Home Search Collections Journals About Contact us My IOPscience

Registration of DRRs and portal images for verification of stereotactic body radiotherapy: a

feasibility study in lung cancer treatment

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2007 Phys. Med. Biol. 52 2157

(http://iopscience.iop.org/0031-9155/52/8/008)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 137.189.170.231
This content was downloaded on 03/10/2015 at 02:51

Please note that terms and conditions apply.

IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 52 (2007) 2157–2170 doi:10.1088/0031-9155/52/8/008

Registration of DRRs and portal images for

verification of stereotactic body radiotherapy: a
feasibility study in lung cancer treatment

Thomas Künzler1, Jozef Grezdo2, Joachim Bogner1,

Wolfgang Birkfellner3 and Dietmar Georg1
1 Department of Radiotherapy and Radiobiology, Medical University Vienna, Vienna, Austria
2 Department of Radiotherapy, St Elisabeth Institute of Oncology, Bratislava, Slovak Republic
3 Center for Biomedical Engineering and Physics, Medical University Vienna, Vienna, Austria

E-mail: thomas.kuenzler@akhwien.at

Received 31 August 2006, in final form 27 December 2006

Published 27 March 2007
Online at stacks.iop.org/PMB/52/2157

Abstract
Image guidance has become a pre-requisite for hypofractionated radiotherapy
where the applied dose per fraction is increased. Particularly in stereotactic
body radiotherapy (SBRT) for lung tumours, one has to account for set-up
errors and intrafraction tumour motion. In our feasibility study, we compared
digitally reconstructed radiographs (DRRs) of lung lesions with MV portal
images (PIs) to obtain the displacement of the tumour before irradiation. The
verification of the tumour position was performed by rigid intensity based
registration and three different merit functions such as the sum of squared
pixel intensity differences, normalized cross correlation and normalized mutual
information. The registration process then provided a translation vector that
defines the displacement of the target in order to align the tumour with the
isocentre. To evaluate the registration algorithms, 163 test images were created
and subsequently, a lung phantom containing an 8 cm3 tumour was built. In
a further step, the registration process was applied on patient data, containing
38 tumours in 113 fractions. To potentially improve registration outcome, two
filter types (histogram equalization and display equalization) were applied and
their impact on the registration process was evaluated. Generated test images
showed an increase in successful registrations when applying a histogram
equalization filter whereas the lung phantom study proved the accuracy of
the selected algorithms, i.e. deviations of the calculated translation vector for
all test algorithms were below 1 mm. For clinical patient data, successful
registrations occurred in about 59% of anterior-posterior (AP) and 46% of
lateral projections, respectively. When patients with a clinical target volume
smaller than 10 cm3 were excluded, successful registrations go up to 90% in AP
and 50% in lateral projection. In addition, a reliable identification of the tumour
position was found to be difficult for clinical target volumes at the periphery of

0031-9155/07/082157+14$30.00 © 2007 IOP Publishing Ltd Printed in the UK 2157

2158 T Künzler et al

the lung, close to backbone or diaphragm. Moreover, tumour movement during

shallow breathing strongly influences image acquisition for patient positioning.
Recapitulating, 2D/3D image registration for lung tumours is an attractive
alternative compared to conventional CT verification of the tumour position.
Nevertheless, size and location of the tumour are limiting parameters for an
accurate registration process.
(Some figures in this article are in colour only in the electronic version)

1. Introduction

The various forms of image guidance (IG) have become important verification tools for
hypofractionated radiotherapy to account for set-up errors or to monitor intra- and interfraction
tumour motion. Verification of the patient’s position can be achieved with traditional methods
such as an electronic portal imaging device (EPID), or through external reflecting markers
or a surface scanner (Verellen et al 2003, Weiss et al 2003, Bert et al 2005, Gierga et al
2005). Internal radio-opaque markers in or near the tumour are as well used for tumour
monitoring. Another option to visualize the tumour directly is ultrasound (Scarbrough et al
2006). However, current developments such as cone beam CT (CBCT) with keV and MeV
beams provide a new generation of image guidance (Pouliot et al 2005, Jaffray et al 2002,
Groh et al 2002, Ford et al 2002, Munbodh et al 2006). With these devices it is possible to
generate traditional fluoroscopic images but their particularly volumetric information offers
new possibilities for image guided and adaptive radiotherapy.
It is well known that the contrast for megavoltage (MV) portal images is low compared
to kilovoltage (kV) images. In the latter energy range the photoelectric effect dominates with
its superior dependence on atomic number compared to the Compton effect which in turn
dominates in the MV range. Nevertheless, for set-up verification a MV portal image provides
good visibility and useful contrast for bony structures. Already Gilhuijs et al (1996) registered
2D portal images to CT data ten years ago, and their automatic 3D analysis of patient set-up
was accurate and robust in about 1 mm and 1◦ for translation and rotation, respectively. In the
meantime a variety of algorithms have been tested for the registration of bony structures (Kim
et al 2001, 2005, Dekker et al 2003, Clippe et al 2003, Sharp et al 2005, Matsopoulos et al
2004, Jans et al 2006).
For lung cancer patients where the tumour is influenced by respiration motion, an extra
monitoring of the target volume movement should be the basis for precise beam delivery.
Especially in stereotactic body radiotherapy for lung lesions with fractional doses around
20 Gy, precise patient set-up and set-up verification accuracy are essential. Currently, in
many departments set-up verification for SBRT is performed with a CT prior treatment. As
long as new IG tools are not available in all clinics there is still the need for improved tumour
verification based on existing and traditionally used equipment, i.e. conventional EPID devices.
However, set-up verification for lung targets based on the bony anatomy is not useful because
the tumour can move with respect to the surroundings and it is a challenge to focus radiation
on the tumour. On the other hand, the difference in density for lung tissue (around 0.2 g cm−3)
and target volume (around 1 g cm−3) leads to an acceptable contrast for image comparison even
in the MV energy range. Focusing on the contrast of moving targets such as lung tumours, the
influence of image noise upon various registration algorithms has already been investigated
(Meyer et al 2006).
The aim of the present study was to investigate whether portal images and digitally
reconstructed radiographs can provide the basis for set-up verification for SBRT patients.
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2159

Treatment planning systems provide DRRs from CT data sets and are generated either by
a conventional raycasting method (Levoy 1990) or alternative techniques, such as lightfield
rendering (LaRose 2001), shear-warp rendering (Lacroute and Levoy 1994), or splat rendering
(Birkfellner et al 2005). Three types of algorithms for image registration, which are well
documented in the literature (Clippe et al 2003, Dekker et al 2003, Penney et al 1998),
were tested for 2D/3D registration. Because of the limited contrast in MV portal images,
segmentation of the region of interest (edge detection) often fails and cannot be used to identify
the position of the target. Our analysis of intensity based registration was performed with
and without filter and the 2D/3D registration software was validated on test images, phantom
studies and finally on patient data.

2. Materials and methods

2.1. Rigid registration

The patient shift (translation vector) in one plane is based on a 2D image registration.
Consequently, two projections from DRR and EPID images are necessary for a 3D analysis, for
example an anterior-posterior and a right-left view. A patient shift for the cranio-caudal (CC)
direction which can be extracted from the AP and lateral view may slightly differ between the
two projections. In the case of small differences, an averaged shift can be applied as long the
algorithm’s outcome seems reliable. For a more detailed discussion, see subsection 2.5 where
limits of a successful registration are defined.
In our feasibility study, we tested rigid registration algorithms which were solely based
on translations, i.e. rotations were neglected. Several aspects support that translations are
sufficient for realignment of SBRT targets before irradiation. The stereotactic body frame
limits rotation of the patient (Wulf et al 2000), targets are often spherical in SBRT and the
PTV including an internal margin already takes rotation into account.
Image registration based on image intensity is a widely used method for 2D/3D image
data in radiotherapy (Hill et al 2001, Maes et al 1997, Park et al 2004, Penney et al 2001). For
pixel intensity registration, the grey values are pixelwise compared and an algorithm calculates
a measure of similarity. Three types of algorithm were tested for our purpose. The portal
image f was compared to the DRR image g, describing pixel grey values f (i, j ) and g(i, j )
for both images. In the following, investigated algorithms are briefly described. To realize the
registration processes, a software program was implemented in Matlab (MathWorks, Natick,
USA) and results were exported to Excel (Microsoft, Redmond, USA) for further statistical
analysis.
Sum of squared differences (SSD).


SSD = [f (i, j ) − g(i, j )]2 (1)
i,j

i.e. the SSD algorithm weights the intensity difference quadratic. The larger the deviation of
pixel values, the higher the penalty for intensity comparison. This similarity measure does
not depend on the absolute pixel value, it responds equally to the same grey value difference
in dark and bright regions (Hill et al 2001).
Normalized mutual information (NMI). Using the definition of joint entropy H


H (f, g) = − pf,g (i, j ) · log[pf,g (i, j )] (2)
i,j
2160 T Künzler et al

(a) (b) (c) (d)

(e) (f)

Figure 1. Various bounding boxes of DRRs of Focal SIM all tissue ((a) and (b)), soft tissue ((c)
and (d)) and of Helax TMS ((e) and (f)).

where p is the probability of intensity which describes the frequency of an intensity pair in the
joint histogram (Maes et al 1997). The mutual information for a portal image f and a DRR
image g is then defined as
NMI(f, g) = H (f ) + H (g) − H (f, g). (3)
Maximizing the mutual information means minimizing H (f, g), being the joint entropy of
both images.
Normalized cross correlation (NCC).

i,j [f (i, j ) − f ][g(i, j ) − g]
NCC =    (4)
2 0.5
i,j [f (i, j ) − f ] i,j [g(i, j ) − g]
2

where f and g are mean values of each image.

The NCC method is a Pearson correlation coefficient and uses the pixel values relative
to the average grey level, so it responds equally to higher and lower intensity and is invariant
to changes of global brightness (Dekker et al 2003). It assumes a linear relationship between
intensity values but is vulnerable to outliers of high intensity.

2.2. DRR and portal image

Three different DRR sets, one calculated by a treatment planning system (TPS), Helax TMS
(V1.6B, Nucletron, The Netherlands) and two calculated by a virtual simulation software,
Focal SIM (V4.2, CMS, St. Louis, USA) were used as the basis for the 2D/3D registration.
With both systems, DRRs can be created by applying different default settings, illustrated in
figure 1. Computed DRRs from the TPS use the predefined window setting lung, whereas
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2161

(a) (b) (c) (d)

(e) (f)

Figure 2. Portal images of the same lung tumour patients displayed in figure 1.

for the virtual software we took default settings all tissue and soft tissue. These settings were
different in terms of windowing, as well as different in contrast and brightness adjustments.
All portal images were acquired with a standard amorphous silicon EPID device (iView
GT, Elekta, Crawley, UK); see some examples in figure 2. With patients the portal image
was created just before irradiation, using a 6 MV beam with the field size of 10 × 10 cm2
and delivering 5 monitor units. The native EPID file format was 16 bit greyscale and was
downgraded to 8 bit to have the same greyscale depth as the 8 bit DRR.
Unfiltered original images were taken for registration in a first step. Subsequently, various
image filters were applied to both, DRRs and portal images, to verify whether the robustness of
image comparison could be improved. First, histogram equalization (HE) was applied which
tries to change the greyscale-frequency histogram in a way to adjust a constant frequency to all
pixel intensities. This method is mostly used when different imaging modalities are compared
(Dekker et al 2003, Kim et al 2005). Subsequently, the portal imaging software also offers
a built-in display equalization (DE) which is a modified HE filter (iViewGT, User’s Manual,
Elekta, Crawley, UK). It identifies any part of the image that is not in the radiation field, and
discards the latter. The rest of the image is then enhanced to increase its useful dynamic range,
i.e. the greyscale values are stretched. The DE filter can be applied to all images that are
exported from the EPID software. Portal images were tested with and without DE.

2.3. Registration process

For any registration process a region of interest (ROI) needs to be selected. It defines the area
of a target image that is compared step by step to a reference image. For the lung lesions,
2162 T Künzler et al

the regions of interest were already available: the CTV and PTV contour of the TPS. For the
registration process, these contours were converted into a circumscribing rectangle and we
call this rectangle a bounding box (BB), for example PTV-BB; see figure 1. In addition to
these two contours we took a manually defined ROI to exclude high intensity regions which
often occur in lateral projections of the thorax when the backbone overlaps the target.
For 2D/3D registrations, the DRR was truncated with the previously defined bounding
boxes and the target itself was entirely within the ROI, centred around the beam axis. To
determine the new position of the tumour before irradiation, the truncated DRR was compared
to the portal image where the tumour could be located at any position in the field of view. For
that purpose, one of the three registration algorithms was applied on native or filtered images,
respectively. To minimize the time for registration, we also truncated the portal images. We
added a 20 mm margin to the ROI rectangle in all directions, so the restricted portal image
was 40 mm larger in height and width than the truncated DRR. Thus we were limited to
detect a tumour displacement of maximum 20 mm which was a reasonable compromise in
this feasibility study.
For the presentation of results, the following definitions are introduced: the reference shift
of an observed target is obtained by two CT data sets: the planning CT and the verification
CT scan shortly before irradiation (3D/3D registration). Displacement refers to differences in
the target position when comparing a DRR to the portal image by various algorithms (2D/3D
registration). On the other hand, deviation refers to the difference between the displacement
and the previously defined reference shift.

2.4. Test images and phantom study

In the presence of structures such as heart, liver or bone, the correct identification of the
tumour in the thorax becomes difficult for 2D/3D registration. To test the robustness of the
algorithms, 163 test images were created with three types of structures having different grey
values, thus mimicking a tumour, the lung and another extra tissue. They were all within a
grey value range of a clinical DRR and PI. The position of the structure tumour was set at six
different locations in reference to the extra tissue, as an example see one tumour position with
different grey values in figure 3. HE filtered and raw images were tested.
The accuracy of the investigated algorithms was then evaluated under more clinical
conditions with an in-house built lung phantom; see a 3D reconstruction in figure 4. It
consists of 6 cm Styrofoam covered with 1 cm solid water slabs on the top and the bottom
to realize a simple lung geometry with the surrounding thorax wall. Within the styrofoam a
polystyrene cube (8 cm3) is mimicking the tumour at the centre of the phantom. Portal images
in an anterior-posterior projection were taken of five different phantom offset positions up to
15 mm. We applied the algorithms and deviations in lateral and cranio-caudal directions were
recorded. The registration was rated successful if the deviation did not exceed a limit of
1 mm, as this millimetre accounts for set-up inaccuracies when positioning the lung phantom
on the treatment table.

2.5. Study based on patient data

Patient data were taken from regular patients with lung lesions, treated with SBRT at the
Medical University Vienna. All patients have been irradiated in a stereotactic body frame
(Elekta, Crawley, UK) combined with an abdominal compression if necessary (Wulf et al
2000).
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2163

Figure 3. A selection of test images, containing three structures: tumour, lung and another extra
tissue influencing the registration.

Figure 4. 3D reconstruction of the physical phantom.

In a first step, a data set A which consisted of 13 targets treated in 35 fractions was used
to test our registration set-up and to address following questions: which type of filter is ideal
for registration, to which extent does the size of the ROI influence the test results, what type
of DRR is feasible for intensity comparison and finally, which type of algorithm delivers the
most accurate results. For all fractions and targets, DRRs and orthogonal portal images were
available.
Next, the ideal combination of parameters (type of algorithm, DRR, ROI and filter
settings) was tested for a larger data set B (which included data set A) in order to determine the
2164 T Künzler et al

robustness and accuracy under clinical conditions. Data set B consisted of 38 targets treated
in 113 fractions. Again, DRRs and orthogonal portal images were available for all fractions
and targets. Finally, data set B was used to analyse whether the tumour position and size have
an influence on the accuracy of the registration. In total, more than 10 000 registrations were
performed for three algorithms, three types of DRR, three ROIs, two types of filter and an AP
and lateral view for each fraction.
2D/3D registration algorithms were evaluated under clinical conditions by comparing
the calculated translation vector to a manual CT registration performed with the BrainSCAN
system (BrainLAB, Heimstetten, Germany). For the manual registration of the tumour tissue,
the verification CT was registered to the planning CT data set and the reference shift of the
target was determined prior each SBRT fraction. This reference shift was used as a standard
for image registration.
However, to define the limit of a successful 2D/3D registration process, several factors
need to be considered. First, the manual CT registration can lead to inaccuracies up to
3 mm because of the inherent CT data resolution. Axial slice resolution of about 1 mm is
higher than resolution in the CC direction where we set a slice thickness of 4 mm. The pixel
resolution can be explained by the field of view of the axial CT slice around 50 cm and the
native DICOM resolution of 512 × 512 pixels. This observer target registration error takes
CT resolution limitations as well as intra-observer uncertainties into account. Additional
inaccuracies originate from the transportation of the patient from the CT to the linac and
associated set-up errors in stereotactic body frame positioning, which are around 3.5 mm
(Wulf et al 2000). Using image data of lung lesions it is obvious that another influence comes
from tumour movements due to breathing at the time of CT and EPID data acquisition. In
order to limit this influence, planning CT, verification CT and portal images were acquired
in shallow breathing. Breathing motion was considered by adding another uncertainty of
3 mm. This number is based on an ongoing study at our clinic where the tumour motion
range is determined in the CC or AP direction during shallow breathing. More specifically,
some of our test patients were undergoing several multislice CT scans at different moments
of the breathing cycle which provided us with information about the target position during
shallow breathing. Assuming that all these above-mentioned uncertainties can be considered
as independent and adding them in quadrature, a 5 mm limit was obtained to define a successful
2D/3D registration.

3. Results

Each registration algorithm needs its specific time for calculation, depending on the
size of the DRR and the portal image. The duration of a single registration process
ranged from a few seconds for the NCC algorithm (an optimized algorithm provided by
Matlab) and some minutes for the SSD and NMI method in the present form of code
implementation.

3.1. Test images and phantom study

When evaluating the influence of histogram equalization or display equalization, filtered
test images showed increased registration accuracy for both filter types. Whereas the NMI
algorithm already achieved a success rate of 100% for unfiltered images, SSD and NCC
improved definitely with the applied HE equalization; see table 1. Furthermore, tested
algorithms were more error-prone with targets close to an interfering structure than a detached
target.
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2165

Table 1. Increase in successful registrations for 163 generated test images when the histogram
equalization filter was applied.

Successful registrations
Algorithm Without filter (%) With HE (%)
SSD 49 69
NMI 100 100
NCC 74 78

Table 2. Phantom displacements and phantom deviations for three algorithms: SSD, NMI and
NCC.
Displacement (mm) Deviation SSD (mm) Deviation NMI (mm) Deviation NCC (mm)

LATERAL CC LATERAL CC LATERAL CC LATERAL CC

0 −15 0.1 0.3 0.3 0.7 0.1 −0.6
5 0 0.2 0.7 0.0 0.3 0.0 0.4
7 2 0.5 0.9 0.6 0.9 0.5 0.7
−10 −4 −0.1 0.2 −0.5 0.6 −0.2 0.6
−12 3 0.0 0.3 −0.3 0.7 −0.1 0.3

Lung phantom registrations were successful for the SSD, NMI and NCC algorithms, i.e.
they were within the predefined 1 mm limit. Furthermore, average deviation for AP and lateral
directions were below 0.6 mm. Table 2 presents the detailed results of a static lung phantom
with a clearly visible target.

3.2. Clinical results

The best results for patient group A were achieved with the normalized cross correlation
algorithm without applying a filter on the DRR and the EPID. Neither the HE filter nor the
DE filter improved any results compared to the unfiltered, original images. In more detail,
the DE filter was not as advantageous as the histogram equalization and was therefore no
longer considered for our testing. The DRRs created by the TPS (Helax) led to the best
registration outcome whereas the DRRs of the virtual simulation software showed almost the
same performance. Comparing various bounding boxes as a region of interest they yielded
similar mean deviations for the NCC algorithm whereas the PTV-BB delivered the best results.
Both graphs in figure 5 illustrate the results for the NCC algorithm, using the Helax DRR
and the PTV-BB as ROI. On the left hand side, the outcome is shown when using the original
images (no filter). The success rate of 58.5% for the AP projection was a bit higher than for the
lateral view which was 46.2%, respectively. The outcome for HE filtered images was worse
than without. Correct registrations with the HE filter occurred only for 55.7% and 41.5% for
AP and lateral projections. In addition, figure 5 displays deviations up to 8 mm.
Tumour deviations for data set B were averaged for patients with three or more fractions
in order to define a measure of success as illustrated in figure 6 for the AP projection.
Summarizing the results for the AP and lateral view, six out of 28 patients showed large
deviations around 10 mm and more. When patients with a CTV smaller than 10 cm3 were
excluded, successful registrations go up to 90% in AP projection and 50% in lateral projection.
In the patient group with a clinical target volume smaller than 10 cm3 the results were not
2166 T Künzler et al

Deviation for the NCC algorithm, no filter Deviation for the NCC algorithm, HE
100% 100%
90% 23 90% 19
26 29
80% 80%
70% 21 70% 28
60% 31 and larger 60% and larger
33
50% < 8 mm 50% < 8 mm
40% < 5 mm 40% < 5 mm
30% 62 30% 59
49 44
20% 20%
10% 10%
0% 0%
ap lat ap lat
(a) (b)

Figure 5. Results of 2D/3D registration for the NCC algorithm, using DRRs from HELAX.
The PTV-BB was taken as the region of interest. The left figure shows results of the original
images whereas the right illustration presents the outcome of filtered images, using the histogram
equalization. The absolute number of deviations is within the bars.

Average deviation in AP projection [mm]

25.0
20.0
15.0
10.0
5.0
0.0
0 20 40 3
60 80 100
CTV [cm ]

Figure 6. Patients with three or more fractions were used to correlate average deviations of the
NCC algorithm and the CTV. In the CTV range up to 10 cm3, the deviation between 2D/3D
registration and the reference displacement is larger.

encouraging. For very small CTVs less than 3 cm3, the planning system cannot create a DRR
with a recognizable tumour and subsequently, the registration was not possible. For the cases
with a clinical target volume between 3 cm3 and 10 cm3, registration was only successful when
the target was localized in the middle of the lung and not too close to the heart, liver, spleen,
backbone or the ribs; see figure 7.
Large targets are more robust to registration algorithms than small ones and the volume
of 10 cm3 seems to be a threshold for a good registration. For larger targets, deviations
stabilize around 5 mm. Furthermore, incorrect registrations could be correlated to the position
of the clinical target volume. For the AP projection, bad results occurred when the CTV was
localized near the liver, spleen or diaphragm. In lateral projections incorrect registration took
place if the CTV was localized in the posterior region of the lung where the influence of the
backbone disables an identification of the tumour position.

4. Discussion

For a daily use of automated registration algorithms in clinical treatments, a fast and reliable
registration process is required to correct for interfraction target movements and set-up
uncertainties. The process should take place within seconds which is achievable for optimized
algorithms nowadays, e.g. the optimized NCC algorithm. Moreover, there are some additional
parameters influencing the reliability for a registration. Particularly, the chosen ROI is a
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2167

Figure 7. Successful registrations (green) and bad registrations (red with arrow) of lung CTVs for
patients with three or more fractions. The robustness of registration depends on the location of the
target.

significant parameter that can strongly affect this process. When defining the region of interest
it is essential to include the complete target into the reference ROI. The contours of the treatment
planning system, being the basis for bounding boxes, do not necessarily cover a complete
structure. However, image completeness of target structures in DRRs was already discussed
(Lemieux et al 1994). He investigated the DRR for patient-to-computed-tomography image
registration and defined the physically realistic image, describing DRRs where rays passing
through the top and bottom slice of the CT are discarded. These non-physical rays should be
eliminated in order to prevent artefacts of truncated surfaces. Accordingly, the limitation to
the physically realistic image constrains the information content of target structures.
In our application-driven development for lung tumour detection, we tried to mitigate
bones in DRR computation. We attempted to focus on lung tumour tissue instead of bony
anatomy, applying the lung window in the TPS (Helax) and for the virtual simulation software
the setting soft tissue. We even tried to define a volume of interest (VOI) for the CT data set
where only the lesion itself was visible. This VOI was then tested as a basis for the DRR
generation in the virtual simulation. Unfortunately, VOI limited DRRs led to unsatisfying
results and the tested algorithms had severe difficulties. We assume that DRR imaging is
badly influenced when reducing the CT data to some slices.
The NMI, the most effective algorithm for synthetic data, and the NCC cost function are
based upon different approaches. Whereas NMI sorts pixel intensities for a joint histogram, the
intensity value itself is lost and frequencies of intensity pairs are the basis for further evaluation.
This correlation of two possible random variables leads to the loss of image information. NMI
is frequently applied for intermodality registration, e.g. MRI and CT and it was successful in
comparing phantom or synthetic data. But it was likely to fail in real patient situations. In
contrast to mutual information, NCC describes a relation between linear dependent variables
and the full image content is taken into account for this value. This could be an advantage of
linear correlation coefficients over statistical approaches when applied on images of the same
modality. However, the robustness of the normalized correlation coefficient is sensitive to
2168 T Künzler et al

high intensity regions because the average image intensity is the central value for computation
of NCC. Khamene et al (2006) tested a variety of registration algorithms, applied on the
Rando phantom and its bony anatomy. The averaged target registration error (TRE) of the
correlation coefficient, an equivalent to our normalized cross correlation, was around 7 mm
whereas their best algorithm, the local normalized correlation (LNC), had an averaged TRE
below 1 mm. For LNC, the image is subdivided into blocks and the correlation coefficient is
computed for each block separately and then averaged to get a scalar value as a result. For
images with high intensity areas, a global correlation coefficient is strongly interfered by bright
regions. Thus, subdividing an image into small parts and applying the correlation algorithm
could improve results significantly. But for images with less intense pixel values, e.g. in lung
tumour registration, NCC still maintains its relevance as a robust registration algorithm.
To evaluate registration algorithms, a reliable golden standard is required. For lung
tumours, this reference standard is difficult to acquire because of cyclic movements of the
tumour. Lung lesion movements have been investigated (Seppenwoolde et al 2002) and a
hysteresis as tumour path was postulated because the tumour follows a different path during
inhalation and exhalation. For a more simple consideration of the movement, the end position
of inspiration and expiration during shallow breathing can be taken to evaluate a so-called
tumour motion range. For imaging in our investigation, the patients were asked to perform
shallow breathing. Hence, the moment of acquiring an image was arbitrary and therefore,
tumour motion range had a considerable impact on the registration process.
The above-mentioned limitations of our clinical feasibility study imply further
investigations. Focusing on the tumour position during imaging for CT or MV portal images,
the deep inspiration breath hold technique (DIBH) could provide a more reproducible tumour
position (Hanley et al 1999). Imaging combined with the DIBH method allows us to reduce
intrafraction tumour motion for a more reproducible positioning of the target (Stock et al
2006).
Recent developments of CBCT will offer new possibilities and have high potential for
future IGRT techniques. Nevertheless, 2D/3D intensity value registration can be used in
clinics without a CBCT option but being equipped with a standard EPID. For these clinical
sites, the use of flat panel devices could be an alternative to the cone beam technique, of course
with some limitations. Today, the amorphous silicon EPID offers portal images with good
contrast and low noise, useful for daily geometric verification even in the 6 MeV energy range.
Moreover, acquiring EPID images is fast compared to CBCT which takes time for a complete
rotation of the gantry and the reconstruction of the volume.

5. Conclusion

For radiotherapy departments where stereotactic RT is applied to lung tumours but no resources
are available for a 3D verification of the tumour position (as CT-linac, 3D IGRT), a 2D/3D
registration concept is an attractive alternative. Its advantages are simplicity and accessibility.
For patients it shortens the time to verify the position from 35 min to several minutes, as
there is no need to acquire a CT and fuse it to reference CT data. Our method of 2D/3D
registration offers a high probability (90%) of identifying and verifying the position of the
lung tumour with a clinical target volume larger than 10 cm3 for AP projections. However, for
lateral projections the probability of correct registrations is significantly lower (50%). Another
limitation for accurate results is the influence of several organs which are not avoidable in
projection images. Further research is needed to evaluate the impact of a deep inspiration
breath hold technique on the success rate of 2D/3D registrations.
Registration of DRRs and PIs for verification of stereotactic body radiotherapy 2169

Acknowledgments

Jozef Grezdo was financially supported by the Marie Curie Fellowship HPMT-GH-01-00318-
05 and Thomas Künzler received a grant from Elekta.

References

Bert C, Metheany K G, Doppke K and Chen G T Y 2005 A phantom evaluation of a stereo-vision surface imaging
system for radiotherapy patient setup Med. Phys. 32 2753–62
Birkfellner W, Seemann R, Figl M, Hummel J, Ede C, Homolka P, Yang X, Niederer P and Bergmann H 2005
Wobbled splatting—a fast perspective volume rendering method for simulation of x-ray images from CT Phys.
Med. Biol. 50 N73–84
Clippe S, Sarrut D, Malet C, Miguet S, Ginestet C and Carrie C 2003 Patient setup error measurement using 3D
intensity-based image registration techniques Int. J. Radiat. Oncol. Biol. Phys. 56 259–65
Dekker N, Ploeger L S and van Herk M 2003 Evaluation of cost functions for gray value matching of two-dimensional
images in radiotherapy Med. Phys. 30 778–84
Ford E C, Chang J, Mueller K, Sidhu K, Todor D, Mageras G, Yorke E, Ling C C and Amols H 2002 Cone-beam CT
with megavoltage beams and an amorphous silicon electronic portal imaging device: potential for verification
of radiotherapy of lung cancer Med. Phys. 29 2913–24
Gierga D P, Brewer J, Sharp G C, Betke M, Willett C G and Chen G T Y 2005 The correlation between internal
and external markers for abdominal tumors: implications for respiratory gating Int. J. Radiat. Oncol. Biol.
Phys. 61 1551–58
Gilhuijs K G A, van de Ven P J H and van Herk M 1996 Automatic three-dimensional inspection of patient setup in
radiation therapy using portal images, simulator images, and computed tomography data Med. Phys. 23 389–99
Groh B A, Siewerdsen J H, Drake D G, Wong J W and Jaffray D A 2002 A performance comparison of flat-panel
imager-based MV and kV cone-beam CT Med. Phys. 29 967–75
Hanley J et al 1999 Deep inspiration breath-hold technique for lung tumors: the potential value of target immobilization
and reduced lung density in dose escalation Int. J. Radiat. Oncol. Biol. Phys. 45 603–11
Hill D L, Batchelor P G, Holden M and Hawkes D J 2001 Medical image registration Phys. Med. Biol. 46 R1–45
Jaffray D A, Siewerdsen J H, Wong J W and Martinez A A 2002 Flat-panel cone-beam computed tomography for
image-guided radiation therapy Int. J. Radiat. Oncol. Biol. Phys. 53 1337–49
Jans H-S, Syme A M, Rathee S and Fallone B G 2006 3D interfractional patient position verification using 2D–3D
registration of orthogonal images Med. Phys. 33 1420–39
Khamene A, Bloch P, Wein W, Svatos M and Sauer F 2006 Automatic registration of portal images and volumetric
CT for patient positioning in radiation therapy Med. Image Anal. 10 96–112
Kim J, Fessler J A, Lam K L, Balter J M and Haken R K T 2001 A feasibility study of mutual information based
setup error estimation for radiotherapy Med. Phys. 28 2507–17
Kim J, Yin F F, Zhao Y and Kim J H 2005 Effects of x-ray and CT image enhancements on the robustness and
accuracy of a rigid 3D/2D image registration Med. Phys. 32 866–73
Lacroute P and Levoy M 1994 Fast volume rendering using a shear-warp factorization of the viewing transformation
SIGGRAPH’94 451–8
LaRose D 2001 Iterative x-ray-CT registration using accelerated volume rendering PhD Thesis Carnegie Mellon
University http://www.ri.cmu.edu/pubs/pub 3762.html
Lemieux L, Jagoe R, Fish D R, Kitchen N D and Thomas D G T 1994 A patient-to-computed-tomography image
registration method based on digitally reconstructed radiographs Med. Phys. 21 1749–60
Levoy M 1990 Efficient ray tracing of volume data ACM Trans. Graph. 9 245–61
Maes F, Collignon A, Vandermaulen D, Marchal G and Suetens P 1997 Multimodality image registration by
maximization of mutual information IEEE Trans. Med. Imaging. 16 187–98
Matsopoulos G K, Asvestas P A, Delibasis K K, Kouloulias V, Uzunoglu N, Karaiskos P and Sandilos P 2004
Registration of electronic portal images for patient set-up verification Phys. Med. Biol. 49 3279–89
Meyer J, Richter A, Baier K, Wilbert J, Guckenberger M and Flentje M 2006 Tracking moving objects with
megavoltage portal imaging: a feasibility study Med. Phys. 33 1275–80
Munbodh R, Jaffray D A, Moseley D J, Chen Z, Knisely J P S, Cathier P and Duncan J S 2006 Automated 2D–3D
registration of a radiograph and a cone beam CT using line-segment enhancement Med. Phys. 33 1398–411
Park H, Bland P H, Brock K K and Meyer C R 2004 Adaptive registration using local information measures Med.
Image Anal. 8 465–73
2170 T Künzler et al

Penney G P, Batchelor P G, Hill D L G and Hawkes D J 2001 Validation of a two- to three-dimensional registration
algorithm for aligning preoperative CT images and intraoperative fluoroscopy images Med. Phys. 28 1024–32
Penney G P, Weese J, Little J A, Desmedt P, Hill D L G and Hawkes D J 1998 A comparison of similarity measures
for use in 2-D–3-D medical image registration IEEE Trans. Med. Imaging 17 586–95
Pouliot J et al 2005 Low-dose megavoltage cone-beam CT for radiation therapy Int. J. Radiat. Oncol. Biol.
Phys. 61 552–60
Scarbrough T J, Golden N M, Ting J Y, Fuller C D, Wong A, Kupelian P A and Thomas C R 2006 Comparison of
ultrasound and implanted seed marker prostate localization methods: implications for image-guided radiotherapy
Int. J. Radiat. Oncol. Biol. Phys. 65 378–87
Seppenwoolde Y, Shirato H, Kitamura K, Shimizu S, van Herk M, Lebesque J V and Miyasaka K 2002 Precise and
real-time measurement of 3D tumor motion in lung due to breathing and heartbeat, measured during radiotherapy
Int. J. Radiat. Oncol. Biol. Phys. 53 822–34
Sharp G C, Kollipara S, Madden T, Jiang S B and Rosenthal S J 2005 Anatomic feature-based registration for patient
set-up in head and neck cancer radiotherapy Phys. Med. Biol. 50 4667–79
Stock M, Kontrisova K, Dieckmann K, Bogner J, Poetter R and Georg D 2006 Development and application of a
real-time monitoring and feedback system for deep inspiration breath hold based on external marker tracking
Med. Phys. 33 2868–77
Verellen D, Soete G, Linthout N, Van Acker S, De Roover P, Vinh-Hung V, Van de Steene J and Storme G 2003
Quality assurance of a system for improved target localization and patient set-up that combines real-time infrared
tracking and stereoscopic X-ray imaging Radiother. Oncol. 67 129–41
Weiss E, Vorwerk H, Richter S and Hess C F 2003 Interfractional and intrafractional accuracy during radiotherapy
of gynecologic carcinomas: a comprehensive evaluation using the exactrac system Int. J. Radiat. Oncol. Biol.
Phys. 56 69–79
Wulf J, Hädinger U, Oppitz U, Olshausen B and Flentje M 2000 Stereotactic radiotherapy of extracranial targets:
CT-simulation and accuracy of treatment in the stereotactic body frame Radiother. Oncol. 57 225–36