time frame. The pre-
Since the time scale of the remodeling process is of the and post-menopausal group as well as the set of
order of months or years, computational modeling offers simulations were analysed and compared using three-
a unique approach to study long term processes without dimensional quantitative bone morphometry and newly
the inconveniences of animal experimentation. Com- developed animation techniques to visualize and moni-
puter modeling is reliable, inexpensive and fast, and thus tor changes in trabecular bone architecture over several
different optimization goals can be tried systematically decades.
until shape and architecture of bone can be reproduced.
Mathematical formulations of Wolff’s law for trabecular
bone remodeling have been based on two general Materials and methods
approaches. The first method is the development of
empirical stress/morphology relationships that relate Two groups of seven trabecular bone specimens were
experimental measures of trabecular architecture to selected from a larger pool of data from the European
finite element estimates of stress distributions [6]. The Union BIOMED1 Concerted Action ‘‘Assessment of
second method is the development of iterative algo- Bone Quality in Osteoporosis’’, where cylindrical bone
rithms to predict load-driven bone adaptation based on biopsies from different anatomical locations were har-
the stress/strain state, again using finite element tech- vested post-mortem from 32 females and 38 males [22].
niques to predict stress and strain distributions Donor age in the BIOMED1 study ranged from 23 to 92
[7,8,9,10,11,12,13,14,15]. years, with a mean age of 68±16 years. The study in-
Since the work of Frost in the early 1960s, bone has cluded samples from the iliac crest (IC, transiliac), the
been considered to remodel in discrete localized regions. femoral head (FH), the lumbar spine (L2 and L4) and the
All cells participating in the remodeling process have calcaneus (CA) for every subject [23]. The two subgroups
been termed the bone multicellular unit (BMU) and are included four trabecular bone cores from the iliac crest
thought to proceed through activation, resorption, and (IC) and three from the lumbar spine (L2) site each and
formation [16]. Remodeling has been proposed to were chosen to represent a pre-menopausal (36.7±9.5
function as a method of preventing the accumulation of years) and a post-menopausal (77.4±3.3 years) group of
fatigue damage and maintaining an adequate supply of women, respectively. The criteria for the pre-menopausal
relatively low mineral density bone to subserve mineral group were an age of less than 50 years and a medical
homeostasis. But what happens when the bone loses its record showing no known bone disorders. For the post-
ability for load-driven adaptation, when microdamage is menopausal group, only women in the 7th and 8th
no longer repaired, as seems to be the case for bone loss decade of their life were included as a consecutive sample.
associated with age, medication or diseases? In these The goal was to have a cross-section of the BIOMED1
cases, the bone is left with its function to renew old population rather than a selected group of individuals
bone, where each remodeling cycle seems to involve a that could be considered normal or osteoporotic.
net loss of bone mass, and as a consequence of this
slightly negative balance between osteoclastic and
osteoblastic activity, thinning of trabeculae and an in- Three-dimensional microstructural imaging
crease in trabecular spacing take place. Therefore, bone
gradually loses its strength over time. To generate the three-dimensional models of micro-
The aim of this project was to investigate if changes structural bone, the selected biopsies were measured
in trabecular architecture as observed with age-related using a micro-tomographic imaging system (lCT 20;
bone loss and osteoporosis can be predicted from a Scanco Medical, Bassersdorf, Switzerland), a compact
computer model that simulates bone resorption after fan-beam type of tomograph, also referred to as
hormone depletion over several decades based on real- desktop lCT. The system was especially designed for the
 S27

non-destructive measurement and analysis of unpro- structures change with age or loading [27,28]. Because
cessed surgical bone biopsies [21]. For each sample, a the spatial resolution of the QCT measuring system
total of 286 micro-tomographic slices was acquired. (250 lm) and the mean trabecular width (80–150 lm)
Measurements were stored in three-dimensional image were of the same order, it was not yet possible to resolve
arrays. The total examination time per specimen was the true shape of individual trabeculae. It was, however,
approximately 3 h for a voxel size of 14 lm. possible to segment the trabeculae due to their large
A three-dimensional Gaussian filter with a limited, separation (400–1200 lm). Nonetheless, to implement a
finite filter support was used to partly suppress the noise computer simulation of age-related bone loss directly
in the volumes. In the next step, the bone tissue was based on osteoblastic and osteoclastic activity, which is
segmented from marrow and solution with the help of a the goal of this project, it seems to be necessary to have a
thresholding procedure. All samples were binarized resolution on the order of 20 lm or better as provided
using the same parameters for the filter width (1.2), the by the desktop micro-tomography system described
filter support [2] and the threshold (10.2% of the maxi- earlier. To come up with a cell-based approach, the
mal attenuation) [21]. For the subsequent simulation, a original SIBA algorithm had to be extended to allow an
subvolume of the originally measured data was selected iterative approach directly associated with an underlying
including only binarized trabecular plates and rods. This physical time frame and to reformulate the laws for bone
selected volume of interest (VOI) was located in the resorption in the simulated bone atrophy.
center of the specimen to eliminate preparation artifacts The extended SIBA algorithm is based on the
on the surface. The size of the VOI was always remodeling sequence first described by Frost [29], which
3.6·3.6·3.6 mm (256·256·256 voxels). consists of three phases: resorption, reversal, and for-
mation. Thomsen et al. [24] used an iterative approach
for their stochastic model and the time interval between
Simulated bone atrophy and age-related bone loss iterations was 1 day. In our approach [30], one iteration
would include all three steps, where bone resorption
Thomsen et al. [24] developed a computer simulation (osteoclastic activity) and bone formation (osteoblastic
model that could assist in predicting the long-term effect activity) would be modeled for each iteration. Never-
of changes in the remodeling process on bone mass, theless, the timeframe for the total remodeling cycle was
trabecular thickness, and perforations. They presented a chosen to equal 197 days to meet the average remodeling
stochastic model of the remodeling process in human cycle time from the Thomsen study. Furthermore, the
vertebral trabecular bone, where the model is based on algorithm is based on the assumption that for normal
histomorphometric and structural data from human bone osteoclastic and osteoblastic activity are in balance
studies. The model describes the variation in bone mass, resulting in an almost stable bone mass and bone
average trabecular thickness and number of perforations structure. After the onset of menopause, which was as-
over an extended period of time. We wanted to take that sumed to be at the age of 50 years, the balance would
approach one step further and introduce a model di- become negative, i.e. more bone is resorbed than
rectly based on realistic three-dimensional trabecular formed, resulting in a net loss of bone mass. No further
bone structures. In order to include the three-dimen- information was needed to run the models. The changes
sional bone architecture in the modeling, we introduced in architecture, we observed, were therefore only the
the concept of non-invasive bone biopsy [25], where the result of the loss in bone mass over the years, which
three-dimensional structure of trabecular bone was ac- means that the architecture as it was present before
quired non-invasively by a high-resolution quantitative menopause already predetermined the outcome in the
CT system for peripheral measuring sites. Based on non- simulated bone atrophy with a given, constant loss of
invasive bone biopsy, we developed a novel bone bone. In the simulation, osteoclastic activity is modeled
resorption algorithm to investigate the biomechanical as an exponential function, i.e. the probability for bone
consequences of age- and disease-related bone loss [26]. being resorbed at a certain location in the bone matrix
In that approach, bone loss was expressed as a decrease has a Gaussian profile. In practical terms, this means
in bone density in combination with topological altera- that the algorithm, applied to the digitized trabecular
tions of the bone structure. In order to be able to control bone structure, is based on constrained Gauss filtration
the process of bone loss, an algorithm to simulate bone of binary data sets, i.e. only bone and non-bone will be
resorption and adaptive processes based on discrete differentiated. The effect of the Gaussian filter is mainly
volume data was developed. The algorithm, also referred controlled by the filter width and the constrained filter
to as simulated bone atrophy (SIBA), generated a set of support. Constrained, in this sense, means that the dis-
microstructural models, all derived from the same ori- crete convolution of the Gaussian and the binary data
ginal structure. Generally, the models varied in respect set is computed only for a limited, finite filter support. In
to both apparent density and the degree of structural our case, we used a filter support of 2, which means that
anisotropy. The algorithm did not include a proposal for a newly computed value only depended on the sum of
a new mathematical formulation of bone adaptation but the weighted values of the point’s direct neighborhood
more based on a phenomenological approach and (5·5·5). This also means that an osteoclast can pene-
observations made by other researchers of how bone trate into the bone only two voxels per cycle, which
S28

corresponds to a perforation depth ranging between 28 determine the values for osteoclastic perforation depth,
and 48 lm depending on the neighborhood configura- the osteoblastic efficiency level and activation frequency.
tion. Since the original structures were based on binary Activation frequency (AF) is defined as the rate at which
data, every data point in the volume has either a value of bone multicellular units (BMU) are formed. Together
one or zero prior to the filtration. The constrained Gauss with the individual cellular function rates at the BMU, it
filtration will result in a gray-scale data set with values determines the percentage of bone turnover [32]. AF was
between zero and one. The more data points are ‘‘set’’ in not part of the original algorithm, where the length of
the weighted neighborhood the higher the resulting gray- one iteration was simply defined by the length of the
scale values and vice versa. Therefore, the application of remodeling cycle (197 days). In the present simulation,
the Gaussian filter introduces a small gradient on the the length of one iteration was defined as 1/AF under the
surface of the structure, whereas values in the center of assumption that each BMU is activated in that time.
the structure, totally surrounded by mineralized bone, Normal activation frequency was assumed at 1/1096
will be unchanged. The total amount of bone mass in a days [24], which corresponded to an iteration length of
local neighborhood (sum of weighted voxels) is the same approximately 3 years. In peri-menopause, activation
before and after the filtration but the three-dimensional frequency was seen to double [33], which resulted in a
distribution of the mass has changed according to the cycle time of 1.5 years in our simulation. In order to
local topology. In our simulation, the four phases in the establish a more realistic time frame and estimation for
remodeling cycle are fully synchronized for all cell the values of the simulation parameters after the onset of
activities in the structure, which means that in a first step menopause, the experimental post-menopausal group
bone is resorbed only and then in a second step bone is was used as a control group. The strategy was to adapt
formed only. There is no temporal overlay of the the osteoclastic efficiency level in a way to match the
osteoclastic and osteoblastic activity, which in reality is volumetric densities of simulation and experimental
not the case and has therefore to be taken as a current group at the end-point of the simulation. It is clear that
limitation of the algorithm. Additionally, the amount of such an approach can only be one of many scenarios of
surface area involved in a remodeling cycle was not what could have happened to these pre-menopausal
controlled, so that in the simulation the whole surface bone specimens with time. Important for us was to see
was affected by either osteoclastic or osteoblastic cell whether the outcome of the simulation would result in
activity, which again in reality is most probably not the reasonable and realistic structures if compared to the
case. The amount of bone that is removed or added in post-menopausal experimental group. The validation of
one iteration step is controlled by setting a global the algorithm was therefore based on both a qualitative
threshold for the filtered data, which must be between 0 (visualization) and quantitative (morphometry) com-
and 1. In our simulation, we found that a threshold of parison between those two groups. For the three-
0.5 (all voxels with a gray-scale value higher or equal to dimensional visualization, an extended marching cubes
0.5 are included) would allow maintenance of the total algorithm was used [25]. The algorithm, which is based
amount of mass. This was defined as the 100% efficiency on a divide-and-conquer approach, triangulates the
level for the osteoblastic activity, which means that the surface of any given voxel array in a fast and secure way.
osteoblasts were able to form as much bone as has been It not only allows perfect triangulation, but also
resorbed by the osteoclasts. A threshold of 1.0 on the smooths the surface effectively.
other hand resulted in the removal of all elements with
at least one zero voxel in their local 5·5·5 neighbor-
hood, which would correspond to a 0% efficiency level. Direct quantitative bone morphometry
The relation between the threshold and the efficiency
level can be expressed as a second order polynomial, In addition to the visual assessment of structural chan-
which means that initially the efficiency is high and is ges with age, morphometric and architectural indices
dropping down quickly with increasing threshold. can be determined from the micro-tomographic exam-
Whereas the original implementation distinguished inations. In the past, structural properties of trabecular
only between pre- and post-menopause [30], where bone have been investigated by examination of two-
changes were assumed to happen only after the onset of dimensional sections of cancellous bone biopsies. Three-
menopause, the current implementation includes three dimensional morphometric parameters are then derived
distinctively different stages, namely the pre-, peri- and from two-dimensional images using stereological meth-
post-menopause. Pre-menopause is the period of time ods [34,35]. While parameters like bone volume density
after bone growth ceases and before menopause begins. (BV/TV) and surface density (BS/TV) can directly be
Although bone loss in this stage is minimal, it is obtained from two-dimensional images, a range of
important to model the structural changes prior meno- important parameters such as trabecular thickness
pause. Without its inclusion, the changes that occur will (Tb.Th), trabecular separation (Tb.Sp), and trabecular
be underestimated. The peri-menopausal stage is a per- number (Tb.N) are to be derived indirectly assuming a
iod of about 5 years after the onset of menopause where fixed structural model, typically an ideal plate or rod
the bone may actually undergo the most changes [31]. model is used. Such assumptions are, however, critical
Experimental values from the literature were used to due to the well-known fact that trabecular bone con-
 S29

tinuously changes its structure type as a result of cavities. Trabecular number (Tb.N) is defined for the
remodeling. A deviation from the assumed model will plate model as the number of plates per unit length. An
lead to an unpredictable error of the indirectly derived alternative geometrical interpretation can be formulated
parameters. This is particularly true in this study, where as the inverse distance between the midsection of the
we actually would like to follow the changes in bone plates. For a general structure, it would be possible to
structure in the course of age-related bone loss. In such a assess the mean distance between the mid axes of the
case, a predefined model assumption could easily over- structure and use the inverse of this measure to calculate
or underestimate the effects of the bone atrophy the mean number of elements per unit length. The mid-
depending on the assessed index. For these reasons and axes of the structure can be extracted from a binary
in order to take full advantage of the volumetric mea- three-dimensional image using different techniques [40].
suring technique, new methods that make direct use of In this implementation, three-dimensional distance
the three-dimensional information are required. To meet transformation to extract the geometrical mid points of
this demand several new methods by means of three- the structure was used [39]. To get a measure for mean
dimensional image processing have recently been pre- distance between the points the direct thickness method
sented, allowing direct quantification of the actual similar to the Tb.Sp calculation was applied, i.e. the
architecture of trabecular bone. In this study, we used separation between the mid-axes was assessed. In addi-
metric and also non-metric parameters entirely based on tion to the computation of the direct metrical parame-
direct three-dimensional calculations. The definitions ters, other non-metric parameters can be calculated to
and methods used for the calculation of the model describe the three-dimensional nature of the bone
independent parameters have recently been developed structure. An estimation of the plate-rod characteristic
and introduced for the microstructural evaluation of the of the structure can be obtained from the structure
BIOMED1 study [36]. model index (SMI) [41]. SMI is calculated by a differ-
In the presented approach, the volume of the tra- ential analysis of a triangulated surface of a structure.
beculae (BV) was calculated using tetrahedrons corre- For an ideal plate and rod structure the SMI value is 0
sponding to the enclosed volume of the triangulated and 3, respectively. For a structure with both plates and
surface used for the surface area calculation. From this rods of equal thickness the value will be between 0 and 3,
measure the relative bone volume (BV/TV) can be cal- depending on the volume ratio between rods and plates.
culated. The total volume (TV) is the volume of the Another parameter often used as an architectural index
whole examined sample. The bone surface area (BS) was is the geometrical anisotropy of a structure, which is
calculated using surface triangulation of the binary data typically determined using mean intercept length (MIL)
[25]. The specific bone surface or bone surface-volume measurements [42]. MIL denotes the average distance
ratio is given by BS/BV. In conventional stereology, between bone/marrow interfaces and is measured by
these primary parameters are then be used to derive tracing test lines in different directions in the examined
other indices such as the trabecular thickness (Tb.Th), VOI. From this measurement, an MIL tensor can be
the trabecular separation (Tb.Sp) or the trabecular calculated by fitting the MIL values to an ellipsoid. The
number (Tb.N) based on the underlying model eigenvalues of this tensor can then be used to define the
assumption. An even more indirect way to calculate degree of anisotropy (DA), which denotes the maximum
these metric parameters is to derive the involved surface to minimum MIL ratio.
area indirectly using methods such as mean intercept
length or two-dimensional perimeter estimation. Such
methods were developed for the two-dimensional anal- Results
ysis of trabecular bone, but are nevertheless used for the
analysis of three-dimensional images [37]. A recent study Bone atrophy simulation
showed that such indirectly derived primary quantities
could vary as much as 52%, depending on which First, a time-zero iteration with no associated time and
method is used [38]. A more direct approach to assess bone loss was used to convert the measured bone spec-
metric parameters from three-dimensional images is imen to a computer model. A sigma value of 0.5, a filter
based on measuring actual distances in the three- support of 2, and a 100% efficiency level were chosen.
dimensional space. Such techniques do not rely on an For each phase of the simulation, a set of parameters
assumed model type and are therefore not biased by had to be determined and implemented. Individual dif-
eventual deviations of the actual structures from this ferences between the specimens were not accounted for.
model. In such a way, mean trabecular thickness All parameters were chosen to be homogenous for each
(Tb.Th) can be calculated directly and without an of the three phases of pre-, peri- and post-menopause.
underlying model assumption by determining a local For the pre-menopausal phase, an osteoclastic penetra-
thickness at each voxel representing bone [39]. The same tion depth of approximately 30 lm (sigma=1.1, sup-
method can be used to calculate the mean trabecular port=2), an osteoclastic efficiency level of 99.7%
separation (Tb.Sp) by applying the thickness calculation (resulting in a decrease in volume density of 0.34% per
to the non-bone parts of the three-dimensional image. year) and an AF of 1/1096 days were chosen. The
The separation is thus the thickness of the marrow number of iterations to reach the age of 50 years, the
S30

assumed onset of menopause, varied between zero and menopausal group on a visual basis. From that, we can
nine depending on the individual age of the donor bone. conclude that the algorithm does in fact produce realistic
The average bone loss that occurred was )5.6% over an structures comparable to the old-aged group. Figures 1
average of 13.4 years of pre-menopause, which was and 2 show examples for the iliac crest and the lumbar
consistent with the values found in the literature for that spine respectively. In the illustrations, the baseline
time period [43]. In peri-menopause the osteoclastic structure of a pre-menopausal woman and the corre-
penetration depth was increased by 40% [24,44] chang- sponding simulation are compared to the structure of a
ing the sigma from 1.1 to 1.5, the osteoclastic efficiency post-menopausal woman for both measurement sites.
level was reduced to 95% and AF was doubled to 1/548 The distinct transformation of the trabecular bone
days [33] resulting in a total of four iterations needed to architecture from a ‘‘plate-like’’ to a more ‘‘rod-like’’
simulate 6 years of peri-menopause. Finally, in post- structure in the course of age-related bone loss can easily
menopause, osteoclastic penetration depth and AF were be noticed. The advantage of simulated bone atrophy is
set to the values for the pre-menopausal phase with a that it generates not only static images at the end-point
sigma of 1.1 and an AF of 1/1096 days, respectively. of the simulation but rather allows to follow the changes
Post-menopause was simulated with a total of eight with age step by step as illustrated in Fig. 3, where we
iterations concluding at the age of 80 years. concentrated on a smaller detailed view of the structural
Simulated bone atrophy was applied to all seven pre- changes from a simulation of an iliac crest bone sample.
menopausal specimens resulting in 12–21 models of For a young and healthy subject as illustrated in the top-
atrophied bone per specimen starting at the individual left corner of the figure we typically find a plate-like
age of the donor bone until a simulated age of 80 years structure with smaller rods connecting the plates for
was reached. The average computer time needed for a both the iliac crest and the lumbar spine. With increas-
full length simulation varied between 180 and 220 min ing age the plate-like structure is partly resolved into a
per specimen using a personal workstation DEC 3000- rod-like structure. The rods forming the trabecular
M300LX (Digital Equipment Cooperation, Maynard, network are more and more thinned until they are
USA). After the simulations, three-dimensional visual- eventually perforated. After they are disconnected they
izations as well as structural parameters using direct disappear quickly, leaving the structure more widely
three-dimensional techniques were computed for the spaced and less well connected.
total of 129 microstructural bone models including pre- These changes were also reflected in the structural
menopausal (7), post-menopausal (7), and all the simu- properties assessed by quantitative bone morphometry.
lations (115). In total, this accounted for a few weeks of The relative bone volume decreased monotonously with
computer time with an approximated 250 minutes of the progression of the simulation for all seven specimens
calculation time per model. (Fig. 4). Right after the onset of menopause, bone was
lost fast, where with the progression of age the losses
slowed down. When we looked at the relative changes
Qualitative and quantitative validation over the course of the simulation, the different specimens
lost different amounts of bone with DBV/TV ranging
In a first attempt to validate the algorithm, we simply from 21% to 68% (Fig. 5). This is an interesting fact,
compared the three-dimensional structures at the end- since the efficiency level for bone formation controlling
points of the simulations with the structures of the post- the net bone loss per iteration was set as identical for all
specimens, and those specimens that lost bone ‘‘fast’’
were not associated with an especially high or low bone
density.
Fig. 1 Three-dimensional trabecular bone architecture of iliac crest This left architecture was the only independent
bone biopsies of a 32-year-old (left), the corresponding ‘‘80-year- factor in the simulation influencing the rate of bone loss,
old’’ simulation (middle) and an 83-year-old (right) woman, which means that the architecture itself determined
respectively. Both the simulation and the structure from the age-
matched group show a relative bone volume of 13% whereas the whether an individual lost bone ‘‘fast’’ or ‘‘slow’’ in our
bone volume at the pre-menopausal baseline was measured at 21% simulation with a constant efficiency level and a constant
 S31

Fig. 2 Three-dimensional trabecular bone architecture of lumbar specimen where we saw an decrease of about 20%, again
spine (L2) bone biopsies of a 50-year-old (left), the corresponding showing the ability of the simulation to pick up indi-
‘‘80-year-old’’ simulation (middle) and an 80-year-old (right)
woman, respectively. Both the simulation and the structure from vidual differences in architecture. Trabecular separation
the age-matched group show a relative bone volume of 7%, showed similar although inverted trends as the primary
whereas the bone volume at the pre-menopausal baseline was parameter of bone volume with changes ranging from
measured at 11% +9% up to +19%. Tb.N on the other side was affected
heavily by the procedure with changes from )17% to
remodeling cycle time. When we looked at specific bone )55%, indicating that the simulated bone loss left the
surface (Fig. 6), we noticed that BS/BV was increasing structure with fewer trabeculae. Structure model index
during the simulation but the effects were not as pro- supported the theory that trabecular bone is trans-
nounced as for bone volume density. Nevertheless, the formed from plate-like to rod-like in the progress of age-
specific surface density at the age of 50 (start of meno- related bone loss with an initial mean value of 1.1 (0.5–
pause) predicted the total relative changes in bone vol- 2.1), indicating more plate-like structures (the value for a
ume density over 30 years with an r2 of 0.96, which perfect plate model is 0) and a mean value of 1.8 (1.1–
means that 96% of the changes in bone density over time 2.7) after the simulation indicating a transition to more
could be predicted by the initial BS/BV measure (Fig. 7). rod-like structures (the value for a perfect rod model is
Mean trabecular thickness was not much affected by 3). The most dramatic changes in SMI were found in the
the simulation (approximately ±10%), except for one iliac crest specimens with an increase of over 150%. The

Fig. 3 Detail from the

simulation of an iliac crest bone
biopsy of a 47-year-old woman.
The simulation illustrates the
changes in bone architecture
over time, where with
increasing age the plate-like
structure is partly resolved and
with that transformed into a
more rod-like structure. In this
process, the rods forming the
network are more and more
thinned until they are
eventually perforated. After a
perforation takes place the
remaining rod disappears
quickly, leaving the structure
more widely spaced and less
well connected
S32

Fig. 7 BS/BV measured at menopause is highly correlated to the

Fig. 4 BV/TV in the course of age-related bone loss as determined total amount of relative bone loss occurring over 3 decades
by simulated bone atrophy. The simulation started at the individual
ages of each donor. The onset of menopause was assumed to be at
the age of 50 years smallest changes were seen in the lumbar spine with an
increase of SMI from 1.8 to 2.4 (+27%).
The results from the quantitative morphometry for
the end-points of the simulation were then compared to
the values of the post-menopausal group directly anal-
ysed after micro-tomographic imaging (Table 1). Since
we used BV/TV to calculate the efficiency level, the two
values for the simulation and the post-menopausal
group had to be in full accordance and showed both an
average decrease of 38% compared to the pre-meno-
pausal baseline measurements. BS/BV, as the other
primary parameter, only increased 12% for the post-
menopausal group and 19% for the simulation group,
respectively. Tb.N decreased 32% for the post-meno-
pausal group and 37% for the simulation group
respectively. Tb.Sp showed again similar results for the
post-menopausal and the simulation group, with in-
creases of 25% and 23% respectively, whereas Tb.Th
showed a slight decrease of 11% for the post-meno-
Fig. 5 Relative changes in BV/TV with respect to the simulated age pausal group and a 4% increase for the simulation. If we
starting at menopause
look at the non-metric parameters, we find very similar
results for the SMI with an increase of 72% and 77%
respectively. DA, on the other hand, changed only
moderately for the post-menopausal group with an in-
crease of 7% whereas in the simulation the degree of
anisotropy changed by 9%.

Discussion

The aim of this project was to develop an algorithm

which allows simulation of the response of the trabec-
ular bone in age-related bone loss and to determine the
biomechanical consequences of such a response based
on realistic three-dimensional models of the trabecular
microstructure. Two groups of seven trabecular bone
specimens were measured micro-tomographically,
including specimens from pre-menopausal and post-
Fig. 6 Relative changes in BS/BV with respect to the simulated age menopausal women, respectively. In order to control
starting at menopause bone loss over time, a novel algorithm to simulate bone
 S33

Table 1 Mean (SD) of assessed

morphometric parameters Parameters Pre (exp.) Peri Peri post Post Post (exp.)

Age (years) 36.6 (9.3) 50 56 77 77.4 (3.3)

BV/TV (%) 17.9 (6.7) 17.3 (7.0) 14.6 (7.0) 11.0 (6.9) 11.0 (4.4)
BS/BV (1/mm) 19.1 (5.4) 18.6 (5.6) 20.2 (6.9) 22.8 (8.8) 21.4 (3.5)
Tb.N (1/mm) 1.6 (0.4) 1.5 (0.3) 1.3 (0.4) 1.0 (0.4) 1.1 (0.3)
Tb.Th (mm) 0.14 (0.03) 0.16 (0.05) 0.15 (0.05) 0.15 (0.05) 0.13 (0.02)
Tb.Sp (mm) 0.61 (0.10) 0.66 (0.07) 0.69 (0.08) 0.75 (0.10) 0.76 (0.10)
DA 1.2 (0.1) 1.3 (0.1) 1.3 (0.1) 1.4 (0.2) 1.3 (0.1)
SMI 1.0 (0.6) 1.1 (0.7) 1.4 (0.7) 1.8 (0.7) 1.7 (0.6)

resorption and adaptive processes was developed. Al- structural level. Nevertheless, it might well be that to
though the simulations were computationally very describe the effects purely associated with age-related
intensive, minimal user interaction was needed to build bone loss in a hormone deficiency model, where bone
and analyze the models. All individual steps were fully seems to have at least partly lost its ability to adapt to
automated. The algorithm, also referred to as simulated alterations in its mechanical environment, load-driven
bone atrophy (SIBA), generated a set of microstructural modeling and remodeling might no longer have signifi-
models, iteratively derived from the original three- cant importance.
dimensional structure. Simulated bone atrophy was used To our knowledge, this is the first time that age-re-
to ‘‘age-match’’ the first and the second group incor- lated bone loss has been described and simulated directly
porating an underlying realistic timeframe for the sim- based on three-dimensional measurements of trabecular
ulation. Using quantitative bone morphometry and bone with the incorporation of a realistic time frame.
three-dimensional animation tools, the changes in bone Although validation will be very difficult, it will be of
density and bone architecture could be monitored over crucial importance for further application of simulated
an average observation time of 43 years until the age of bone atrophy. The only way to really validate any
80 years. The structures at the simulated age of 77 years remodeling theory would be the incorporation of in vivo
were then compared qualitatively and quantitatively to follow-up measurements providing not only important
the structure of the post-menopausal group with an experimental data on the time-scale and structural
average age of 77.4 years. Good correspondence was properties of atrophied bone but also allowing validation
found between the two groups. No significant changes of each remodeling step in the simulation. Kinney et al.
could be detected between parameters from the simula- [20] developed a CT-scanning method demonstrating
tion and the experimental group. Differences on the sequential changes in bone structure at a resolution of a
order of 10–20% can easily be accounted for given the few microns. However, this resolution is only possible in
huge biological variations of up to 50% in the different animal experiments, due to dose considerations. Since we
structural properties of trabecular bone as seen in the are interested in the prediction of the individual fracture
analysis of the BIOMED1 study [36]. risk of patients, the final aim must be the non-invasive
Although there are several limitations to this study, assessment of the bone architecture in human subjects in
we believe that we were able to demonstrate that it is vivo. In order to assess the structural changes in the
possible to simulate, at least certain aspects, of age-re- progress of rapid bone loss, Müller et al. [45] introduced
lated bone loss using the proposed mechanism. The a method to assess and analyze cancellous bone based
implemented approach in combination with micro- upon three-dimensional peripheral computed tomogra-
structural trabecular bone models allowed a successful phy in vivo. In their study, structural parameters could
transformation of ‘‘young-and-normal’’ to ‘‘old-and- be reproduced in vivo with a coefficient of variation of
osteopenic’’ bone on a microstructural level. Another less than 0.4% demonstrating the potential of high-res-
advantage of the proposed model was that an identical olution tomography to detect structural changes in tra-
remodeling law and time step was applied to both iliac becular bone during age-related bone loss or therapeutic
and lumbar trabecular bone, which raises the question and diagnostic trials. The nominal resolution provided
whether there is a universal law for all types of bone with by the system was 170 lm, which would not be sufficient
respect to the effects on the cellular level. One of the for validation purposes. Today, there are new in vivo
limitations of simulated bone atrophy is that at the systems coming into the clinical market [46] providing
moment no experimental data is available to backup the even higher resolutions down on the order of 90 lm.
proposed cellular mechanisms and that no mechanical Whether this resolution will be high enough allowing
feedback loop is included in the algorithm. This is both simulations of bone adaptation with the possibility
especially important for simulation of bone modeling to non-invasively follow bone remodeling and adapta-
related to osteogenesis or fracture healing in healthy tion in vivo will have to be investigated in future studies.
subjects with the ability to react to such changes in the One other future goal of the project will be to
overall loading patterns. For this reason, one of the investigate the consequences of bone atrophy on the
future goals must be a combination of load and hor- mechanical behavior of the trabecular bone over
mone driven modeling and remodeling on a micro- time. Generally, the influence of the microstructural
S34

adaptation on the biomechanical competence of the 17. Odgaard A, Andersen K, Melsen F, Gundersen HJ (1990) A
bone can be expressed as a function of its anisotropic direct method for fast three-dimensional serial reconstruction. J
Microsc 159:335–342
bone properties on the continuum level. It has been 18. Feldkamp LA, Goldstein SA, Parfitt AM, Jesion G, Kleerek-
shown that the mechanical properties of trabecular bone oper M (1989) The direct examination of three-dimensional
can be predicted using microstructural large-scale finite bone architecture in vitro by computed tomography. J Bone
element analyses (FEA) [47,48,49]. The application of Miner Res 4:3–11
19. Bonse U, Busch F, Gunnewig O, Beckmann F, Pahl R,
such simulations in combination with FEA of non- Delling G, Hahn M, Graeff W (1994) 3D computed X-ray
destructively assessed trabecular bone structures will tomography of human cancellous bone at 8 microns spatial
help to understand the influence of age-related bone loss and 10(-4) energy resolution. Bone Miner 25:25–38
on the mechanical behavior of trabecular bone and will 20. Kinney JH, Lane NE, Haupt DL (1995) In vivo, three-
also be used to validate similar findings from follow-up dimensional microscopy of trabecular bone. J Bone Miner Res
10:264–270
patient measurements in vivo. 21. Rüegsegger P, Koller B, Müller R (1996) A microtomographic
system for the nondestructive evaluation of bone architecture.
Acknowledgements The author gratefully acknowledges support Calcif Tissue Int 58:24–29
from the Swiss National Science Foundation (no. 823A-043040). 22. Dequeker J (1994) Assessment of quality of bone in osteopo-
The author would also like to thank Jennifer Barragan for her help rosis—BIOMED I: fundamental study of relevant bone. Clin
with the coding of the simulated bone atrophy algorithm. Rheumatol 13:7–12
23. Müller R, Rüegsegger P (1997) Micro-tomographic imaging for
the nondestructive evaluation of trabecular bone architecture.
Stud Health Technol Inform 40:61–79
24. Thomsen JS, Mosekilde L, Boyce RW, Mosekilde E (1994)
References Stochastic simulation of vertebral trabecular bone remodeling.
Bone 15:655–66
1. Melton LJ, 3rd, Chrischilles EA, Cooper C, Lane AW, Riggs 25. Müller R, Hildebrand T, Rüegsegger P (1994) Non-invasive
BL (1992) Perspective. How many women have osteoporosis? J bone biopsy: a new method to analyse and display the three-
Bone Miner Res 7:1005–1010 dimensional structure of trabecular bone. Phys Med Biol
2. Kleerekoper M, Villanueva AR, Stanciu J, Rao DS, Parfitt AM 39:145–164
(1985) The role of three-dimensional trabecular microstructure 26. Müller R, Rüegsegger P (1996) Analysis of mechanical prop-
in the pathogenesis of vertebral compression fractures. Calcif erties of cancellous bone under conditions of simulated bone
Tissue Int 37:594–597 atrophy. J Biomech 29:1053–1060
3. Kanis JA (1994) Osteoporosis and its consequences Osteopo- 27. Mosekilde L (1988) Age-related changes in vertebral trabecular
rosis. Blackwell Science, Oxford, pp 1–21 bone architecture—assessed by a new method. Bone 9:247–250
4. Ray NF, Chan JK, Thamer M, Melton LJ, 3rd (1997) Medical 28. Parfitt AM (1992) Implications of architecture for the patho-
expenditures for the treatment of osteoporotic fractures in the genesis and prevention of vertebral fracture. Bone 13:S41–47
United States in 1995: report from the National Osteoporosis 29. Frost HM 1969 Tetracycline-based histological analysis of
Foundation. J Bone Miner Res 12:24–35 bone remodeling. Calcif Tissue Res 3:211–237
5. Wolff J (1892) Das Gesetz der Transformation der Knochen 30. Müller R, Hayes WC (1997) Biomechanical competence of
(The law of bone remodelling). Springer-Verlag, Berlin, Ger- microstructural bone in the progress of adaptive bone
many remodeling. In: Bonse U (ed) Developments in X-ray
6. Cheal EJ, Snyder BD, Nunamaker DM, Hayes WC (1987) tomography, vol. 3149. San Diego, California, pp 69–81
Trabecular bone remodeling around smooth and porous im- 31. Young R, May H, Murphy S, Grey C, Compston JE (1996)
plants in an equine patellar model. J Biomech 20:1121–1134 Rates of bone loss in peri- and postmenopausal women: a 4-
7. Hart RT, Davy DT, Heiple KG (1984) Mathematical modeling year, prospective, population-based study. Clin Sci (Lond)
and numerical solutions for functionally dependent bone 91:307–312
remodeling. Calcif Tissue Int 36:S104–109 32. Eriksen EF, Langdahl B (1995) Bone remodeling and its con-
8. Fyhrie DP, Carter DR (1986) A unifying principle relating sequences for bone structure. In: Odgaard A, Weinans H (eds)
stress to trabecular bone morphology. J Orthop Res 4:304–317 Bone structure and remodeling, recent advances in human
9. Carter DR (1987) Mechanical loading history and skeletal biology, vol. 2. World Scientific, Singapore, pp 25–36
biology. J Biomech 20:1095–1109 33. Brockstedt H, Kassem M, Eriksen EF, Mosekilde L, Melsen F
10. Carter DR, Orr TE, Fyhrie DP (1989) Relationships between (1993) Age- and sex-related changes in iliac cortical bone mass
loading history and femoral cancellous bone architecture. and remodeling. Bone 14:681–691
J Biomech 22:231–244 34. Schenk RK, Olah AJ (1980) Histomorphometrie. In: Kuhlen-
11. Huiskes R, Weinans H, Grootenboer HJ, Dalstra M, Fudala B, cordt R, Bartelheimer H (eds) Hanbuch der inneren Medizin
Slooff TJ (1987) Adaptive bone-remodeling theory applied to VI/a!, Knochen, Gelenke, Muskeln. Springer-Verlag, Berlin,
prosthetic-design analysis. J Biomech 20:1135–1150 pp 437–494
12. Cowin SC, Moss-Salentijn L, Moss ML (1991) Candidates for 35. Parfitt AM, Drezner MK, Glorieux FH, Kanis JA, Malluche
the mechanosensory system in bone. J Biomech Eng 113:191– H, Meunier PJ, Ott SM, Recker RR (1987) Bone histo-
197 morphometry: standardization of nomenclature, symbols, and
13. Weinans H, Huiskes R, Grootenboer HJ (1992) The behavior units. Report of the ASBMR Histomorphometry Nomencla-
of adaptive bone-remodeling simulation models. J Biomech ture Committee. J Bone Miner Res 2:595–610
25:1425–1441 36. Hildebrand T, Laib A, Müller R, Dequeker J, Rüegsegger P
14. Huiskes R, Ruimerman R, van Lenthe GH, Janssen JD (2000) (1999) Direct three-dimensional morphometric analysis of
Effects of mechanical forces on maintenance and adaptation of human cancellous bone: microstructural data from spine,
form in trabecular bone. Nature 405:704–706 femur, iliac crest, and calcaneus. J Bone Miner Res 14:1167–
15. Ruimerman R, Van Rietbergen B, Hilbers P, Huiskes R (2003) 1174
A 3-dimensional computer model to simulate trabecular bone 37. Goulet RW, Goldstein SA, Ciarelli MJ, Kuhn JL, Brown MB,
metabolism. Biorheology 40:315–320 Feldkamp LA (1994) The relationship between the structural
16. Frost HM (1964) The laws of bone structure. Charles C. and orthogonal compressive properties of trabecular bone.
Thomas, Springfield J Biomech 27:375–389
 S35

38. Simmons CA, Hipp JA (1997) Method-based differences in the 44. Eriksen EF, Mosekilde L, Melsen F (1985) Trabecular bone
automated analysis of the three- dimensional morphology of resorption depth decreases with age: differences between nor-
trabecular bone. J Bone Miner Res 12:942–947 mal males and females. Bone 6:141–146
39. Hildebrand T, Rüegsegger P (1997) A new method for the 45. Müller R, Hildebrand T, Häuselmann HJ, Rüegsegger P (1996)
model independent assessment of thickness in three-dimen- In vivo reproducibility of three-dimensional structural proper-
sional images. J Microsc 185:67–75 ties of noninvasive bone biopsies using 3D-pQCT. J Bone
40. Goldak JA, Xinhua Y, Knight A, Lingxian D (1991) Con- Miner Res 11:1745–1750
structing discrete medial axis of 3-D objects. Int J Comput 46. Müller R (2002) The Zurich experience: one decade of three-
Geom Appl 1:327–339 dimensional high-resolution computed tomography. Top Magn
41. Hildebrand T, Rüegsegger P (1997) Quantification of bone Reson Imaging 13:307–322
microarchitecture with the structure model index. Comput 47. Van Rietbergen B, Weinans H, Huiskes R, Odgaard A (1995) A
Methods Biomech Biomed Eng 1:15–23 new method to determine trabecular bone elastic properties and
42. Harrigan TP, Mann RW (1984) Characterization of micro- loading using micromechanical finite-element models. J Bio-
structural anisotropy in orthotropic materials using a second mech 28:69–81
rank tensor. J Mater Sci 19:761–767 48. Müller R, Rüegsegger P (1995) Three-dimensional finite ele-
43. Birkenhager-Frenkel DH, Courpron P, Hupscher EA, Cler- ment modelling of non-invasively assessed trabecular bone
monts E, Coutinho MF, Schmitz PI, Meunier PJ (1988) Age- structures. Med Eng Phys 17:126–133
related changes in cancellous bone structure. A two-dimen- 49. Boyd SK, Müller R, Zernicke RF (2002) Mechanical and
sional study in the transiliac and iliac crest biopsy sites. Bone architectural bone adaptation in early stage experimental
Miner 4:197–216 osteoarthritis. J Bone Miner Res 17:687–694