801
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
CT series
The first step of finding the 6 DOF patient alignment is to
perform an initial alignment (pre-alignment in figure 1). In this
pre-alignment it is not possible to find the alignment in full 6
degrees of freedom, but the CT series has to be projected into
the digital flat-panel (DFP) plane only a few times.
However, through the pre-alignment we gain a good approach
to the real patient pose. Next, we identify portions of the DR
images that can be excluded from further image comparison.
We can identify these regions in the planes of the DRR images
as well. They are then excluded from the rendering process, to
speed-up the computation.
In the last step, the full 6 DOF alignment, we fine-tune the
initially detected pose to achieve the desired accuracy (here 0.5
mm). Therefore, image comparisons between DRRs for
different alignments are performed and the similarity is
maximized. As this procedure requires a large number of
consecutive CT projections, it benefits from the fact, that areas
could be excluded from the rendering process.
3.1 Pre-alignment
To find the patient alignment we first perform a step that we call
pre-alignment. This is done by 2D registrations of two X-ray
images to the respective DRRs. The results of the registrations
are inversely projected into 3D space and used to update the
DRRs with the new alignment.
3.1.1 The Registration Process: There exists a wide range
of gray-value based image comparators in the scope of
registration. As methods like cross-correlation or usage of
difference images are not applicable for images that differ in
much more aspects than contrast and intensity, we decided to
use Mutual Information (MI) as image similarity measure
(PLUIM et al., 2003).
A joint histogram is built-up by reading the gray-values of both
images at the position of two overlaid pixels. A cell of the two-
dimensional histogram is then incremented by one at the
respective coordinates, defined by the two gray-values. The
Mutual Information value MI is calculated by equation 1:
MI(R,F) = H( < R) + H( < F)-H{R,F) (1)
where M1(R,F) = Mutual Information value
R, F = reference (DR) and floating image (DRR)
H(R), H(F) = Shannon Entropies of the images
H(R,F) = Joint Entropy of R and F
The negative MI value is minimized by a Downhill Simplex
minimization algorithm as described in (Press et al., 1982). The
three free transformation parameters are the floating image
shifts in X- and Y- direction of the image plane and rotation of
the image plane around its normal vector.
3.1.2 Inverse Projection: After each registration, the results
are back-projected into a common coordinate system. The
underlying geometry is shown in figure 2.
Figure 2. Geometry of the treatment equipment
Figure 2 depicts only the relevant parts of the equipment. The
image detectors and the X-ray tubes determine the geometric
properties that are of essential importance for the DRR
rendering and for the inverse projection of the registration
results. The patient table determines the coordinate system used
for patient alignment.
3.1.3 DRR Update and User ROIs: The DRRs are created
by ray-tracing. When scattering is neglected, the intensity of an
X-ray passing through the respective object is given by the line
integral along the virtual X-ray:
- \f(x,mx+b)ix
I = I 0 *e ■“ (2)
where I 0 - intensity of the X-ray at the source
/ = intensity of the DRR gay-value
we choose I 0 to normalize the expression in equation 2 to a
resulting intensity range of