Information Technology Reference
In-Depth Information
spinous processes or on the corners of the vertebral body may be added to obtain a
more re
ned and detailed geometry of the 3D vertebrae by deforming generic
models using an epipolar geometry [
45
]. However, it is dif
cult to identify with
precision low-level primitives such as exact points and to match them accurately on
a pair of views. Thus the repeatability of this procedure can not be assured. As
discussed in Sect.
3.2
, the manual identi
cation of landmarks is a long and complex
process, which cannot ensure repeatability over multiple raters.
4.2 Contour-Based Methods
Local correspondence also relies on the assumption that a point on an object surface
appears the same in the biplanar images in which it is visible. However due to the
intrinsic properties exhibited by the X-ray modality, local correspondence is not
necessarily a reliable feature for 3D bone reconstruction. Non-stereo corresponding
contours (NSCC) methods have been proposed for the 3D reconstruction of ana-
tomical objects demonstrating few corresponding features on the biplanar X-rays,
such as for long bones (femur) or the pelvis [
38
]. The approach optimizes 3D
deformations of prior models by minimizing the object
s projection from manually
identified 2D contours. These techniques demonstrated promising results but are
still limited to the manual identi
'
cation of curves along the edges of long bony
structures. For these reasons, variational methods with a region-based component
have been applied to multi-view stereo reconstruction as an alternative to local
correspondence [
65
]. Unlike local correspondence, there is no matching of points
between pairs of images for consistency, but instead the comparison is integrated
over regions. This can not only improve the precision of reconstruction results
by incorporating additional data such as high-level corresponding geometrical
primitives (curves, surfaces), but reduces manual intervention required to identify
speci
c anatomical landmarks on the X-rays.
4.3 Statistical Methods
In order to reduce inaccuracies on the 2D localization of landmarks and to be a
clinically useful procedure, previous studies were conducted to propose more
automated methods. Initial attempts were based on vertebral template matching
[
46
,
56
], and feature-based by using active shape models (ASM) [
57
] or Hough
transforms [
23
,
64
], to detect dominant characteristics (corners, edges) from the
vertebral shape body. Still, these techniques were ineffective towards noise and
varying appearance in shape. Statistical shape models, and more recently 2D-3D
registration methods, have been the focus of increased attention for the 3D
reconstruction of the human spine. A variety of methods have been proposed in the
previous years for image to physical space (patient) registration. While some have
