Information Technology Reference
In-Depth Information
In order to better analyze spine shapes and potentially compare different groups
of patients to
find commonalities, one has to aggregate the data from a large number
of patients and create statistical shape models. These models can then be used in a
variety of ways, but we will concentrate on two important classes of applications in
this chapter.
First, statistical shape models can be used in a descriptive fashion. In this
context, they are used to describe, visualize, or summarize a large number of
complex 3D models. They enhance healthcare workers
under-
standing of how the spine shape varies for different groups of people and allow
them to act accordingly. As an example, one can compare the spine shape of
patients with and without a brace to slow down the progression of scoliosis. It is
then possible to
or researchers
'
'
fine-tune the brace itself based on the comparison results.
Second, statistical shape models can be used to assist in image analysis tasks.
Tasks such as 3D model reconstruction, registration, or the labeling of anatomical
parts can be tedious to perform manually and dif
cult to perform automatically
without a prior shape model. Because they implicitly encode what constitute a valid
spine model, statistical shape models can be used to constrain the possible solutions.
This reduces the solution space for image analysis algorithms, which translates into
better accuracy or faster algorithms that necessitate fewer human interventions.
There are several ways to represent the shape of the spine. Therefore, there are
several ways to create statistical shape models of the spine. One possibility is to
have clinicians derive clinically relevant indices from the 3D models and then
compute statistics based on these indices. This avenue was exploited in the context
of studies on scoliosis. Several studies [
11
13
,
30
,
41
] examined the variations in
the clinical indices used by physicians to quantity the severity of the deformations.
These indices have the advantage of enabling physicians to quickly and easily
assess the severity of the scoliosis. However, they also present many problems.
First, most clinical indices are global to the whole spine, and thus do not provide
spatial insight about the local geometry. Furthermore, most of the indices (including
the Cobb angle) are computed using 2D projections, where a signi
-
cant part of the
curvature could be hidden (since the deformity is three-dimensional). In addition,
they describe the characteristics of the deformation, rather than the shape itself. It is
usually not possible to move backward from clinical indices and compute a 3D
model that could be compared with radiographs. This situation makes clinical
indices far from being ideal for assisting image analysis algorithms. Finally, clinical
indices are created based on the experience of clinical practitioners in order to
describe certain types of deformations. Thus, clinical indices that are relevant to one
pathology may be useless for another.
An alternative to clinical indices is to describe the geometry of the spine directly
and not indirectly via the characteristics of a pathology. For instance, one can build
statistical shape models based on a collection of prede
ned 3D points located on the
spine. This idea is attractive because conventional multivariate probabilities and
statistics can be leveraged to analyze and use the data. Thus, there are a large
