Articulated Statistical Shape Models of the Spine - Spinal Imaging and Image Analysis - page 321

models, as well as their accuracy will vary. However, clinical imperatives are likely
to decide the imaging modality and posture of the patient.
Volumetric modalities such as computed tomography (CT), magnetic resonance
(MR), and even 3D ultrasound can be processed by registering template vertebral
models, manually identifying a prede

ned set of anatomical landmarks, or com-
puting spine-based coordinate systems as part of a curve planar reformation
approach [ 42 ]. The vast majority of CT or MR scanners require the patient to lie
down. Therefore, one has to be very conscious that the variability encoded in the
articulated statistical model will represent the anatomical variations for this par-
ticular posture.
Traditional bidimensional modalities can also be used to recreate articulated
models of the spine. More than one image will generally be necessary to remove the
ambiguities that are inherent to 2D images. The most common case in this category
is the reconstruction of 3D models of the spine from multiple radiographs (see
Chap. 5 of this topic for an in-depth discussion). In essence, a human expert
provides a computer program with indications about the matching locations in the
different radiographs. A three-dimensional model can then be reconstructed by
performing triangulation on the matched image coordinates. The way the human
expert interacts with the computer program and how the computer program uses
these interactions to perform matching varies signi cantly from one system to

another.
Many methods that use an articulated statistical shape model can be used to
extract new articulated models with relative ease. For instance, Klinder et al. [ 25 ]
and Rasoulian et al. [ 35 ] used articulated models as part of segmentation methods
of the spine applied to CT (computed tomography). The three-dimensional recon-
struction of the spine from multiple radiographs was also performed with the help
of an articulated statistical shape model [ 4 , 29 ], and even ultrasound segmentation
has been considered for an articulated biomechanical shape prior [ 20 ].
In order to illustrate the different concepts presented in this chapter and dem-
onstrate the use of certain techniques, we will use a database of approximately 300
scoliotic patients that was collected at the Sainte-Justine Hospital (Montreal,
Canada). These cases were all examined using stereo-radiographs of the spine. Six
anatomical landmarks were manually identi ed by a skilled technician (the same

landmarks presented in Fig. 1 ) on each vertebra from T1 ( (first thoracic vertebra) to

L5 (last lumbar vertebra) on the two radiographs (a posterior-anterior and a lateral
radiograph). Then, the 3D coordinates of the landmarks were computed using a
triangulation procedure. The accuracy of this method was previously established to
be 2.6 mm [ 1 ]. Once the landmarks are reconstructed in 3D, each vertebra can be
rigidly registered to its upper neighbor. The resulting rigid transformations can then
be used to build the articulated representation.