Professional Documents
Culture Documents
G. Shechter
Philips Medical Systems and Technologies, Haifa, Israel
Th. Köhler
Philips Research Laboratories, Hamburg, Germany
A. Altman
Philips Medical Systems and Technologies, Haifa, Israel
R. Proksa
Philips Research Laboratories, Hamburg, Germany
共Received 6 January 2004; revised 14 April 2004; accepted for publication 27 May 2004;
published 14 July 2004兲
A new approximate method for the utilization of redundant data in helical cone-beam CT is pre-
sented. It is based on the observation that the original WEDGE method provides excellent image
quality if only little more than 180° data are used for back-projection, and that significant low-
frequency artifacts appear if a larger amount of redundant data are used. This degradation is
compensated by the frequency split method: The low-frequency part of the image is reconstructed
using little more than 180° of data, while the high frequency part is reconstructed using all data.
The resulting algorithm shows no cone-beam artifacts in a simulation of a 64-row scanner. It is
further shown that the frequency split method hardly degrades the signal-to-noise ratio of the
reconstructed images and that it behaves robustly in the presence of motion. © 2004 American
Association of Physicists in Medicine. 关DOI: 10.1118/1.1773622兴
Key words: cone-beam CT, cone-beam reconstruction, helical CT, PI, wedge
2230 Med. Phys. 31 „8…, August 2004 0094-2405Õ2004Õ31„8…Õ2230Õ7Õ$22.00 © 2004 Am. Assoc. Phys. Med. 2230
2231 Shechter et al.: The frequency split method for helical cone-beam reconstruction 2231
FIG. 1. Geometry of the data acquisition system 共left兲 and the beam shape
after WEDGE-beam transformation.
p 1 共 ,  ,h 兲 ⫽ 冕0
⬁
f 共 S 共 兲 ⫹lE 共 ,  ,h 兲兲 dl, 共3兲
FIG. 2. Exemplary overscan weighting functions for different illumination
windows according to Eq. 共13兲. Illumination windows 关 f : l 兴 from top to
where E(,  ,h) is the unit vector pointing from the source bottom are 关⫺/2:/2兴, 关⫺2/3:2/3兴, 关⫺:兴, and 关⫺1.3:1.3兴.
position S() to the detector element D(,  ,h).
冕
ning, which is also called WEDGE-beam-transform.10 Each 1
row of the detector is handled separately. The new data are f 共 x 兲⫽ w 共 ,x 兲 p 4 共 ,u 共 ,x 兲 ,h 共 ,x 兲兲 d , 共9兲
2 I共 x 兲
parametrized by the view angle , which is the angle of the
parallel lines projected onto the xy plane, the distance of the where the functions u( ,x) and h( ,x) provide the
lines from the rotation axis u and finally the unchanged co- coordinates of the object point x projected onto the WEDGE
ordinate h, see also Fig. 1: detector, and the integration is performed over the set
I(x) of all projection angles, where the object point
⫽⫹  , 共4兲
x is within the cone. The weighting function w( ,x) has
u⫽R sin  , 共5兲 to fulfill the following normalization condition for all
0 苸 关 0: 兴 :
p 2 共 ,u,h 兲 ⫽p 1 共 ⫺arcsin共 u/R 兲 ,arcsin共 u/R 兲 ,h 兲 . 共6兲
Geometrically, the cone beam is transformed into a
wedge-shaped beam, which gave rise to the name of the
冕 兺
I 共 x 兲 兵 i苸N 其
␦ 共 0 ⫺i 兲 w 共 ,x 兲 d ⫽1. 共10兲
␣共 x 兲⫽ 再 , l 共 x 兲 ⫺ f 共 x 兲 ⬎2 ,
共 l 共 x 兲 ⫺ f 共 x 兲兲 ⫺ , else,
共11兲
再
T 共 x 兲 ⫽ l共 x 兲 ⫺ f共 x 兲 ⫺
1, else
, l 共 x 兲 ⫺ f 共 x 兲 ⬎2
共12兲
冦
0, F 2 ⫽0.95, F 3 ⫽0.05, F 4 ⫽0.12.
T 共 x 兲共 ⫺ f 共 x 兲兲 / ␣ 共 x 兲 , f 共 x 兲⭐⬍ f 共 x 兲⫹␣共 x 兲,
⫽ T共 x 兲, f 共 x 兲 ⫹ ␣ 共 x 兲 ⭐ ⬍ l共 x 兲 ⫺ ␣ 共 x 兲 , WEDGE reconstruction, the product of the ramp with M ( )
would be used in the filter step. We use two parameters F 1
T 共 x 兲共 l 共 x 兲 ⫺ 兲 / ␣ 共 x 兲 , l共 x 兲 ⫺ ␣ 共 x 兲 ⭐ ⬍ l共 x 兲 ,
and F 2 to parametrize the shape of M. Here, F 1 is the start
0, l共 x 兲 ⭐ . and F 2 the end of a cosine shaped fall-off according to
共13兲 for ⭐F 1 Ny ,
冦
1,
The shape of this weighting function is illustrated in Fig. 2
for different illumination window widths.
M F 1 ,F 2 共 兲 ⫽
1
2 冉
1⫹cos 冉
共 ⫺F 1 Ny兲
共 F 2 ⫺F 1 兲 Ny
冊冊 ,
共14兲
B. The frequency split method for F 1 Ny⬍ ⭐F 2 Ny ,
0, else,
If the original WEDGE method is used without redundant
data, we observe only very little cone-beam artifacts, even with Ny being the Nyquist frequency. In the frequency split
for large cone angles.15 However, if it is applied to scans method, we use two functions M Flf ,F ( ) and
3 4
with a small pitch and a large cone angle, we observe sig- M Fhf ,F ,F ,F ( ) to reconstruct the low-frequency and the
1 2 3 4
nificant artifacts for both, aperture weighted WEDGE and high frequency image, respectively. The two parameters F 3
angular weighted WEDGE. This is shown in Sec. III for and F 4 are used as F 1 and F 2 in Eq. 共14兲 to control the
angular weighted WEDGE. The artifacts have typically a resolution of the low-frequency image. The modulation
very low spatial frequency in planes perpendicular to the transfer function for the high-frequency image is simply
rotation axis. This has been observed also for other 3D re-
construction algorithms.12,13 This motivated the following M Fhf
1 ,F 2 ,F 3 ,F 4
共 兲 ⫽M F 1 ,F 2 共 兲 ⫺M Flf 3 ,F 4 共 兲 . 共15兲
approach to handle redundant data: Two separate images are
The modulation transfer functions are illustrated in Fig. 3.
reconstructed. The first one contains the low spatial fre-
As Kudo et al.,13 we reconstruct the low frequency image
quency components in planes perpendicular to the rotation
not from strictly nonredundant data, but we use a small
axis. This is reconstructed using no or only little redundant
amount of redundant data 共which is called ⌬ by Kudo
data. The second one contains the high spatial frequencies
et al.兲. The amount is another parameter for the reconstruc-
and this is reconstructed using all redundant data. It may be
tion. The low frequency image is reconstructed using an an-
noted that one advantage of using redundant data is the re-
gular range of ⫹ centered around the PI illumination win-
duction of aliasing artifacts in the xy plane. Since these ar-
dow I PI(x). The reason to use some redundant data is to
tifacts show up in high spatial frequencies, the frequency
reduce the sensitivity of the algorithm to patient motion. This
split method still maintains this feature. 共Sampling artifacts
will be demonstrated in Sec. III. The high frequency image is
in z direction,6 also known as windmill artifacts,19 are of
always reconstructed using all redundant data.
course not reduced by the quarter detector offset and are also
not expected to be suppressed by the frequency split method
since they show up in the medium frequency range.兲 III. EVALUATION
A modification of the ramp filtering 关Eq. 共8兲兴 is tradition- Head imaging is probably the most demanding applica-
ally used to control the resolution in the xy plane by multi- tion for cone-beam algorithms because of the very complex
plication with a modulation transfer function 共MTF兲. The bone structures inside the head, which can easily cause cone-
frequency split is realized by a modification of this MTF beam artifacts. Thus, we picked a head phantom for the
function. Suppose we want to achieve an MTF given in the evaluation of cone-beam artifacts. For the analysis of the
frequency domain as a function M ( ). For standard sensitivity to motion, we took the cardiac phantom described
TABLE I. Definition of the scanners used in the simulations. L is the distance TABLE II. RMS noise level in HU calculated in a circular region of interest
from the source to the detector, R the distance of the source to the rotation containing 5025 voxels on the left-hand side of the phantom. Images were
axis, and H the detector height, and P relative pitch. reconstructed using standard WEDGE. They are shown in Fig. 4.
in the lower row. Here, the effect is even stronger because WEDGE. However, for the higher cutoff frequency F 4
approximately three times more data were used by the recon- ⫽0.3, we observe a small increase of the noise level by 3%.
struction for the right image than for the left one. This causes
also the significant reduction in the noise level, see Table II. B. Sensitivity to motion artifact
If all redundant data are used, the resulting noise level is In order to analyze the sensitivity to patient motion, we
almost the same for both pitches, as expected since the simu- used the cardiac phantom as described in Ref. 20. The phan-
lated dose for the two cases is the same. The argument about tom had 60 beats per minute and a rest period of 20%. The
the same dose is strictly true only for object points on the simulated time for one rotation of the scanner was 420 ms.
rotation axis, where the ratio illumination window is the The other scan parameters are the same as for the low pitch
same as the ratio of the relative pitches. Since the region, scan of the head phantom. Results for the frequency split
where we computed the noise level is slightly off-center, a method for different values of the overscan and the cutoff
small difference in the noise level is expected. frequency F 4 are shown in Fig. 7. For comparison, Fig. 8
shows the result for the standard WEDGE method. The goal
2. Frequency split WEDGE of the evaluation here is not to recover the small spheres
Figure 5 shows exemplarily the low frequency, the high well—we do not evaluate a cardiac CT algorithm—but to get
frequency, and the final image for the frequency split method as little artifacts as possible in the surrounding area. It might
applied to the low-pitch scan. Results for the frequency split be seen as a simulation of a thorax scan, where we do not
method applied to the low-pitch scan are shown in Fig. 6 for want to get artifacts inside the nonmoving tissue from a
different values of the overscan angle and cutoff frequency moving stent placed in a coronary artery.
F 4 . The rms noise values are summarized in Table III. Comparing the upper and the lower row of Fig. 7, we
We observe, that the low-frequency artifacts are com- observe that the sensitivity to patient motion increases with
pletely suppressed by the frequency split method for both an increased cutoff frequency F 4 . This can be explained by
cutoff frequencies and ⭐/3. For a larger overscan of the fact that a low cutoff frequency implies that more fre-
⫽2/3, the dark and bright shades near the inner ear, which quency components are reconstructed using the full overscan
are present in the standard WEDGE reconstruction in the range. The sensitivity is reduced if a larger overscan angle
lower right image of Fig. 4 become slightly visible. is used. However, the biggest decrease in motion artifacts is
The noise measurement confirms that for the low achieved by going from ⫽0 共left column of Fig. 7兲 to
cutoff frequency F 4 ⫽0.15, the dose utilization of the ⫽/3 共middle column兲. Increasing the overscan angle fur-
frequency split method is virtually the same as for standard ther to ⫽2/3 共right column兲, has only a little effect on the
TABLE III. RMS noise level in HU calculated in a circular region of interest TABLE IV. Artifact levels in the motion phantom. The RMS deviation from
containing 5025 voxels on the left-hand side of the phantom. Images were the true background value was calculated in the square indicated in Fig. 8.
reconstructed using the frequency split method. They are shown in Fig. 6. The first three columns are related to the images shown in Fig. 7, full
For all cases, we used F 3 ⫽0. overscan is equivalent to standard WEDGE shown in Fig. 8. For all cases,
we used F 3 ⫽0.
⫽0 ⫽/3 ⫽2/3
⫽0 ⫽/3 ⫽2/3 Full overscan
F 4 ⫽0.15 5.76 5.76 5.77
F 4 ⫽0.30 5.92 5.90 5.90 F 4 ⫽0.15 2.6 1.3 1.1 1.0
F 4 ⫽0.30 4.8 1.4 1.2 1.0
motion artifacts. This conclusion obtained from a visual in- Alternatively, it is valid to say that these Radon planes con-
spection is also supported by a quantitative analysis of the tain only low-frequency information in planes perpendicular
motion artifacts. We calculated the root-mean-square devia- to the rotation axis.
tion from the true background value in a rectangular region The observation that the WEDGE method exhibits low-
indicated in Fig. 8. The results are summarized in Table IV. frequency artifacts may be explained by the facts mentioned
The frequency split method applied to WEDGE becomes above. In the WEDGE method, redundancies are treated like
the standard WEDGE method for the both limits, F 4 →0 and in 2D CT. The fact that some Radon planes are not measured
→max. Thus, the result of the standard WEDGE method as multiple times is neglected. The structure of the resulting
shown in Fig. 8 represent the best possible result that we can artifacts can be related by the Fourier slice theorem to an
obtain with the frequency split method applied to WEDGE. improper weighting of these Radon planes.
It appears that an overscan angle of ⫽/3 is sufficient to The frequency split method is a heuristic approach to
reach almost the same level of artifact suppression as the mitigate this effect. Cone-beam artifacts are reduced signifi-
standard WEDGE method. cantly, see Figs. 5 and 6. This improvement is achieved over
a wide range for the parameter values F 4 and .
The choice for the overscan parameter and the cutoff
IV. DISCUSSION frequency is obviously a trade-off between the performance
Like in many other approximate methods, the way redun- of the algorithm with respect to cone-beam artifacts, dose
dant data are used in the WEDGE method is motivated and utility, and robustness with respect to patient motion. If we
justified by two-dimensional 共2D兲 arguments: Rays passing a consider cone-beam artifacts, it appears that the overscan
certain object point are considered to be equivalent if they must not be chosen large, see right-most images in Fig. 6.
are parallel in the projection onto the plane perpendicular to The aspect of dose utilization pushes the cutoff frequency
the rotation axis. Of course, redundancy in cone-beam CT is down and the overscan to large values. Similarly, motion
not related to line integrals but to Radon planes. As a matter artifacts are better suppressed for a low cutoff frequency F 4
of fact, not all Radon planes are measured multiple times for and a large overscan angle, see Fig. 7. However, it appears
a low-pitch helical CT scan.17,5 In particular, planes, which that these parameters are not very critical for the scanner that
have only one intersection point with the helix, can be mea- we investigated here. With a proper choice of the cutoff fre-
sured only once. Geometrically, these planes are almost per- quency and the overscan angle, namely F 4 ⫽0.15 and
pendicular to the rotation axis. According to the 3D Fourier ⫽/3, we can maintain the good dose utilization and sup-
slice theorem, these Radon planes contain predominantly in- pression of motion artifacts of the WEDGE method, while
formation about spatial frequencies along the rotation axis. reducing cone-beam artifacts significantly.
1
K. C. Tam, S. Samarasekera, and F. Sauer, ‘‘Exact cone beam CT with a
spiral scan,’’ Phys. Med. Biol. 43, 1015–1024 共1998兲.
2
M. Defrise, F. Noo, and H. Kudo, ‘‘A solution to the long object problem
in helical cone-beam tomography,’’ Phys. Med. Biol. 45, 623– 643
共2000兲.
3
A. Katsevich, ‘‘Analysis of an exact inversion algorithm for spiral cone-
beam CT,’’ Phys. Med. Biol. 47, 2583–2597 共2002兲.
4
A. Katsevich, ‘‘Theoretically exact FBP-type inversion algorithm for spi-
ral CT,’’ SIAM 共Soc. Ind. Appl. Math.兲 J. Appl. Math. 62, 2012–2026
共2002兲.
5
C. Bontus, T. Köhler, and R. Proksa, ‘‘A quasiexact reconstruction algo-
rithm for helical CT using a 3-Pi acquisition,’’ Med. Phys. 30, 2493–2502
共2003兲.
6
T. Köhler, R. Proksa, C. Bontus, M. Grass, and J. Timmer, ‘‘Artifact
analysis of approximate cone-beam CT algorithms,’’ Med. Phys. 29,
FIG. 8. Result for the standard WEDGE method and the low-pitch scan of 51– 64 共2002兲.
7
the cardiac phantom at z⫽3 mm. The rectangle indicates the region where C. Bontus, R. Proksa, J. Timmer, T. Köhler, and M. Grass, ‘‘Movement
the motion artifact level is calculated, see Table IV. Level 0 HU, window artifacts in helical CT cone-beam reconstruction,’’ Proceedings of the
100 HU. Sixth International Meeting on fully three-dimensional image reconstruc-
tion in radiology and nuclear medicine, 2001, pp. 199–202.
8
K. Taguchi, ‘‘Temporal resolution and the evaluation of candidate
Since Kudo et al.13 observed similar results for the aspect algorithms for four-dimensional CT,’’ Med. Phys. 30, 640– 650
of cone-beam artifacts and dose utility by applying the fre- 共2003兲.
9
quency split method to a different cone-beam reconstruction G. Shechter, T. Köhler, A. Altman, and R. Proksa, ‘‘High-resolution im-
ages of cone beam collimated CT scans,’’ Nuclear Science Symposium
method,13 it seems to be a very generic approach to improve
Conference Record, IEEE, 2003.
image quality of approximative methods for low-pitch cone- 10
H. K. Tuy, ‘‘3D image reconstruction for helical partial cone beam scan-
beam CT. ners,’’ in Proceedings of the 3D’99 Conference 共Egmond aan Zee, The
As already pointed out by Kudo et al.,13 the additional Netherlands, 1999兲, pp. 7–10.
11
computational cost is relatively low, because the reconstruc- G. Shechter, G. Naveh, A. Altman, R. Proksa, and M. Grass, ‘‘Cardiac
tion of the two images can be in fact combined into a single image reconstruction on a 16-slice CT scanner using retrospective ECG-
gated, multicycle 3d back-projection algorithm,’’ Proc. SPIE 5032, 1820–
back-projection loop, all the geometrical calculations during
1830 共2003兲.
back projection have to be carried out only once. The pre- 12
E. Tanaka, S. Mori, K. Shimizu, E. Yosikawa, and H. Murayama, ‘‘Mov-
processing time is increased by a factor of approximately ing slice septa and pseudo 3-D reconstruction from multi-ring PET,’’
1.5, because the most time consuming part here is the Fou- Phys. Med. Biol. 37, 661– 672 共1992兲.
13
rier transformation and we only need an additional inverse H. Kudo, F. Noo, M. Defrise, and T. Rodet, ‘‘New approximate filtered
Fourier transformation of the projection data. backprojection algorithm for cone-beam helical CT with redundant data,’’
Nuclear Science Symposium Conference Record, IEEE, 2003.
14
M. Magnusson, ‘‘Linogram and other direct Fourier methods for tomog-
V. SUMMARY raphic reconstruction,’’ Tech. Rep. Ph.D. thesis No. 320, Linköping Uni-
The angular weighted WEDGE method shows artifacts versity, 1993.
15
for a simulated 64-row scanner if redundant data are used. T. Köhler, C. Bontus, K. Brown, D. Heuscher, M. Grass, G. Shechter, and
R. Proksa, ‘‘Evaluation of helical cone-beam CT reconstruction algo-
These artifacts are more serious if the pitch is chosen small. rithms,’’ Nuclear Science Symposium Conference Record, IEEE, 2002,
The artifacts show up as bright or dark shades with very low pp. 1217–1220.
frequency components within the xy plane. 16
D. Heuscher and K. Brown, ‘‘Redundant data and exact helical cone-
We introduced the frequency split method as a generic beam CT,’’ Proceedings of the 3D’03 Conference, Saint Malo, France,
method to suppress these artifacts. We applied it to the an- 2003, pp. Tu AM2– 4.
17
gular weighted WEDGE reconstruction algorithm and ana- R. Proksa, T. Köhler, M. Grass, and J. Timmer, ‘‘The n-PI-method
for helical cone-beam CT,’’ IEEE Trans. Med. Imaging 19, 848 – 863
lyzed it with respect to artifacts, noise, and sensitivity to 共2000兲.
patient motion. The method introduces some additional pa- 18
P. E. Danielsson, P. Edholm, J. Eriksson, M. Magnusson-Seger, and H.
rameters, which need to be selected carefully. In the case of Turbell, ‘‘The original PI-method for helical cone-beam CT,’’ Proceed-
a 64-row scanner, we found a set of parameters, for which ings of the 3D’99 Conference 共Egmond aan Zee, The Netherlands, 1999兲,
the method suppresses artifacts very efficiently without in- pp. 3– 6.
19
creasing the noise level. The algorithm appears to be quite M. D. Silver, K. Taguchi, I. A. Hein, K. S. Han, M. Kazama, and I. Mori,
‘‘Windmill artifact in multislice helical CT,’’ Proc. SPIE 5032, 1918 –
robust with respect to patient motion. For other scanner or
1927 共2002兲.
other base methods than WEDGE, other parameter values 20
M. Kachelriess, S. Ulzheimer, and W. A. Kalender, ‘‘ECG-correlated im-
may lead to the best trade-off between the mentioned quality age reconstruction from subsecond multi-slice spiral computed tomogra-
measures. phy scans of the heart,’’ Med. Phys. 27, 1881–1902 共2000兲.