Information Technology Reference
In-Depth Information
6 Conclusion
In this chapter, we presented a comprehensive framework for the three dimensional
reconstruction of spines from radiographs. This
first section included the calibration
of the X-ray scene using a data-driven approach which extracted matched features
between the biplanar images to estimate the geometric parameters of the setup. This
was followed by a hybrid statistical and image-based approach to generate a 3D
model of the spine from these calibrated X-rays.
The self-calibration approach uses high level shape primitives extracted from the
natural content of the images, such as the spine silhouettes, to establish a reliable
correspondence between the pair of X-ray images, and furthermore determine the
2D-3D relationship of the radiographic scene. Geometrical-based features were
proposed to optimize the spine correspondences on the biplanar radiographs in
order to obtain a global calibration of the acquisition system. The results con
rm
that using intensity, surface and geometrical-based components correlated with
prior knowledge information enables the segmentation of the spine
s global shape
on the X-ray images. Results have shown that these high-level primitives help to
automatically self-calibrate the radiographic setup by using representative shape
related features such as the visual hull reconstruction, and are a viable and accurate
procedure for the 3D reconstruction of the spine. The proposed automatic technique
allows generating more accurate 3D vertebra shapes compared to the manual
identi
'
cation and matching of landmarks performed by an operator based on epi-
polar geometry. Furthermore, an image-based calibration technique incorporates
information on orientation features which were previously unavailable. While the
accuracy of the method is promising for the extraction of meaningful 3D clinical
data, errors may be propagated from the segmentation of the spine silhouettes or
from patient motion between the sequential biplanar acquisitions which can affect
the convergence and
final accuracy of the geometrical parameters. Still, the
approach allows the automatic X-ray calibration for the 3D reconstruction of sco-
liotic spines, which was difficult to perform with previous methods that require a
calibration object and manually identi
ed landmarks.
The hybrid statistical and image-based 3D reconstruction approach presented in
this chapter was anchored on the statistical distribution of a scoliotic population and
automatically segment scoliotic vertebrae using 3D level set surface evolution
techniques from enhanced biplanar images. The proposed method offers a more
reliable approach to this problem by integrating statistical, image-based and mor-
phological knowledge, and therefore becomes a suitable tool for clinical assessment
of spinal deformities. The proposed approach presented a suf
cient level reliability
to correctly detect the rotation and location of scoliotic vertebrae so it can be used in
clinical trials. The method presented in this chapter generates models similar to
those obtained from manual identi
cation. However, the manual approach is a
tedious and error prone procedure and does not guaranty 100 % accuracy. Therefore
the differences exhibited in the experiment may come from the identi
cation errors
provided from the manual landmarking. Furthermore, the proposed framework uses
