A STEREOPHOTOGRAMMIC SYSTEM TO POSITION PATIENTS
FOR PROTON THERAPY
Neil Muller
Department of Mathematical Sciences, University of Stellenbosch, Stellenbosch, South Africa
Evan de Kock, Ruby van Rooyen and Chris Trauernicht
Medical Radiation Group, iThemba Labs, Somerset West, South Africa
Keywords:
Stereo vision, proton therapy, visually guided robotics.
Abstract:
Proton therapy is a successful treatment for lesions that are hard to treat using conventional radiotherapy, as
the radiation dose to nearby critical structures can be tightly controlled. To realise these advantages, the patient
needs to be accurately positioned, and monitored during treatment to ensure that no motion occurs. iThemba
LABS uses a fixed beam-line, and moves the patient using a suitable positioning device. In this paper, we
discuss several aspects the stereo vision based system used to both determine the position of the patient in the
room, and to monitor the patient during treatment.
1 INTRODUCTION
Proton therapy is a useful treatment method as the
dose distribution properties of the proton beam al-
low higher doses to be delivered to the target volume
with lower doses to the surrounding tissue. Due to the
high cost associated of treatment, it is often reserved
for lesions that are difficult to treat with conventional
radiotherapy techniques, such as inter-cranial lesions
or lesions close to critical structures (see for example
(Webb, 1993)).
iThemba LABS
1
has been involved with proton
therapy for over ten years. Due to cost restrictions,
iThemba LABS uses a fixed beam-line to deliver the
proton dose, and a motorised chair with 5 degrees
of freedom to position the patient (see (Jones et al.,
1995)). A second treatment vault is being built that
uses a robot-controlled manipulator to position the pa-
tient. The position of the patient during setup and
treatment is monitored by a number of cameras and
stereo techniques are used to calculate the patient’s
position. A critical issue is the high accuracy re-
quired.
1
Laboratory for Accelerator Based Science
2 PROTON VAULTS AT ITHEMBA
LABS
The treatment environment (see Figure 1), shows the
treatment layout. We have 9 cameras observing the
patient during treatment. During a treatment session,
at least 3 cameras will be used to calculate the position
of the markers on the mask (see Figure 2(a)). The
cameras are positioned around the beam isocenter at
an average distance of 2m from the isocenter. The
positions were chosen to maximise the volume visible
to the cameras while not obstructing activities in the
vault.
The position of the target volume is determined
from the CT scan used to plan the treatment. So that
the position of the target volume can be related to the
mask, the patient is scanned wearing the mask. When
the patient is in the reference position for the scan, a
stereo system, calibrated to CT scanners coordinate
system (see (de Kock et al., 2002)), determines the
position of the markers on the mask. Thus the rela-
tionship between the target volume and the markers
on the mask is known.
The stereophotogrammetric (SPG) system is de-
signed to be used in both vaults, and the control
systems for the chair and the robot are designed to
present a common interface to the SPG system. The
538
Muller N., de Kock E., van Rooyen R. and Trauernicht C. (2007).
A STEREOPHOTOGRAMMIC SYSTEM TO POSITION PATIENTS FOR PROTON THERAPY.
In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IU/MTSV, pages 538-541
Copyright
c
SciTePress
Figure 1: The treatment vaults..
(a) Treatment
Mask.
(b) The SPG System’s
View of the Treatment
Mask.
Figure 2: The treatment mask.
SPG system is responsible both for positioning the
patient correctly, and monitoring the patient during
treatment to ensure that no excessive movement oc-
curs.
3 THE
STEREOPHOTOGRAMMETRIC
SYSTEM
A central design aim of the current project is to
achieve the required level of accuracy using commod-
ity hardware. Thus we choose to use conventional
PAL CCD cameras, and use frame-grabbers based off
the venerable BT-878 chipset.
Since we need to monitor the patient position in
real time and react to patient motion, we restrict our-
selves to a subset of the cameras during a treatment
session to minimise the computational load. We also
try to ensure that the marker detection problem is as
simple as possible, by using retro-reflective markers,
illuminated by green light-emitting diodes (LEDs)
mounted next to the cameras, and narrow band-pass
filters on the cameras, to minimise the effect of the
background on the image, and to ensure we have good
contrast between the markers and the background.
The markers are 10mm in diameter. A typical image
of the mask is shown in Figure 2(b).
The markers used are circular, and, because of the
lighting conditions, the contrast between the markers
and the background is high. Thus the markers can be
found by searching for bright regions, and the marker
centroid can be found using ellipse fitting techniques.
Each candidate marker is only accepted if the size is
within acceptable limits, and the deviation of the ob-
served shape of the marker from an ideal ellipse is
sufficiently small (see (van Rooyen, 2003)).
4 CALIBRATION
Calibration of the SPG system needs to achieve two
goals. We need to calibrate the parameters of the cam-
eras, and their relative positions, so we can obtain ac-
curate results from the stereo reconstruction. Since
the SPG system must position the patient with respect
to the beam, which is fixed in the room, calibration
needs to establish the relationship between the room
coordinate system and the coordinate system of the
cameras.
4.1 Distortion
As we are using conventional zoom lenses, radial dis-
tortion is a factor. We use a two-stage approach to
estimate the distortion. We determine an accurate
base distortion model from observing a suitable pat-
tern (shown in Figure 3(a)) from a known position.
(a) Distortion Correction
Pattern.
(b) The Calibration
Cube.
Figure 3: Calibration patterns.
This is designed to be easy for our software to pro-
cess. By observing the extent to which the lines of
dots curve, the parameters of the distortion model can
be obtained. We use a standard model for radial dis-
tortion, namely
(x
u
, y
u
) = (x
d
, y
d
) + (
x
d
, y
d
)δ(κ) (1)
where
r
2
d
= x
2
d
+ y
2
d
= (x
d
c
xr
)
2
+ (y
d
cyr)
2
(2)
and
δ(κ) = κ
1
r
2
d
+ κ
2
r
4
d
+ ... (3)
Since the relationship between the camera and the
pattern is known, and the pattern is known to a high
degree of accuracy, the only unknown parameters af-
fecting the image are the distortion coefficients and
the focal length. These can be estimated accurately
using standard non-linear optimisation techniques.
We restrict the model to a fairly low-dimensional
polynomial for numerical stability. While we do not
directly model tangential distortion, we allow the cen-
tre of distortion to differ from the centre of the image,
as this is a good approximation for the decentring dis-
tortion (see (Stein, 1997)).
This base distortion model is expected to change
over time, as many events will require the cameras to
be re-adjusted in the vault. Since such changes to the
model should be small, and as removing the camera to
recalculate the model in the fixed rig is inconvenient,
we wish to update the model. Thus we mount the
pattern in Figure 3(a) on a portable planar object. This
is held so that it roughly fills the field on vision of the
camera.
By extracting the lines from the pattern, we can
test if the model is still valid. If the model is no
longer valid, we can use our knowledge of the pat-
tern to calculate an updated distortion model, (similar
to (Tamaki et al., 2002)). Since the pattern is held so
that it is approximately face-on to the camera, we can
use that and current distortion model as a reasonable
initial estimate, and the optimiser converges to the up-
dated solution quickly.
From this updated distortion model, we can then
decide whether the drift from the base model has be-
come too large, in which case we remove the camera
and recalculate the base distortion model, or else we
simply use the updated model (see (van Rooyen and
Muller, 2004) for more details).
4.2 Camera Calibration
Since we need to calibrate the cameras to the surveyed
coordinate system in the room, we use a single cube
to calibrate all the cameras. This cube is mounted on
a specially constructed jig which can be reliably posi-
tioned at the reference position. The correspondence
between the cube position and the room was obtained
by surveying the room with the cube in position, and
is known to a high degree of accuracy. The relation-
ship between the cube and the beam-line is regularly
checked using a theodolite mounted along the beam-
line. We use different cubes for each vault, and the
jig design is slightly different between each vault so
there is no possibility of using the incorrect cube.
Each face of has a large number of circular mark-
ers, whose positions are known from the surveying re-
sults. Each face also as a distinctive pattern of squares
and triangles which is used to automatically label all
the points detected on the cube (see figure 3(b)). From
the number of markers visible, we can determine if it
has an adequate view of the volume around the beam
isocenter.
The current distortion model is used to correct for
distortion before calibration, and then calibration pro-
ceeds using Tsai’s method (see (Tsai, 1987) for de-
tails).
5 POSITIONING THE PATIENT
For each target volume, several beams, each with dif-
ferent entry points are planned. To treat a beam, the
patient is positioned so that the target volume is posi-
tioned at the beam isocenter, and the entry point lies
along the beam axis, between the beam source and
the target volume. A collimator is inserted into the
beam line to shape the beam to the target volume,
and, once the patient is in position, the collimator is
rotated to match the profile of the target volume. In
a single treatment session, the patient will be treated
using multiple beams.
Before treatment, the patient is secured to the po-
sitioning system and moved to a standard reference
position. In this position, most of the markers on the
mask can be observed, and stereo vision techniques
are used to determine the 3D position of a subset of
the markers on the mask.
These markers are then registered to the marker
positions in the reference position of the CT Scan.
This determines the transformation between the CT
scanner reference position and the current patient po-
sition. From this, the current position of the target
volume and entry point can be determined. The re-
quired movement to position the patient is calculated
and sent to the positioning system.
This procedure is repeated until the patient is in
position. This final position is verified by comparing
an X-ray taken along the beam path with the predicted
X-ray view (see (van der Bijl, 2006)).
Positioning the patient for the next beam in a ses-
sion follows the same procedure, except that, since
the current position of the patient is known, there is
no need to return to the reference position.
6 MONITORING THE PATIENT
Monitoring the patient is a subset of the positioning
loop. Since the patient is monitored continually after
being positioned, and the frame-rate is high in com-
parison to the speed of the movements of interest, we
can a simple nearest neighbour comparison to track
markers.
On each iteration, the current position of the target
volume and the entry point are calculated and com-
pared to the required positions. If the differences ex-
ceed the specified tolerances, the patient is declared
to be out of position and the beam is interrupted.
Monitoring continues for some time after this to
determine if this is a transient event. If the patient
moves back into position after a short delay and re-
mains in position for a suitable period, the treatment
is resumed, otherwise the beam is aborted, and the
patient is repositioned.
7 ACCURACY OF THE SYSTEM
To assess the accuracy of the SPG system in reason-
ably realistic circumstances, we ran a series of ex-
periments using a dummy mask and using the mo-
torised chair. We constructed 8 treatment beams for
this mask, with the target point of each beam coinci-
dent with one of the markers on the mask. The accu-
racy was measured using a theodolite mounted along
the beam-line.
The mask was positioned for each beam using the
SPG system, and then the chair was moved until the
marker chosen as the target point was centred in the
view of the theodolite. The displacement required to
centre the marker was recorded for each beam, and
this was repeated for all the beams, using several dif-
ferent camera combinations for each beam. These ex-
periments were repeated daily over 5 successive days.
Because the of position of the theodolite, the dis-
placement along the beam-line, could not be ob-
served, so we only measured the error orthogonal to
the beam-line.
The average error was found to be 0.305mm with a
standard deviation of 0.219mm. Since the slice thick-
ness typically used for planning patients at iThemba
LABS is 1mm, this is well within the required toler-
ances.
The results for each beam are listed in Table 1.
Table 1: Accuracy Test Results.
Beam Avg d Std Dev d
1 0.436 0.116
2 0.187 0.088
3 0.175 0.105
4 0.416 0.185
5 0.158 0.074
6 0.182 0.086
7 0.247 0.119
8 0.265 0.160
All 0.305 0.219
8 CONCLUSIONS
In this paper, we describe several aspects of the
stereophotogrammetric system used for patient po-
sitioning at iThemba LABS. By exploiting the con-
trolled natured of the environment, we can obtain
good results using simple approaches. In addition to
positioning the patient, we can monitor the patient and
respond if the patient moves during treatment. The
system achieves acceptable accuracy using commod-
ity hardware.
REFERENCES
de Kock, E., O’Kennedy, B., and Muller, N. (2002). Cali-
brating a stereo rig and CT scanner with a single cal-
ibration object. In Greiner, G., Niemann, H., Ertl, T.,
Girod, B., and Seidel, H.-P., editors, Vision, Modeling,
and Visualization 2002.
Jones, D. T. L., Schreuder, A. N., Symons, J. E., R
¨
uther,
H., van der Vlugt, G., Bennet, K. F., and Yates, A.
D. B. (1995). Use of stereo-photgrammetry in proton
radiotherapy. In R
¨
uther, H., editor, Proc. Int. FIG.
Symp. on Photogrammetry in Enigineering Surveying,
pages 138–152.
Stein, G. (1997). Lens disortion calibration point correspon-
dences. In IEEE Conference on Computer Vision and
Pattern Recognition, pages 602–608.
Tamaki, T., Yamamura, T., and Ohnishi, N. (2002). Uni-
fied approach to image distortion. In Kasturi, R., Lau-
rendeau, D., and Suen, C., editors, 16th International
Conference on Pattern Recognition, pages 584–587.
IEEE, IEEE Computer Society.
Tsai, R. Y. (1987). A versatile camera calibration technique
for high-accuracy 3D machine vision metrology using
off-the-shelf TV cameras and lenses. IEEE Journal of
Robotics and Automation, RA-3(4):323–344.
van der Bijl, L. (2006). Verification of patient position for
proton therapy uisng portal x-rays and digitally recon-
structed radiographs. Master’s thesis, University of
Stellenbosch.
van Rooyen, R. (2003). Fast, robust detection of circular
retro-reflective targets. In Proceeding of the Four-
teenth Annual Symposium of the Pattern Recognition
Association of South Africa, pages 21–26.
van Rooyen, R. and Muller, N. (2004). A method for effi-
ciently re-estimating camera distortion parameters. In
et al., F. N., editor, Proceeding of the Fifteenth An-
nual Symposium of the Pattern Recognition Associa-
tion of South Africa. Pattern Recognition Association
of South Africa.
Webb, S. (1993). The physics of three-dimensional radio-
therapy: Confromal radiotherapy, radiosurgery and
treatment planning. Insitute of Physics Publishing,
Bristol and Philadelphia.