Information Technology Reference
In-Depth Information
3.1 Linear Calibration Techniques
The
first 3D reconstruction techniques were based on stereo-radiography, such as
the Direct Linear Transform (DLT), which have been widely used for this appli-
cation [
2
,
10
,
12
,
44
]. These methods require the use of a large calibration object
with
fiducial markers of known 3D coordinates and their projection on the 2D
images to estimate the intrinsic parameters of the radiographic con
guration.
However, this setup makes the reconstruction algorithm vulnerable to patient
motion between the exposures by creating inconsistencies in the calibration and
patient stereo geometries, in addition to being cumbersome in a routine clinical
setup. To overcome these limitations, Cheriet et al. [
8
] proposed an explicit cali-
bration procedure to estimate the X-ray source and
film geometric con
guration
relative to the patient
s frame of reference from the content of the images (using a
non-rigid calibration vest). The general idea was to adjust the geometric parameters
that describe the radiographic setup in such a way as to minimize landmark retro-
projection errors. The algorithm used matched calibration landmarks identi
'
ed on a
pair of X-rays and approximate geometric parameters to calibrate the images, while
still requiring a calibration object to iteratively update the radiographic system
s
'
parameters until it converges to a stable state which re
ects a valid solution. These
approaches are not suitable for the new generation EOS systems (Sect.
2.2
).
fl
3.2 Non-linear Calibration Techniques
Methods have been proposed to enable the 3D reconstruction of the spine from
biplanar X-ray images without requiring a bulky object or calibration vest, thus
introducing the 3D evaluation of spinal deformities in a wide range of clinical
setups [
28
]. Still, they require an expert to manually identify and match landmarks
on the anatomy to calibrate and subsequently reconstruct a model in 3D. In fact, to
generate a 3D model of the patient
s spine from biplanar radiographs, certain points
(anatomical landmarks) on the vertebrae within the image have to be located in
order to obtain a 3D model of the scoliotic spine using a triangulation algorithm
[
12
]. Typically, this identi
'
cation is performed manually by an expert operator and
consists of locating anatomical landmarks on a coronal and sagittal radiograph
(Fig.
5
). However, it is dif
cult to accurately identify low-level primitives such as
exact points and to match them accurately on a pair of views. Thus the repeatability
of this procedure cannot be assured. Furthermore this task is time-consuming,
tedious and error-prone, and the quality of the 3D reconstruction is directly linked
with the precision of 2D localization. Panjabi discussed in detail errors that arise
when manually identifying anatomical landmarks [
48
].
Due to these pitfalls, clinical 3D assessment of the deformity during a patient
s
visit is therefore not possible. Moreover, because current self-calibration techniques
rely on single point correspondences between the biplanar images which offer
'
