The generation of digitally reconstructed radiographs

with six parameters1

Xin Li, Ling-hong Zhou2, Xin Zhen, Wen-ting Lu

School of Biomedical Engineering
Southern Medical University
Guangzhou, China
[email protected]

Abstract—The generation of digitally reconstructed radiographs devices. The second solution to reduce patient setup errors is
(DRRs) is a significant procedure in several medical imaging called Portal Imaging, which comes from Electronic Portal
applications such as 2D/3D image registration. Also, digitally Imaging Devices (EPID). The simulation films (SF) is used as
reconstructed radiographs play an important role in modern a reference for comparison with portal images. Charles Yang et
radiation therapy. They are planar X-rays that can be generated al.[2]summed up several disadvantages: (1) it requires a
from the same data sets used for the planning CT scan, taken separate setup of the patient, on a different couch, with
with the patient in the treatment position. To reduce the patient remarking of the reference centre; (2) it involves transcribing
setup errors, image registration, such as 2D/2D, 2D/3D, 3D/3D error in the transferal of the treatment volume from the CT scan
image registration, was introduced. Essentially, 2D/3D
to the SF; (3) simulation film, being a diagnostic X-ray, is
registration is a procedure that the DRRs resulted from a 3D CT
restricted in its ability to display mainly bony contours;
data set are to be matched with the real X-ray images by
applying a rigid transformation. This means that we should look (4) tumor volume and other soft-tissue volumes of interest
for a translation (expressed by three parameters tx, ty, tz) and a
cannot be easily and accurately marked on the SF; and (5) in
many instances it is necessary to employ additional measures
rotation (expressed with three parameters as well θx, θy, θz) in
such as intravenous (6) contrast (to display the kidneys), oral
space. At last, a fast DRR generation technique, called three-
dimension Bresenham line generation algorithm, with which
and rectal contrast, anal marker, skin marker etc. in order to
satisfactory images are obtained would be proposed. enhance the usefulness of SF.
Digitally reconstructed radiographs (DRRs) can overcome
Keywords-radiation therapy; digital reconstructed radiographs; all the disadvantages that are associated with conventional SF.
2D/3D image registration; three-dimension Bresenham algorithm Thus, DRRs appears. However, good-solution DRRs is
difficult to obtain because of hardware and software limitations.
I. INTRODUCTION
DRRs is a key technique in radiotherapy virtual simulation,
Radiation therapy has become one of the three main cancer which is used in radiation treatment planning system (TPS)
treatment modalities, together with surgery and chemotherapy. primarily. Also, it plays an important role in delineation of
To be efficient, radiation therapy must deliver a maximum dose targets [3]. Recently, DRRs has been becoming a significant
of X-rays (now produced by linear accelerators) to the tumor tool that corrects patient’s setup errors and position changes
while sparing surrounding normal tissue. Before the beginning during treatment in image guided radiation therapy (IGRT).
of treatment, physicians and physicists have to establish a
Radiotherapy Treatment Planning (RTP). The RTP defines the DRR volume rendering, also called simulated X-ray
number of beams, their size, their shape, their tilt and the beam volume rendering, is a direct volume rendering technique that
energy. This is now done in 3D with the help of a computed consists of simulating x-rays passing through the reconstructed
tomography (CT) scan of the patient. However, delivering X- CT volume based on the absorption only optical model. There
ray doses is a fractionated process. For instance, at least 35 are many methods implemented in generating DRR. For an N×
daily fractions are necessary to treat prostate cancer. But, N×N size volume, most direct volume rending techniques, such
installing the patient in exactly the same position every day is as ray casting [4], splatting [5], shear-warp [6], have Ο(N3)
very difficult. This position is defined by the CT-scan used for time complexity. Fourier volume rending [7] has Ο(N2logN)
establishing the RTP. Thus, the main difficulty is the day-to- time complexity, but it doesn’t support prospective projection.
day reproducibility of the patient setup. D. Sarrut et al. [1] Monte Carlo volume rending [8] is suitable for large data set,
reported that mean setup errors are between 5.5mm and 8mm as it is independent of the size of datasets and related to the
with a maximum, though rarely reported, of 18mm or 16mm. number of the samples, but the convergence of this means is
Even using of immobilization devices displacements remain slow. Some light-field methods need less time, nevertheless,
important: 22%are between 5 and 10mm and 57% are over they require large memory to store the pre-computed value.
4mm. The first solution to improve patient setup is patient However, For DRR generation, only direct methods are apt,
immobilization. However they do not eliminate all errors and because full voxel information is needed for the simulation of
fail to show any improvement with the use of immobilization

1 Project supported by: The Scientific Research Foundation of the Higher Education Institutions of Guangdong Province, China (Grant
No. CGZHZD0713); Key scientific and technological research programs of Guangdong Province, China (Grant No. 2008A030102011).
2 Corresponding author: Ling-Hong Zhou, Tel: 86-20-61648291, E-mail: [email protected].

978-1-4244-4713-8/10/$25.00 ©2010 IEEE

physical effect of X-ray attenuation, though it’s time- the entrance can be found. However, if the distance between
consuming. the exit and the source is no more than the one between the
entrance and the source, the value of the corresponding pixel in
II. MATERIALS AND METHODS DRR would be set into zero.
If the entrance and the exit (with different distance) are
A. Pre-processing CT Data known, three-dimension Bresenham algorithm would be
The voxel matrix composed by CT images is not applied to calculate the matching pixel value in DRR.
symmetrical originally, as the size of pixels in XY plane differs
from the one in Z direction .Therefore, it is necessary to D. Three-dimension Bresenham algorithm
construct symmetrical voxel matrix which has the same length, 2D Bresenham algorithm is a convenient and useful line
width and height for convenient computation. Interpolation can algorithm in Computer Graphics. The idea of this method can
be a friendly solution to this problem. be explained with Figure1. There are lots of pixels, such as
pixel C, D, E, which are ready for be chosen. Point B that lies
B. Establishing the Coordinate System between C and D is the real position the line pass through. The
Data coordinate system: the origin is located in the center task need to be done is just that selecting C or D to instead of B.
of the CT datasets, and X, Y, Z axis runs parallel with the Obviously, the most direct way is comparing the distance
length, width and height of the CT datasets separately. between B and D with the one between B and C. If BD is
greater than BC, C would be chosen. Otherwise, D is the
chosen one. As well, comparing the position of B to the one of
⎛ ⎞ A, which is in the middle between C and D, is another quick
⎛X ⎞ ⎛ X '⎞ ⎜ tx ⎟ means to judge.
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜Y ⎟ = R (θ x , θ y , θ z ) ⎜ Y ' ⎟ + ⎜ t y ⎟ (1)
⎜Z ⎟ ⎜ Z '⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎜t ⎟ D
⎝ z⎠
Display coordinate system: It’s used to show the result. In
our implement, this can be counted as Data coordinate before
rigid transformation. A point (X′,Y′,Z′) in display coordinate
B
system can be changed into the corresponding point (X,Y,Z) in
data coordinate system through a transformation matrix: The
A
expression of the rotate matrix R(Θx,Θy,Θz) in (1) is (2) and
the (tx, ty, tz)T represents the translation vector.
E C
R (θ x , θ y ,θ z ) =
⎛ Cy × Cz Cx × Sz + Sx × Sy × Cz Sx × Sz − Cx × Sy × Cz ⎞
⎜ ⎟ (2)
⎜ −Cy × Sz Cx × Cz − Sx × Sy × Sz Sx × Cz + Cx × Sy × Sz ⎟ Figure 1. 2D Bresenham algorithm
⎜ Sy − Sx × Cy Cx × Cy ⎟
⎝ ⎠
After that, we could generalize 2D Bresenham algorithm to
where
three-dimension straight line.
Sx = sin θ x , Cx = cos θ x
Sy = sin θ y , Cy = cos θ y Supposing that the entrance coordinate is (x0, y0, z0) and the
exit coordinate is (x1, y1, z1), the differencing in three
Sz = sin θ z , Cz = cos θ z
directions could be expressed like (3), (4), (5):
C. Computation of Digtal Reconstructed Radiographys
With the help of the transformation matrix, the position of Δx = x1-x0 (3)
virtual illuminating source and imaging plane in display
coordinate system could be mapped into the data space.
Δy = y1-y0 (4)
Assuming the resolution of DRR is M×N, what to do next
is computing the intersections of every ray and the data. The
intersections computation can be designed like this: First, six Δz = z1–z0 (5)
points can be obtained, which exists both in the line and on one
face of the data. Then these intersections would be divided into
Without loss of generality, let Δx>=Δy>=Δz>0. Then
two groups, the positive group and the negative one, which is
equation (6) and (7) defines two variables.
determined by the sign of inner product between the line vector
and the normal vector of the face. If the point in the positive
group is in the data, it would be chosen as the exit. Similarly, judge1 = Δy/Δx-0.5 (6)
 judge2 =Δz/Δx-0.5 (7)

At last, the flow chart of three-dimension Bresenham

algorithm will be given in Figure2.

Given x=x0,
Δx=x1-x0, Δy=y1-y0, Δz=z1-z0

Obtain judge1 and judge2:

judge1 = Δy/Δx-0.5 judge2 =Δz/Δx-0.5

Yes No
judge1>=0? No Figure 3. DRR with the parameter (0, 0, 0, 0, 0, 0)

y=y; y=y+1;

judge1=judge1+Δy/Δx judge1=judge1+Δy/Δx-1

Yes No
judge2>=0?

z=z; z=z+1;

judge2=judge2+Δz/Δx judge2=judge2+Δz/Δx-1

x=x+1

Yes Figure 4. DRR with the parameter (0, 0, 0, 0, 0, 90)

x<=x1?

No

End

Figure 2. The flow chart of the three-dimension Bresenham algorithm

III. RESULTS AND CONCLUSION

A. Results
Our algorithm was implemented in C++ and tested on a
2.66GHz Intel Celeron PC with 1G RAM, and its
implementation was based purely on the software without
exploiting any supporting 3D hardware. The dataset is a
512×512×170 CT scan of the abdomen.
Figure3—Figure6 show the DRRs of the pixel size
512×512 generated from the CT dataset of different positions.

Figure 5. DRR with the parameter (0, 0, 0, 90, 45, 0)

 The preprocessing time was about 10s and the production I shall extend my thanks to Mr. Zhen xin and Miss Lu wen-
of DRRs consumes 7s more or less. ting for all their kindness and help. I would also like to thank
all my postgraduate classmates who have helped me to develop
B. Conclusion the fundamental and essential academic competence in our
DRRs with good quality in different position can be institute.
generated. And three-dimension Bresenham algorithm has a Last but not least, I’d like to thank all my friends, especially
nice performance on speeding up the generation process. my three lovely roommates, for their encouragement and
support.

REFERENCES

[1] D. Sarrut, S. Clippe “Fast DRR generation for intensity-based 2D/3D

image registration in radiotherapy” Elsevier Science 2003
[2] Charles Yang, Michael Guiney, Peter Hughes, Samuel Leung, Kuen
HoeLiew, James Matar and George Quong “Use of digitally
reconstructed radiographs in radiotherapy treatment planning and
verification” Australasian Radiology (2000) 44, pp439-443
[3] Zhen Xin, Zhou Ling-hong, Wang Zhuo-yu, Chen Chao-min,
“Reconstruction of contour lines in TPS using Delaunay Triangulation.
Bioinformatics and Biomedical Engineering,” The 1st International
Conference on Bioinformatics and Biomedical Engineering, Wuhan,
China, 2007, pp.1022-24.
[4] M. Levoy, “Display of surfaces from volume data,” Computer Graphics
and Applications, 8(3), 1988,pp.29-37.
[5] Westover L, “Footprint evaluation for volume rendering,” Proceedings
of SIGGRAPH, 24(4) ,1990,pp.367-376.
[6] Philippe Lacroute, Marc Levoy, “Fast Volume Rendering Using a Shear-
Warp Factorization of the Viewing Transformation,”
Figure 6. DRR with the parameter (128, 128, 128,- 90, 0, 0) Proc.SIGGRAPH’94, Orlando, Floirda,1994, pp.451-458.
[7] T. Malzbender, “Fourier volume rendering ACM Trans. Graph,”
1993,pp.233-50.
ACKOWMLEDGMENT [8] B. Csebfalvi, L. Szirmay-Kalos, “Monte Carlo volume rendering,” Proc.
IEEE Visualization Conf. 2003 (Seattle)(New York: Institute of
First and foremost, I would like to show my deepest Electrical and Electronics Engineers),2003,pp.449-56
gratitude to my supervisor, Dr. Zhou ling-hong, a respectable,
responsible and resourceful scholar, who has provided me with
valuable guidance.