f

Morphometric evaluations of
personalised 3D reconstructions and
geometric models of the human spine
C..i~. Aubin 1-t J, Dansereau 1"2 F. Parent 1 H. Labelle z
J. A. de Guise 2"3
1Department of Mechanical Engineering, I~cole Polytechnique, PO Box 6079, Station 'Centre-Ville',
Montr6al, Quebec, Canada, H3C 3A7
2Research Center, Sainte-Justine Hospital, 3175 COte Sainte-Catherine Road, Montr6al, Quabec,
Canada H3T 1C5
aDepartment of Automated Production Engineering, t~cole de Technologic Sup6rieure,
1110 Notre-Dame 0., Montreal, Quabec, Canada, H3C 1K3

Abstract--In the past, several techniques have been developed to study and analyse the
3D characteristics of the human spine: multi-view radiographic or biplanar 3D reconstruc-
tions, CT-scan 3/9 reconstructions and geometric models. Extensive evaluations of three
of these techniques that are routinely used at Sainte-Justine Hospital (Montreal, Canada)
are presented. The accuracy of these methods is assessed by comparing them with
precise measurements made with a coordinate measuring machine on 17 thoracic and
lumbar vertebrae (T1-LS) extracted from a normal cadaveric spine specimen. Multi-view
radiographic 3D reconstructions are evaluated for different combinations of X-ray views:
lateral (LA T), postere--anterior with normal incidence (PAO~ and postero-anterier with 20~
angled down incidence (PA20~ The following accuracies are found for these reconstruc-
tions obtained from different radiographic setups: 2.1 + l.Smm for the combination with
PAO~ views, and 5.6+4.5mm for the PAO~ stereopair. Higher errors are
found, in the postero-anterior direction, especially for the PAO~-PA20~ view combination.
Pedicles are found to be the most precise landmarks. Accuracy for CT-scan 3D
reconstructions is about 1.1 +O.Smm. As for a geometric model built using a multiview
radiographic reconstruction based on six landmarks per vertebra, accuracies of about
2.6• for landmarks and 2.3• for morphometric parameters are found. The
geometric model and 3D reconstruction techniques give accurate information, at low X-ray
dose. The accuracy assessment of the techniques used to study the 3D characteristics of
the human spine is important, because it allows better and more efficient quantitative
evaluations of spinal dysfunctions and their treatments, as well as biomechanical model-
ling of the spine.

Keywords---3D reconstructions, Multi-planar X-rays, CT-scan, Geometric model, Spine,

Co-ordinate measuring machine, Digital X-rays, Vertebral morphometry
Med. Biol. Eng. Comput., 1997, 35, 611-618

1 Introduction The most common vertebral landmarks that are recon-

structed in 3D are the centre of vertebral body endplates and
TI~ SPINE is a very complex anatomical structure, the patho-
the inferior and superior bases of both pedicles. Graphical
logical deformations of which have traditionally been evalu-
representations of such reconstruction techniques are generally
ated in 2D using postero~nterior and/or lateral view X-rays.
limited to 3D points corresponding to recomlnacted anatomi-
To study and analyse spinal 3D characteristics, stereoradio-
cal landmarks, which are generally connected by straight lines
graphic or biptanar reconstruction techniques have been
to give a wire-frame representation. However, such 3D repre-
developed (KRATk~, 1975; BROWN et aL, 1976; OLSSON et
sentatious are difficult for clinicians to interpret because of the
aL, 1976; ~ M A R S H et al., 1980; DE SMs et al., 1980;
low number of reco~trueted anatomical landmarks and reeon-
STOKESet al., 1980, 1987; DANSEREAUetal., 1990; PLAMON-
.struction errors on landmark coordinates (implicit in any
DON and GAONON, 1990; H.~QUET et al., 1992).
stereoradiographic or biplanar reconstruction technique).
These representations generally result in non-realistic geome-
trical shapes of the reconstructed vertebral structxtres that are
Correspondence should be addressed to Professor Aubin.
email: aubin@grbb.potymtLca. difficult for clinicians to interpret and may generate distorted
First received 6 September 1995 and in final form 20 May 1997.
images as a resalt.
A personalised parametric model of the spine was devel-
9 IFMBE:1997 oped by the authors (DANsEARAU et al., 1993) to produce a

Medical 8t Biological Engineering & Computing November 1 ~ 7 611

more realistic representation relating to these six landmarks
(obtained using two or three X-ray views) using adapted
simple geometric primitives (elliptic cylinders, prisms etc.).
These modifications allow schematic representations of the 3D
shape of the spine and thorax that can then be visualised and
easily analysed by clinicians. However, other needs for geo-
metric refinements and more accuracy have appeared with the
development of a personalised biomechanical finite element
model of the spine and thorax (AUqSlN et aL, 1995a;
DESCRJMES et aL, 1995) to study orthopaedic or surgical
treatments of scoliosis (AUBINet al., 1995b).
A geometric model was developed (AUBrN et al., 1995a) to
refine the representation of the reconstructed geometry, without
digitising more landmarks on multi-view X-rays (which is a
very tong and tedious task in the context of routine clinics).
This geometric model is constructed by combining six land-
marks obtained from the multi-view radiographic 3D recon-
structions with serial CT-scan 3D reconstructed models, using
the 'connecting cube' algorithin of DE GUISE and MARTEL
(1988) (similar to the 'marching cube' algorithm developed by
LORENSEN et al., (1987)) of typical human dry vertebrae
(taken from a 42-year-old male cadaver). These typical refined
models of each vertebra are deformed using mathematical
transformations such as translations, rotations and scaling, in
addition to dual kriging interpolation techniques (TROCHU,
1993). The resulting geometric spinal model allows a more
realistic representation of all thoracic and lumbar vertebrae,
which is useful for visual interpretation by clinicians, calcula-
tion of geometrical clinical indices and finite element model-
ling needs.
The question of concern now, is how accurate all these
reconstructions and modelling techrdques are. Preliminary
evaluations have been partially made in the past for multi-
view radiographic 3D reconstructions (ANORt~ et aL, 1992;
1994; DANSEREAUet aL, 1993), CT-scan 3D reconstructions
(DE GUISE and MARTELet al., 1988) and geometric modelling
(AUBIN et aL, 1995a). However, the accuracy of the overall
modelling approach, as well as its individual techniques
(multi-view radiographic and CT-scan reconstructions), is
still not well documented. There is also a need to know the
best combination of X-ray views (ANDR~ et al., 1994) to
minimise the X-ray dose and the time needed for digitisation
and to maximise the accuracy of the 3D multi-view radio-
graphic reconstruction and the geometric model.
To address these questions and problems, this paper
presents complete in vitro morphometric evaluations and
accuracy assessments of multi-view radiographic 3D recon-
structions, CT-scan 3D reconstructions and geometric models
of the normal non-pathological spine. The overall purpose of
such evaluations is to establish optimum methods to document
spine geometry.
2 M a t e r i a l and m e t h o d s
All evaluations and validations presented in this paper were
made on 17 vertebrae of a non-pathological thoraeo-lumbar
(T I-L5) dry cadaveric human spine presenting no specific
deformities. The general surface aspects of this spine were
quite good, despite the presence of some porosity and a few
osteophytes. Age, sex, weight and height of the specimen were
not known.
The spine was fixed in a radio-transparent acrylic frame that
was designed to immobilise the specimen during most acqui-
sitions and that allowed all measurements to be taken easily
(Fig. 1). The global coordinate system (which will be called
the 'reference coordinate system') adopted by the Scoliosis
Research Society (STOKES et aL, 1994) was used. This
612 Medical & Biological Engineering & Computing
coordinate system is defined in such a way that the x-, y-
and z-axes represent, respectively, the anterior, left and cranial
directions of the spine.
2.1 3D reference measurements
Precise measurements of all cadaveric vertebrae were
obtained using a 3D coordinate measuring machine*, the
accuracy of which is close to 0.1 ram. These measurements
represent the 'gold standard' and were used as a reference to
assess the reconstruction techniques and the geometric model.
The acquisitions were made in two different steps.
The complete spine was first fixed in the acrylic frame (Fig.
1). 10 identification landmarks were put on each vertebra and
digitised using the coordinate measuring machine (CMM).
This allows the 3D position of each vertebra in the reference
coordinate system to be determined.
In the second step (after all the other acquisitions presented
in the following Sections) more acquisitions were made
on each isolated vertebra fixed in a smaller franle. In
addition to the same ten identification landmarks, 250--350
more points (depending on vertebra level and size) were
measured on the vertebral body (100-150 points), pedicles
(about 20 points each), articular facets (13 points each),
spinous and costo-transverse processes (about 40 points) and
costo-vertebral and costo-transversc facets (five and nine
points, respectively). These points were required for the
following evaluations and especially to allow the comparison
of homologous information and to reduce identification errors
(see computer algorithm described in the following section).
Each vertebra was then replaced in the reference coordinate
system using a least square repositioning approach to mini-
mise 3D distances be~een the corresponding 10 identification
landmarks.
2.2 Evaluation o f multi-view radiographic 3D reconstructions
Multi-view X-rays of the complete cadaverie spine in the
acrylic frame were taken using the method developed by
DANSEREAU et al. (1990), according to the protocol routinely
used at Sainte-Justine Hospital (Montrdal, Canada). Three
Fig. 1 Cadaveric spine specimen in radio-transparent ac~.lic
frame,
*PFtl0M Probe Head mounted on a 30 digitizer G90C, LK Tool USA
inc.
November 1997
digital X-rays were taken from different angles: a conven-
tional clinical postero-anterior view (PA0~ a 20 ~ angled
down postero--anterior view (PA20 ~ and a lateral view
(LAT). 21 anatomical landmarks (Fig. 2) were selected and
digitised (superior and inferior bases of both pedictes, tips of
spinous, transverse and articular processes, centre, most ante-
rior, most posterior, left and right lateral points of each
vertebral body endplate) using a computer assisted program
on a Silicon Graphics workstation1". They were then recon-
structed in 3D with the DLT algorithm (MART_AN, 1976) using
three different combinations of X-ray views:
(a) PA0 ~ + PA20 ~
(b) PA0 ~ + L A T
(c) PA0 ~ + PA20 ~ + LAT (Fig. 3a).
The coordinates of the reconstructed landmal$:s were trans-
formed into the reference coordinate system using geometrical
transformations (translations and rotations) and a least square
algorithm to minimise spatial differences between the 21
reconstructed landmarks and the corresponding measured
landmarks on the specimen.
The positions of the reconstructed landmarks (from the
multi-view X-rays) were then compared with the 3D reference
measurements. 3D reconstruction errors were computed for
the corresponding 21 landmarks by means of Euclidean
distances and also in terms of errors in the x-, y- and z-
directions. The Euclidean distance between the landmarks was
chosen because it has a physical significance and represents
the absolute value of the difference in positions. A computer
algorithm was developed and used to minimise the errors
related to the identification of precisely the same homologous
anatomical vertebral landmarks using both the reconstruction
and the CMM methods. This algorithm determined, by linear
interpolation, the appropriate reference landmark that is the
closest to the reconstructed landmark.
2.3 Evaluation o f CT-scan 3D reconstntctions
Serial radiographic CT-scan slices of i mm thickness were
taken at each millimetre of the complete cadaveric spine in its
acrylic frame. Each slice image was then segmented, and the
vertebral outline of each slice was numerically extaacted and
superimposed to obtain the CT-scan 3D reconstruction
h i
P r
t g
U S
t_
e surements using Euclidean distances between corresponding
Fig. 2 Vertebral landmarks usedjbr validations: (a--d)superior and
inferior bases of both pedictes; (e) tips of spinous, ~ g)
transverse and (h-k) articular processes, (l, m) centre, (n, o)
most anterior, 09, q.) most posterior, (r, s) left and (t, u) right
lateral points of each vertebral body endplate
tSilicon Graphics inc., USA
Medical & Biological Engineering & Computing November 1997
Fig. 3 (a) Multi-view radiographic 3D reconstruction (six land-
marks per vertebra) using PAO~ T view combina-
tion. (b) CT-scan 3D reconstruction. (c) Geometric model ]
(built using six reconstructed landmarks per vertebra
obtained using PAtT-PA20~-LA T views)
(Fig. 3b) according to the 'connecting cube' algorithm devel-
oped by DE Guise and MARTEL (1988).
This CT-scan 3D reconstruction was then transformed in the
reference coordinate system. The same 21 landmarks (Fig. 2)
were visually identified on each CT-scan reconstructed verte-
bra and digitised using an interactive graphics computer tool
on a Silicon Graphics workstation, which allows the operator
to obtain their 3D coordinates. These landmarks were then
compared with the 3D reference measurements. CT-scan
reconstruction errors were computed using Euclidean dis-
tances between corresponding landmarks. A similar computer
algorithm to the one presented in Section 2.2 was also used to
minimise homologous landmark identification difficulties,
except for the vertebral body endplate centre landmarks,
which were placed approximately in the middle of the end-
plates.
2.4 Evaluation o f the geometric model
The geometric model 1 (Fig. 3c) was built using a hybrid
method that combines six landmarks per vertebra (superior
and inferior bases of both pedicles and endplate centres
obtained from multi-view radiographic reconstruction techni-
que with PA0~176 views) with serial CT-scan 3D
reconstructions of typical human vertebrae (AUBrN et aL,
1995a). Each of these vertebrae was deformed using geome-
trical transformations as well as interpolation and extrapola-
tion techniques (dual kriging approach with linear drift and
covariance, equivalent to 3D splines; (TROCHU, 1993)) to fit
the six reconstructed landmarks as well as possible. In this
model, the reconstructed vertebral landmarks were adjusted,
by means of a correction algorithm that has been described
elsewhere (DANSEREAU et al., 1993), to minimise digitising
and reconstruction errors.
The same 21 landmarks (Fig. 2) were extracted from the
model and were then compared with the 3D reference mea-
landmarks. A similar computer algorithm to the one presented
in Section 2.2 was also used to minimise homologous land-
mark identification difficulties. Geometrical parameters on the
vertebral body, pedicles, spinous, transverse and articular
processes (Fig. 4) were also computed and compared with
the same parameters calculated on the reference measured
vertebrae.
613
 averaged for the 17 vertebrae and merged for pedicles,
vertebral bodies, processes and articular facet tips.

3.1 Results o]" multi-view radiographic 3D reconatruction

W~
Fig. 4 Vertebral geometric parameters used for model validation
To evaluate whether the six reconstructed landmarks are
sufficient to produce an adequate model, another geometric
model (model 2) was built using the 21 reconstructed land-
marks (also obtained from the PA0~176 combina-
tion, but without the correction algorithm). Similar
comparisons of geometrical parameters were made, and a
Student's t-test was performed to compare the two geometric
models.
evaluation
The first three columns of Table 1 present, for the 21
reconstructed vertebral landmarks, 3D multi-view radio-
graphic reconstruction errors obtained from different multi-
view combinations (PA0~ ~ PA0O-LAT and PA0 ~
PA20~ With a mean error of 5.6 mm ( + 4.5 ram), the
PA0~ ~ stereopair is clearly less accurate than other
methods. Smaller mean errors were obtained for the two
combinations that include the lateral view (2.1 + 1.5 mm for
both methods). For all reconstructions, pedicles always have
the smallest mean errors, followed by vertebral body endplate
centres. Articular facet tips were always the worst recon-
structed landmarks. Errors due to the identification of exactly
the same landmarks on the X-rays and with the 3D measure-
ments were substantially reduced (up to 4 mm) by the utilisa-
tion o f the dedicated computer algorithm presented in Section
2.2.
The 3D distances (errors) between the reference and the
reconstructions were divided into components along the x-, y-
and z-directions. The absolute values of the errors in each
direction are presented because there was generally no sys-
tematic relative to random errors, except for the superior tips
of the pedicles which were ~1 mm above the real surface. For
the case of the y- and z-directions, mean errors were found to
be relatively small and of almost the same magnitude for all
combinations of X-ray. multi-views (Figs. 5b and c). Errors
were slightly higher for z than for y co-ordinates. However,
errors were generally greater in the postero-anterior direction
(x-coordinates), in particular for the PA0~-PA20 ~ combination
(Fig. 5a). Introduction of the lateral view allows an improve-
ment in reconstruction accuracy in this direction.

3.2 Results o f CT-scan 3D reconstruction evaluation

3 Results
Table 1 presents 3D errors for multi-view radiographic and
CT-sean 3D reconstructions as well as for geometric model I.
3D errors for the 21 vertebral landmarks were evaluated by
means of euclidean distances between reference measurements
and reconstructed or model landmarks. These errors were
The fourth column of Table 1 shows, for the same recon-
structed vertebral landmarks, 3D CT-scan computed recon-
struction errors that are relatively homogeneous for the
different landmarks, except for the eentres of vertebral end-
plates where higher errors were found. These results were
found to be about the same in the x-, y- and zMirections (not
shown in Table l).

Table 1 Three-dimensional mean errors (and standard deviations) obtained.from multi-view radiographic and C'T-scan 3D reconstructions
and from model I for the 17 thoracic and lumbar vertebrae and for different vertebral landmarks (absolute values in ram).
Multi-view radiographic 3D CT-scan 3D Geometric
rcconstmction reeonsm~tion model 1
meau c . o ~ (standazd dcvi~ons)

V~'tebral landmarks PA0~ ~ PA0~ PA0*-PA20*-LAT (standard deviations) (standard deviations)

Pediclcs 2.0 1.2 1.2 0,8 1.7
U,I) (0.7) (O7) (0.5) (0.9)
Ve~brat body:
endplate ccntres 5.0 1.5 1.5 2.4 2.0
(2.8) (0.7) (0.7) (0.6) (1.0)
eotm~txr 6.0 2.0 2.0 0.9 2.3
(4.3) 0.2) (1.2) (0.7) (1.0)
Spinmm and 5.8 2.4 2.4 1.4 4.9
Wansverve processes (4.1) (1.5) (1.5) (1.1) (3.8)
Tips of articular 5.8 3.1 3.1 I. 1 3.8
facets (4.1) (1.9) (1.9) (0.8) (2.2)
Global 5.6 2. I 2. I 1.1 2.6
(4.5) 0.5) 0.5) (0.8) (2.4)

614 Medical & Biodogieal Engineering & Computing Ncammber 1997

3.3 Results of geometric model evaluation and ~ 2.1 4-1.5 ram. These errors are smaller than those of
other published studies carried out o n vertebral landmark
The last column of Table 1 presents model 1 3D errors for
reconstructions with conventional X-ray technology (ANDRE
the same vertebral landmarks that were used in preceding
et al., 1992, 1994; RAB and CHAO, 1977; STOKES et al., 1987).
reconstructions. The lowest errors were found for pedicles and
Even if digital radiography is currently not widely available, it
the vertebral body (under 2.3 ram), whereas higher errors were
shows how better quality and resolution of X-ray images can
found for tips of articular facets and processes. Errors due to
improve 3D reconstruction results. These results also indicate
the identification of homologous landmarks on the geometric
that for 3D spine reconstruction, the PA20 ~ X-ray could be
model and with the 3D measurements were substantially
omitted without significantly affecting the 3D reconstruction
reduced by the utilisation of the dedicated computer algorithm
results. However, this PA20 ~ view is necessary for the 3D
presented in Section 2.2. These errors were decreased by up to
reconstruction of the rib cage (DANSERAU et al., 1990; STOKES
15 m m for the spinous process tips which were identified as
and DANSEREAU, 1988), which is not discussed in this paper.
more posterior, in one case, and more caudal, in the other case.
For multi-view radiographic reconstructions, the most accu-
Errors for model 2 were not included in this Table as
rate landmarks were found to be pedicles followed by verteb-
comparable landmarks returned by the model were the same
ral bodies. This is in agreement with results found by ANDRg
as those for multi-view radiographic reconstruction (PA0 ~
et al. (1994). As for other reconstructed landmarks, they were
PA20~
generally poorly visible on X-rays. This was particularly the
To evaluate the intrinsic morphology and geometry
case for articular facet tips and processes and may explain
obtained from models 1 and 2, parameters were computed
their large reconstruction errors. However, these landmarks
for each reference and modelled vertebra. Average differences
are not commonly reconstructed. The analysis of errors in the
are presented in Table 2. Mean differences and standard
x-, y- and z-directions (positive or negative values, not pre-
deviations are generally higher for model 1 than for model 2
sented in Fig. 5) shows no systematic errors except for the
(2.3 + 2 . 0 r a m against 1.8 4-1.6 ram). According to the Stu-
superior tips of the pedicles ( ~ I mm above the real surface),
dent's t-test, this difference was globally statistically signifi-
which implies that radiographic landmarks identified on plane
cant (p = 0.02). The lowest differences were found for pedicle
X-ray images generally correspond to vertebral surface points.
and vertebral body parameters. Orientations and sizes of
It also points out that the reported errors are related mostly to
articular facets were similarly predicted by both models
the reconstruction method.
(p > 0.26). Differences were higher for spinous and transverse
Identification of homologous landmarks between X-rays
processes.
and 3D measurements obtained using the CMM required a
computer algorithm to minimise errors that can sometimes be
as much as 4mm. This highlights the necessity of such an
algorithm for this kind of validation. In spite of this algorithm,
4 Discussion identification problems could persist for some landmarks
Validation of multi-view radiographic reconstruction meth- while this minimisation is carried out with 3D reference
ods shows that PA0~176 and PA0~ digital X- points that are picked up along vertebral edges that are not
ray combinations give the lowest errors, which are identical sharp. These 3D reference points do not cover all rounded

Table 2 Differences between vertebral parameters evaluated on models and

reference measurements (average (absolute values) for 17 thoracic and lumbar
vertebrae).
Geometric model 1 Geometric model 2
Geometric vertebral mean errors, mm mean errors, mm
parameters* (standard deviations) (standard deviations)
Pedicles'f:
Dr.a 1.2 (0.9) 0.8 (0.5)
Hr~ 2.9 (1.4) 1.4 (1.4)
%oa 2.3 ~ (2.0") 1.9~ (l.5 ~)
Vertebral body:
3em,m 1.8 (0.9) 1.3 (1.0)
6~#tm 2.5 (1.5) N/A (2.1)
Hboar 1.3 (0.8) 1.2 (0.7)
Spinous transverse processes:
Dtpj 2.2 (1.5) 1.1 (0.9)
7.8 ~ (5.6 ~) 11.6~ (5.3 ~
#~ 2.3 ~ (1.8 ~ 4.2 ~ (3.4 ~)
Dn-~t 6.2 (2.9) 4.5 (1.9)
Articular facetst++:
nr,,,~m 0.08 (0.10) 0.07 (0.06)
n,~,, 0.t7 (022) 0.14 (0.16)
H~ 2.0 (1.5) 2.2 (1.9)
Wr,c** 1.3 (0.7) 1.3 (0.8)
GlobaH"t 2.3 (2.0) t.8 (1.6)
* See Fig. 4 for legend of parameters
t Average for right and left aides
~,Average for superior and irderior tips
** Awrage for thoracic veatebrae only
~I'3D distance parameters only

Medical & Biological Engineering & Computing November 1997 615


Fig. 5 Multi-view radiographic reconstruction errors along x-, y-
and z-coordinates (absolute values). (a) x-direction (pos-
tero-anterior); (b) y-direction (left-righO; (c) z-direction
(caudal--cranial). (It) pedic; ([]) endplate centres; ([])
endplate contours; ( H ) spinous and transverse processes;
([]) articular facets
surfaces. Better results could then be expected, if sharply
defined landmarks were used, or if the minimisation could be
carried out with surface modelling of the vertebral envelope
defined using CMM digitised points.
The multi-view radiographic evaluation presented in this
paper was made on an isolated cadaveric spine, because
this is one of the most convenient methods to assess the
accuracy of the geometric model and of the two reconsmlction
techniques (it is not possible to perform such tests on living
subjects, for ethical reasons). Obviously it introduces some
limitations to the methods and conclusions, and greater errors
should be expected when it is applied in vivo to real subjects,
because of visibility attenuation due to trunk tissues, which
may increase landmark identification errors (HIrOM~SH et
al., 1980).
To address some of these limitations, four thoracic verteb-
rae were submerged in a water container of human trunk
dimensions and they were similarly X-rayed using the same
protocol. As expected, the clarity of the radiograph was
reduced. In this case, the 3D reconstruction errors were greater
but still moderate (2.94-0.6mm for the pedieles, and
1.74-0.8mm for geometric vertebral parameters, using the
PA0~ combination). For living subjects with tabs and
complex organ structures overlying the spine, the clarity of the
radiograph may be more altered, and similar effects on 3D
reconstruction errors might be expected. As for specific
applications for spinal diseases, larger errors should also be
(;16 Medical & Biological Engineering & Computing
expected for the same reasons presented above and also
because of intrinsic deformations: and spatial angularity of
vertebrae (SKALLI et al., 1995).
Mean CT-scan reconstruction errors were found to be
9,o 1.1 4- 0.8 mm, which is slightly under unpublished
preliminary evaluations (1.24- 1.1 ram) performed by the
authors, using direct measurements made with a vernier
calliper on 28 vertebrae extracted from seven human cadaveric
lumbar spines. This level of error was expected as it is in
the order of magnitude of the CT images used to generate
the 3D reconstructions. As for vertebral body centres, it
was very difficult to uniformly identify these landmarks on
CT-scan 3D reconstructions and to take homologous points
when they were not defined on sharp edges. These points were
ther~ visually identified as welI as possible on vertebral end-
plates, and may explain the higher comparison errors found for
these landmarks. Obviously, the CT reconstruction method
could not be used for the whole spine of living subjects, for
irradiation reasons, although limited segments can be recon-
structed this way. Nevertheless, this reconsm~ction method is
of particular interest when combined with a radiographic
reconstruction technique (i.e. the geometric model used in
this study, (AtrOIN et aL, (1995a)).
As for the evaluation of geometric model 1, mean errors
of about 2.6mm were found. These errors are in agreement
with preliminary evaluations made by the authors (AtmrN
et al., 1995a) on two spinal segments cr7-T8, L2-L3),
extracted from a normal cadaveric specimen different from
the one used in the present study. Some of these errors
are intrinsic to the method used for this comparison. Anato-
mical landmark identification on the reference spine and
on the CT-scan 3D reconstruction that was used to generate
the model may be quite different. This was highlighted
for spinous process tip landmarks which were identified
as too posterior in one case, and too caudal in the other
case (difference of up to 15mm). This difference was
considerably reduced by the utilisation of the computer algo-
rithm that was specially designed to take into account this
problem.
Other differences can be explained by the intrinsic geome-
try of the vertebrae used with the geometric model, which are
deformed by the kriging technique to fit the reconstructed
landmarks (control points) as well as possible. For these
landmarks, the model restores the same points and their
errors are then those of the multi-view radiographic
reconstruction technique and correction algorithm. As
for other points given by the model, their accuracy is related
to the number and position of the control points (DELORaME
et al., 1995) as welt as to the pathological deformity of
the vertebra (LANDRY et aL, 1996). Higher errors should
be expected for extrapolated landmarks (processes and
articular facets) than for interpolated points. In model 1,
all six reconstructed landmarks that were used to deform
the vertebrae were located in the anterior part of the
vertebra.
Errors of geometric model 2 were the same as those for
multi-view radiographic reconstruction (PA0~
This model was generally more accurate than model 1.
However, model 2 needs the reconstruction of 21 points per
vertebra, which could be quite a long and tedious task in the
context of routine utilisation of this approach. In addition,
some landmarks are quite difficult to identify on all X-rays (in
particular articular facets and processes (HrNMARSH et al.,
1980; A_r~R~ et aL, 1994)). However, this study did not focus
on other approaches that could use an intermodi~tte number of
landmza'ks (between 7 and 20). This could lead to a more
optimised way to maximise accuracy while minimising digi-
tising-time cost.
November 1997
 The influence of the choice of the mathematical parameters
used with the dual kriging technique (number of control
points; drift and covariance, (TROCrtU, 1993)) was not inves-
tigated. This influence is still not well-documented (DELORME
et al., 1995), and the mathematical parameters may not be the
optimum ones. Work is are in progress to investigate this
particular aspect.
5 Conclusions
Extensive evaluations of 3D personalised reconstructions
and geometric models of a normal non-pathological spine
have been made using metrological techniques. The methods
developed in this study are quite general and can be easily
adapted to determine the accuracy of any other 3D model or
reconstruction technique.
Results for multi-view radiographic 3D reconstructions
of the spine showed that the PA0~ stereopair is the
optimum method. This conclusion is important when
3D reconstructions of the spine are needed, because this
combination implies lower radiation (with digital X-rays),
time and cost levels. Pedicle tips and vertebral body endplate
centres were confirmed as the best landmarks for reconstruc-
tion.
As for CT-scan 3D reconstruction, its general accuracy was
found to be ~ 1.1 5:0.8mm and almost the same in all
directions. To our knowledge, this is one of the first morpho-
metric evaluations of the accuracy of CT-scan 3D reconstruc-
tion of the human spine.
The evaluation of geometric models built by a combination
of vertebral landmarks obtained by multi-view radiographic
reconstruction, with CT-scan reconstruction of typical
cadaveric vertebrae, shows that the ufilisation of more land-
marks allows greater accuracy but is a more 'expensive'
method (digitising-time cost). Geometric models are poten-
tially very important for clinicians and biomechanicians
because they produce quick, realistic and accurate representa-
t.ions of the spine at low cost and X-ray dose (only two X-rays
needed).
The multi-view radiographic and CT-scan 3D reconstruc-
tions as well as the geometric models of the spine, were
confirmed to be valuable tools that can be used to institute
accurate and reproducible quantitative evaluations of spinal
dysfunctions, as well as finite element modelling. ActuaUy, the
geometric model was used to produce a new-generation
personalised finite-element model of the trunk to simulate
orthotic and surgical treatments, as weU as to investigate
spinal pathologies (AUBIN et aL, 1995b). The 3D reconstruc-
tions and the geometric model were also used to assess the 3D
efficacy of the Boston brace for the treatment o f scoliosis
(LABELLE et aL, 1996). The sensitivity and accuracy o f
geometric measurements made with 3D reconstructed geome-
tries of the spine or models can now be evaluated in terms o f
the accuracies determined by this study.
Acknowledgments-- This research was funded by NSERC (Canada).
We owe special thanks to Andr~ Choinib,m and Yvan Petit for the CT-
scan reconstruction data and to Philippe Labelle for his assistance
with numerical X-ray digitisation.
References
ANDI~, B., D~a~SE~_AU, J., and L,~ELL~, H. (1992): 'Effect of
radiographic landmark identification errors on the aeem'acy of
thre~-dimensional reconstruction of the human spine," Med. BioL
Eng. Comput., 30, pp. 569-575
Medical & Biological Engineering & Computing November 1997
ANDI~, B., DANSEREAU,J., and LABELLE,H. (1994): 'Optimised
vertical stereo-base radiographic setup for the clinical three-dimen-
sional reconstruction of the human spine,' J. Biomech., 27, pp.
1023-.1035
AUBIN, C.-E., DESCRIMES, J.-L., DANSEREAU, J., SKALL1, W.,
LAVASTE,F., and LABELLE,H. (1995a): 'Mod61isation g6om6tri-
que du rachis et du thorax pour l'analyse biomdeanique par
616ments finis des d~formations scoliotiques,' Ann. Chit., 49, pp.
749--761
AUBIN, C.-E., COTE, B., DANSEREAU, J., DESCRIMES, J.-L., and
LABELLE,H. (1995b): 'Personalized evaluations and simulations
of orthotic treatments for scoliosis' 30th Annual meeting of
Seoliosis Research Society, Asheville, USA
BROWN, R. H., BURST~nq, A. H., NASH, C. L., and SCHOCK,C. C.
(1976): 'Spinal analysis using a three-dimensional radiographic
technique,' J. Biomech., 9, pp. 355-365
DANSEREAU, J., BEAUCHAMP,A., DE GUISE, J. A. and LABELLE,
H. (1990): 'Three-dimensional reconstruction of the spine and
the rib cage from stereoradiographic and imaging techniques.'
16th Conf. of Can. Soc. Mech. Eng., Toronto, Canada, 2, pp.
61-64
DANSEREAU,J., LAI3ELLE,H., and AUBIN, C.-I~. (1993): '3-D perso-
nalised parametric modeling of reconstructed scoliotic spines.' Int.
Symposium on Computer simulation in biomechanics, Montlignon,
France, boanfe. 1.6-1.9
DE GbqSE, J. A., and MARTEL,Y. (1988): '3D biomedical modeling
merging image processing and computer aided design.' IEEE Eng.
Med. Biol. Soc., 10th Annual Int. Conf., New Orleans, USA, pp.
426-427
DELORME, S., DANSEREAU, J., AUBIN, C.-t~., LABELLE, H., and
DE GUISE, J. A. (1995): 'Simulation of soliotic vertebral
deformities: influence of control points on the prediction of
the vertebral morphology. 5th International Symposium on
Computer simulations in biomechanics, Jy-v~kyl~i, Finland, pp.
24--25
DESCRIMES, J.-L., AUBIN, C.-I~., BOUDREAULT,F., SKALLI, W.,
ZELLER, R., D~NSEREAU,J., and LAVASTE,F. (1995): 'Modeling
of facet joints in a global finite element model of the spine:
mechanical aspects' in Three-dimensional analysis of spinal defor-
mities, studies in health technology and informafics (lOS Press) 15,
pp. 107-112
DE SMET, A. A., TARLTON,M. A., COOK, L. T., FRITZ, S. L., and
DWYER, S. J. III (1980): 'A radiographic method for three-dimen-
sional anlaysis of spinal configuration,' Radiology, 137, pp. 343-
348
HECQUET,J., LEGAYE,J., and S~,.~rrlN, J.-J. (1992): 'Logiciel 'Rachis
91ITM,
" .) Int. Symposmm - on 3-D Scoliotic Deformaties (Editions
de l't~cole Polytechnique and Gustav Fisher Verlag, Montreal) pp.
26--33
HINDMAItSI-I,J., LARSSON,J., and MATTSON,O. (1980): 'Analysis of
changes in the scoliotic spine using a three-dimensional radio-
graphic technique,' J. Biomex'h., 13, pp. 279--290
KRATgu V. (1975): 'Analytical X-ray photogrammetry in scoliosis,'
Photogramraem'a, 31, pp. 195--210
LABELLE, H., DANSERF~U, J., BELLEFLEUR,C., and POn'RAS, B.
(1996): ' ~ - d i m e n s i o n a l effect of the Boston brace on the
thoracic spine and rib cage,' Spine, 21, pp. 60-64
LANDRY,C., DE GUISE,J. A., DANSEREAU,J., LABELLE,H., SKALLt.
W., ZELI.~lt, IL, and LAVeS'rE, F. (1996): 'Analyse infographique
des d~formations tridirrmensionelles des vert~bres seoliotiques,
Ann. Chit., 50, p. 732
LORENSI~, W. E, and CLIVE,H. E. (1987): 'Marching cubes: a high
resolution 3D surface construction algorithm,' Computer. Graph.,
21, pp. 163--169
MARZ~, G. T. (1976): 'Rational design for close-range photogram-
merry,' Ph.D. thesis, Department of Civil Engineering, University
of Illinois at Urbana-Charnpaign, USA
OLLSON,T. H., SELVrK,G., and WlLLNER,S. (1976): "Kinematics of
spinal fusion,' Invest. Radiol., 11, pp. 202-209
~AMOt,~ON, A,, and GA~,~O~, M. (1990): 'Evaluation of Euler's
a~agles with a least squares method for the study of lumbar spine
motion,' J. Blamed. Eng., 12, pp. 143-149
R.~, G. T., azut Crt~O, E. Y. (1977): 'Verification of roemgeno-
graphic landmarks in the lumbar spine,' Spine, 2, pp. 287-293
617
SKALLI,W., LAVAS'TE,F. and DESCRIMES,L-L. (1995): 'Quantifica-
tion of three-dimensional vertebral rotations in scoliosis--What are
the true values?, Spine, 20, pp. 546-553
STOKES, I. A. F., MEDLICOTT, P. A., and WrLDER, D. G. (1980):
'Measu, ea,ent of movement in painful intervertebral joints,' Mea~
BioL Eng. Comput., 18, pp. 694-700
STORES, I. A. F., BIOALOW,L. C., and MOP,ELA,'qD, M. S. (1987):
'Three-dimensional spinal curvature in idiopathic scoliosis,' J.
Orthop. Res., 5, pp. 102--I 13
STOKES, I. A. F., and DANSEREAU,J. (1988): 'Measurements of the
three-dimensional shape of the rib cage,' J. Biomech., 21, pp. 893-
901
STORES, I. A. F. et al. (1994): 'Three-dimensional terminology of
spinal deformity,' Spine, 19, pp. 236--248
TROCHU, F. (1993): 'A contouring program based on dual kriging
interpolation,' Eng. Comput., 9, pp. 160-177
618 Medk~ll & Biological Engineering & Computing
Author's biography
Dr. Carl-Eric Aubin is currently Assistant Pro-
fcssor at Ecole Polytechnique (Montr6al, Canada)
in the Department of Mechanical Engineering,
and Associate Researcher at the Research Center
of Sainte-Justine Hospital (Montr6al, Canada).
He recieved B.Eng. and Ph.D. degrees in
mechanical engineering (biomechanics) from
the Ecole Polytechnique. Prior to taking up his
present Faculty appointment, he did a short post-
doctoral research Fellowship in surgical instrumentation modelling at
the Department of Orthopaedics and Rehabilitation of the University
of Vermont. His research interests include orthopaedic biomechanics,
geometrical and finite element modelling of the spine and thorax,
clinical evaluation and biomeebanical simulations of orthotics and
surgical correction of scoliosis deformities, the pathomechanisms of
spinal and thoracic deformities, and seating assistive technology
(rehabilitation engineering).
November 1997