The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
masked out; d) enlarged jaw area of the masked out 
X-ray image; e) corresponding DRR that will be 
masked by the validity mask from the DR image. 
As visible in figure 4-d, areas containing high information 
density with respect to the Block Entropy are preserved. This 
enables us to perform a Mutual Information based comparison 
of the remaining gray-values. 
3.3 6 DOF Alignment Computation 
Relying on two 2D transformations, only 2 of the 3 possible 
rotations for a spatial full 6 DOF correction can be calculated. 
This is because object rotation around the vector vertical to the 
two X-ray beamlines results in no rotational component on the 
flat-panels after the object is projected. The rotation only results 
in a change of the image contents (see figure 5). . 
t t 
Unrotated DRRs Rotation not 
visible in plane 
Figure 5. Original DRRs and DRRs rotated by 45° around 
different room axes 
We use a Downhill Simplex approach with 6 free parameters (3 
translations and 3 rotations) to maximize the combined Mutual 
Information between the DR and the DRR for both stereoscopic 
image pairs A and B. The single Mutual Information values are 
combined as given by equation 4: 
MI(DR a * VM A , DRR a (T) * VM A (T)) 2 + 
M1(DR B * VM B , DRR B (T) * VM B (T)) 2 (4) 
—» max 
where A = image pair for X-ray flat-panel A 
B = image pair for X-ray flat-panel B 
T = sextuplet of the free transformation parameters 
VM = validity mask for the respective image pair 
For the optimization we need to re-render each DRR 
approximately 100 times. To reduce rendering time we mask 
out pixels of the DRR. This is done according to the validity 
mask built-up from the entropy map and if applicable, a user 
defined ROI. Finally we obtain our optimal transformation T 
and can now realign the patient support equipment or the 
radiation source for treatment, e.g. as proposed by (Yue et al., 
2006). 
4. RESULTS 
To test our approach we compared the proposed algorithms to a 
reference algorithm, which uses the full image domain to 
perform 6 DOF alignment correction. The reference algorithm is 
described in (Selby et al., 2007). We did not apply a user 
defined ROI for the performed tests, to stay independent from 
human input. However, using regions of interest usually does 
not degrade the pose estimation results, except the ROI has 
been defined in inappropriate areas of the images or contains 
not enough data for a stable registration. 
A head-phantom CT series with 1 mm slice distance and a 
human pelvis-phantom CT with 0.8 mm slice distance, each 
with 0.49 mm in slice X- and Y- direction have been used to 
test the pose correction approach. The respective kilovoltage X- 
ray images where acquired with Varian PaxScan 4030R flat- 
panels (40 cm x 30 cm X-ray image size at 3200 x 2304 pixels 
resolution) from within a real treatment environment. 
The performance of the alignment estimation could be increased 
by means of 4 up to 8, depending on the number of DRR 
creations necessary (the more iterations with DRR renderings 
had to be computed, the higher was the speed-up). The DRR 
creation itself could be sped-up by approximately factor 10, by 
masking out the pixels which possess low average information 
compared to the whole image. The creation of the entropy map, 
which in contrast to the rendering and matching algorithms, was 
not optimized for a multi-processor system required several 
seconds of additional computation time. Yet the additional time 
expense could be compensated easily by the accelerated 
rendering. 
For the head dataset we achieved the same accuracy as with the 
reference algorithm. The differences between the resulting pose 
parameters were a = ± 0.2 mm for translations and a = ± 0.1° 
for rotations. 
For the pelvis phantom, containing more unwanted artifacts 
than the head data, we could increase the accuracy and 
reliability of the pose estimation and were even able to compute 
an acceptable patient alignment at a large initial translation 
error of 20 mm, where the reference algorithm failed. 
In figure 6 (left) the remaining translational alignment errors are 
shown for the 6 DOF reference algorithm (approach A) and our 
new approach (approach B). These tests have been performed 
with the human head-phantom dataset on a 2,66 GHz Dual Core 
workstation for a range of initial misalignments between 1.0 
mm and 30 mm. Concerning the remaining misalignments it is 
not obvious, which approach to prefer. However, the 
computation time can be reduced dramatically (figure 6, right) 
due to the reduced rendering efforts coming with the methods 
introduced with this paper.