The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
800 
from within the radiation device reflect the real patient 
alignment. 
During the time consuming and error prone manual procedure 
the alignment of the respective body region may change, which 
leads to degradation of the treatment results. Furthermore, 
manual alignment cannot be done for six degrees of freedom 
(DOF), because rotational misalignments can hardly be detected 
and quantified accordingly by visual evaluation of the 2D 
images. For this reasons automatic patient alignment procedures 
are necessary. 
For automatic alignment correction of patients in particle 
radiation treatment devices we register two stereoscopic digital 
radiographic images (DRs) taken from within the treatment 
device from different viewing angles with the respective 
projections of a high-resolution reference CT series, the 
planning CT. The projections are computed for an initial 
treatment set-up, starting with the expected patient alignment. 
The results of the rigid registration are then back-projected into 
the coordinate system of the patient table. The resulting 
correction vector in 6 degrees of freedom is used to move the 
table and to bring the patients tumor in the correct position for 
radiation treatment. 
Two major problems occur when computing the patients pose in 
6 DOF. One is that the rotation around the axis perpendicular to 
both central rays of the X-ray equipment axes (which is in most 
cases the table roll axis) cannot be computed directly from the 
2D projections. This is because these rotations do not lead to 
detectable movements of the contents of the 2D images, but to 
implicit changes of the images, which cannot be interpreted by a 
registration process that relies purely on 2D images. The 
solution to this problem is to maximize the image similarity 
between the DRs and new projections of the CT, depending on 
the free parameters for the 6 degrees of freedom patient 
alignment (3 translations and 3 rotations). This approach 
implies a large number of CT projections, actually done by ray 
tracing and therewith leads to high calculation times, even with 
optimized rendering techniques. 
The other problem is that the comparison of the images suffers 
from image contents, that for example are present in the DRs, 
but not in the DRRs, e.g. parts of the patient fixation 
equipment. 
To improve performance of the 6 DOF alignment detection and 
to reduce the influence of inherent deviations of the image 
contents on the registration process, we propose a modified 
approach, relying only on parts of the respective X-ray images 
and the CT scan. Therefore we initially perform a 5 DOF 
correction to gain a good estimation of the patient pose. Then 
we find regions in the X-ray images, which are expected to lead 
to stable and reliable registration results. All other regions are 
excluded from the DRR rendering process as well as from the 
image similarity maximization. This allows us to reduce 
computation time and to enhance the reliability of the pose 
estimation process. 
2. RELATED WORK 
In (Jeongtae et al., 2001) it is already shown, that mutual 
information is a suitable measure to find an estimation of the 
patient set-up error in radiotherapy. We make use of this 
measure to compare X-ray images with the respective 
reconstructed radiographs. However, to gain a full 6 DOF 
alignment, a large number of DRRs have to be rendered and 
must be compared to the X-ray images. This is normally done at 
high costs of computation time and reduces the acceptance of 
the full 6 DOF alignment in clinical applications. 
One suggested solution to this problem is given in (Birkfellner 
et ah, 2003). They propose to perform several 2D to 2D image 
registrations between DR images and DRRs. The resulting 3D 
transformation is then computed by inverse projection. The 
DRRs are updated as soon as the hypothetical 3D 
transformation of the patient, and therewith the CT scan, 
reaches a certain amount of translation or rotation. The 
complete pose estimation process can be sped-up by means of 
factor l A at the cost of some tenths of a degree in rotation 
accuracy. However, only ±1.6 millimetres accuracy could be 
reached in translation. 
In (Selby et ah, 2008) full 6 DOF alignment correction is shown 
for a single X-ray image and a single DRR image. Using only a 
single X-ray image could reduce rendering time by 'A compared 
to the stereoscopic approach, but suffers from low translational 
accuracy in direction of the X-ray axis (axis from the X-ray 
source to the centre of the digital flat-panel). 
(Rohlfinga et ah, 2004) propose what they call progressive 
attenuation fields to speed-up the rendering process for pose 
estimation. As many DRRs with only slight deviations have to 
be created, each ray through the volume is computed only one 
time. Once computed, each ray represented by the result of a 
line integral is stored in a hash table and can be reused. To 
reduce the number of stored rays, interpolation is applied. 
Through this approach the pose detection could be sped-up by 
factor Vi compared to the same algorithm using standard ray 
casting to render the DRRs. Unfortunately, accuracy and 
reliability suffer from large initial patient set-up errors and with 
16 mm initial misalignment, only about 30% of the tested cases 
led to correct results. 
For our approach we aim to achieve reliable results for at least 
20 mm of initial misalignment. Thus we perform an initial 
5+1D pose detection similar as in (Birkfellner et al., 2003), 
which is based on 2D to 2D image registrations with updating 
the DRRs at certain steps of the process, to reflect the real 
alignment of the patent relative to the X-ray equipment. This is 
done until the transformation cannot be further optimized. 
To enhance accuracy in 6 DOF we then perform automatic 
image comparisons with new DRRs, rendered for each tested 
3D patient pose. This is done as described in (Selby et al., 
2008), but we use two X-ray images to ensure an acceptable 3D 
accuracy. To reduce rendering time and the influence of certain 
areas of the image, we select parts of the X-ray image, which 
are suitable for image matching. The DRR is ray-traced only in 
these areas and'all other parts of the image are excluded from 
the processing. 
In the selected areas, the rendering is done without reducing 
image resolution or radiometric quality to avoid degradation of 
the reachable accuracy. 
3. METHODS 
In figure 1 we first give a brief overview over the whole pose 
estimation process as performed by our approach. After that, the 
relevant working steps will be explained in more detail.