You are on page 1of 11

IEEE TRANSACTIONS ON IMAGE PROCESSING. VOL. I. NO. I .

JANUARY 1992 77

Backprojection by Upsampled Fourier Series


Expansion and Interpolated FFT
Makoto Tabei and Mitsuhiro Ueda

Abstract-A fast backprojection method through the use of apply this technique to the above CT reconstruction, a
interpolated fast Fourier transform (FFT) is presented. The strong obstacle still exists to precise reconstruction. The
computerized tomography (CT) reconstruction by the convo-
lution backprojection (CBP) method has been proved to pro-
CT reconstruction requires the “de-blur’’ filtering of the
duce precise images. Specifically, the realization of “de-blur” projection data. The filter has a response proportional to
filtering by linear convolution is crucial for precise reconstruc- the absolute value of the frequency. The singularity of
tion. However, the backprojection part of the conventional CBP such a filter (the derivative of frequency response is not
method is not very efficient: it leads to high computational com- continuous at the origin) leads to a slowly decaying spa-
plexity, and arbitrary control of the frequency characteristics
tial response, and it causes significant aliasing with the
of the interpolation function is not possible. In this paper, we
propose an alternative approach to interpolating and backpro- DFT.
jecting the convolved projections onto the image frame. First, In contrast, the convolution backprojection (CBP) al-
the upsampled Fourier series expansion of the convolved pro- gorithm ingeniously separates the filtering operation into
jection is calculated. Then, using a Gaussian function, it is two steps: the projection data are linearly convolved with
projected by the aliasing-free interpolation of FFT bins onto a the de-blur filter in the spatial domain to realize the sin-
rectangular grid in the frequency domain. The total amount of
computation in this procedure for a 512 x 512 image is 1 / 5 of gular frequency response, then it is interpolated (low-pass
the conventional backprojection method with linear interpola- filtered) and backprojected onto the image frame [ 11-[3],
tion. This technique also allows the arbitrary control of the fre- [5]. This algorithm has been shown to be accurate and
quency characteristics. amenable to implementation. However, the backprojec-
tion is a very time consuming process, especially when
higher order interpolation is employed to realize good low-
I. INTRODUCTION pass filter characteristics. This computational complexity
is due to the fact that the contribution of the convolved
A COMPUTATIONALLY efficient method of comput-
erized tomography (CT) reconstruction that follows
directly from the projection theorem is to fill the two-di-
projection at each projection angle must be evaluated and
summed at each pixel in the image frame.
As a result, we propose an alternative approach to in-
mensional (2-D) Fourier space by the one-dimensional
terpolating and backprojecting the convolved projection
(1-D) transforms of the projections and then to take the
onto the image frame. The procedure is as follows. The
2-D inverse Fourier transform. This type of reconstruc-
convolved projection is transformed using a 1-D fast
tion technique relies on an interpolation of discrete Four-
Fourier transform (FFT) and is replicated at higher fre-
ier transform (DFT) bins to convert projection data from
quencies to perform upsampling [6], [7]. (The interpola-
a polar to a raster grid in the frequency domain. However,
tion generally requires upsampling before filtering out
the interpolation which is based on an ordinary fill-in con-
higher frequency components.) It is then multiplied with
cept does not work well with DFT. Therefore, it is prone
the frequency response of the interpolation function. The
to yielding an inferior reconstructed image [ 11-[3].
resultant sequence provides the Fourier series coefficients
Recently, we have proposed an interpolation technique
of the convolved projection to be backprojected. Each of
of the DFT bins based on a deconvolution technique using
the coefficients is given along a radial line in polar coor-
Gaussian functions [4]. This procedure can interpolate
dinates, and the Gaussian function is used to project them
DFT bins precisely, i.e., it produces arbitrary frequency
onto a rectangular grid. When the convolved projections
1- and 2-D sinusoids with the DFT. However. when we
for all angles have been projected in the frequency do-
main, a 2-D inverse fast Fourier transform (IFFT) oper-
Manuscript received May 26, 1990; revised October 23, 1990. ation is performed and correction of its result by division
M. Tabei was with the Research Laboratory of Electronics, Massachu- with the Gaussian function produces a reconstructed im-
setts Institute of Technology, Cambridge, MA. He is now with the Preci-
sion and Intelligence Laboratory, Tokyo Institute of Technology, Midori- age.
ku, Yokohama, 227 Japan. The computation required in this procedure for a 512
M. Ueda is with the International Cooperation Center for Science and x 512 image is 1 / 5 of that required for conventional
Technology, Tokyo Institute of Technology, Meguro-ku, Tokyo, 152 Ja-
pan. backprojection with linear interpolation. The use of any
IEEE Log Number 9104330. higher order interpolation functions does not affect the

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
78 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. I. NO. I. JANUARY 1992

computational complexity, since it only requires a re- Y


placement of the tables representing the frequency re- 4
sponse of the interpolation function.
In the following, we describe only the backprojection
part of the reconstruction procedure as the linear convo-
lution part is exactly the same as that described in the lit-
erature [ 11, [3]. In Section 11, we explain the interpolation
of the convolved projection through the use of its Fourier
series expansion. The fast evaluation of the Fourier series
sum at each pixel in the image frame, i.e., backprojec- t
tion, is discussed in Section 111. The results are shown in Fig. 1. Coordinate system for backprojection.
Section IV. In Section V we compare with other reported
work.
filter F to cause higher frequency components to decay
11. INTERPOLATION BY FOURIER SERIES smoothly. The result is that the basis for interpolation be-
In CT reconstruction by the CBP method, the recon- comes local rather than becoming the rippled and slowly
structed imagef(x, y), where (x, y) are the coordinates in decaying form of the sin (x)/x function.
the image plane, is given by the following backprojection If we choose F as a “half-band filter,” which has an-
of the convolved projection q(0, r ) : tisymmetric characteristics about the folding frequency,
this process using the DFT and IDFT satisfies the inter-
polation property, i.e., the weightings are 1 at the sam-
pling bin and 0 at the other bins [6]. This is a good char-
acteristic; however, it is not necessarily required in most
+y . sin (r 6)) applications, such as CT reconstruction [7]. On the other
hand, it is crucial that the frequency response converge to
double or higher order zeros to avoid the Gibbs phenom-
where NO is the number of projections, 7r(iO/No)is the enon by the truncation of the Fourier series.
discretized angular coordinate 0, and r is the coordinate The frequency characteristics of two half-band filters F2
which is given by the (signed) distance from the origin to and F3 are given as follows [6]:
the projection of (x, y) onto the projection line at angle
a ( i o / N O )(see Fig. 1). In the digital system, q ( r ( i o / N o ) ,
r ) is given only at the sampling points of r , i,. A r ( A r is
a sampling interval of r ) . Thus (1) requires the interpo-
lation of q with r from its values at the sampling points.
Here we interpolate through the use of a Fourier series
coefficient Q(?r(iO/Ne),kr), which is given by multiplying
the frequency response of the interpolation function F by
the DFT of q: The frequency response of these filters and their inverse
Fourier transforms (which serve as a basis for interpola-
tion) are shown in Fig. 2. F3 is almost identical to the
modified cubic spline function given in [8].
We also give the frequency response of the linear in-

exp ( - j 2 r e) terpolation FL to compare the reconstructed image with


that of conventional backprojection by linear interpola-
tion:

q (
R -;O’ r) =
Nr- 1

c
kr = - N r
Q (
R -:0’ ) (
k, exp j21r-N r L r k r )

(3)
_ -Nr A r Nr
Ir 5 - A r . (4)
2 2
The function FL is not a half-band filter and does not sat-
The relation between (2) and (3) is similar to that of the isfy the interpolation property when it is used in a finite
DFT and IDFT; however, ( 3 ) uses 2Nr terms of (2), which sum; however, it has the double zero at the sampling fre-
is twice as many as the Nr used by the IDFT. This enables quency, like F2 and F3.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
TABEI AND UEDA: BACKPROJECTION BY UNSAMPLED FOURIER SERIES EXPANSION 79

0.0 0.5 1.0 1.5 2.0

Normarized frequency
(a)
where G: is a 2-D Gaussian function in the wrapped and
normalized form which we define as follows:
G?,,(s,,s,, P , M,, M,)
m m

= c c
m, = --03 m, = -a

-3 -2 -1 0 1 2 3
Bin distance (./AT) P
(b)
(12)
Fig. 2. F z , F3 and their inverse Fourier transforms. (a) Spectrum. (b) In-
verse Fourier transform of (a). In this equation, only the first two arguments of G i . are
variables and P is a positive constant which is propor-
tional to the square of the width of the Gaussian function
Putting ( 3 ) into ( 1 ) with x and y discretized as i, A x and in the frequency domain. The last two arguments are in-
i, A y , we obtain the following: teger constants which are equal to DFT data lengths.

NO-I N,-I
T
f(i, A X , i, A y ) = -
No ig = O k , = - N ,

N, A r

L L

N, N,
-- si,<-
2 2
The direct evaluation of (8) over all points of i, and i,
requires 2NxN,NrNflsummations. This is a significant in- Equation (1 1) is based on the deconvolution technique;
crease in comparison with the N,N, No required by the therefore, it works well only over a limited range of val-
conventional backprojection in the spatial domain. There- ues of 0, i.e., if we choose to be excessively small, an
fore, the described procedure does not have any signifi- aliasing error caused by the periodicity of the DFT in-
cance without the existence of a fast computational pro- creases; and if we choose P to be excessively large, a
cedure for the sinusoidal functions. rounding error increases due to the compensation by the
exponential function after the DFT. However, if we
choose an appropriate value of 6, this process gives a sta-
111. FASTBACKPROJECTION ble and suprisingly precise value on the interval given by
The 2-D complex exponential function of frequencies (9) and (10).
f, and fy over the range given by (9) and (10) is synthe- In practice, this procedure still has the problem of eval-
sized through the use of a 2N, X 2N, point inverse DFT uating the infinite sum in ( 1 2 ) . However, we can truncate

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
80 IEEE TRANSACTIONS ON IMAGE PROCESSING. VOL I . NO. I. JANUARY 1992

the Gaussian distribution without affecting the result, if


the values become sufficiently small. (See Appendix A for -1
the discussion on the choice of 0 and truncation condi- ----
tion.)
Now we put (11) into (8), and obtain
f(ix A x , i, AY)

(b)
Fig. 3 . Circular shift of the data.
2N,-12Nr-1/ Nd-I Nr-l / . \

(a) (b)
Fig. 4. Generation of Fourier series coefficient. (a) DFT of q . (b)
Q(a(is/WU .

where

c, = N,
N,
~
e

*
AX
Ar

If we choose C, = Cy = 1 , the size of the reconstruction


region becomes equal to the size of the projection region.

IV. RESULTS
The convolved projection q of length N , can be calcu-
lated by the linear convolution of the projection data with (b) (C)
the filter function using a 2N, point FFT [ 11. Here we use Fig. 5. Backprojection in frequency domain. (a) Backprojection in fre-
a Shepp-Logan type filter function [ 5 ] . quency domain. (b) Backprojection before compensation. (c) Backprojec-
To calculate the Fourier series coefficients Q , con- tion after compensation.
volved projection q is circularly shifted by N , / 2 (Fig.
3(a)), and its N , point 1-D real FFT is used to give N , / 2 Ny according to (9) and (10) (see Fig. 3(b); only its central
+ 1 points of complex data. It is then replicated at the N, X N, points are shown in Fig. 5(b)). Its division by
higher frequencies by complex conjugate folding, and the Gaussian function produces the backprojection (Fig
multiplied by the frequency response of the interpolation 5(c)). In actual reconstruction, the convolved projection
function (Fig. 4(a)). for all projection angles are projected in the frequency
The resultant sequence is an N, +
1 point complex domain before the 2-D FFT.
Fourier series set of coefficients (Fig. 4(b)). It is projected The Gaussian function in the frequency domain is cal-
with blurring by the Gaussian function onto the 2N, X culated with 0 = 4.65 on a 7 x 7 point distribution. It
+
(N, 1) complex grid along the corresponding projection satisfies relative precision better than lop3 (see Appen-
line (Fig. 5(a)). The Fourier series extends to twice the dexes A and B). In this case, the magnitude of the expo-
folding frequency; therefore, the higher frequency com- nential correction term after the DFT is less than 3 in the
ponents exceed the boundary, which should be treated by reconstruction circle; therefore, the increase of rounding
complex conjugate folding or wrapping. error is negligible. Furthermore, note that while this cor-
Then a 2-D real inverse FFT is calculated to produce rection magnifies the arithmetic rounding error, it does
2N, x 2N, real data. It is then circularly shifted by N, and not magnify the noise in the projection data. (In general,

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
TABEI AND UEDA: BACKPROJECTION BY UNSAMPLED FOURIER SERIES EXPANSION

(b) =
F i g . 6. Reconstructed image. (a) -0.1-2.1. (b) 0.99-1.04.

TABLE I
R E C O N S T R U C TEI O
RNROR

Proposed
Conventional
(linear) interpolation F, F. F,

Inside Skull 8.421 x IO-' 8.357 x 10 7.983 x I0 ' 7 507 x IO-'


Whole Area 6.249 x IO-' 6.242 x I O - ' 5.859 x 10 ' 5.259 x IO-'

the effective bit length for arithmetic is far larger than that T A B L E I1
for projection data.) COMPUTATIONAL
T I M E(SECONDS 1)
BY SUN SPARKSTATION

A 2-D Gaussian function is separable to the product of ( A) (B)


two 1-D Gaussian functions which can be calculated by Conventional Proposed
second-order recursion [9] (Appendix B) . Therefore, the Nr X Ns Nr x N , B-P B-P (A)I(B)
major part of the computation is the generation of a 2-D 64 x 64 64 X 64 0.9 1.6 0.6
Gaussian function as the product of a complex Fourier 128 X 128 128 X 128 7.3 6. I 1.20
series coefficient and two 1-D Gaussian functions. It takes 256 x 256 256 x 256 58.4 25.0 2.34
512 X 512 512 X 512 465.5 fQ1.7 4.58
around 2 X 7 X 7 X N , 2: 100 N, real multiplications
and additions for every projection line. Therefore, the to-
tal number of operations is 100 A ",' multiplications and
additions. This is a significant reduction from the NoN,Ny before sampling. For bandlimited projection data, the ad-
multiplications and 3NBN,N, additions of conventional vantage of F3 will become more apparent.
backprojection with linear interpolation. Table I1 shows the computational times. From the above
Fig. 6 shows a 128 x 128 reconstructed image of the discussion, the proposed method should work 10 times
Shepp-Logan phantom with interpolation function FL. In faster than conventional backprojection for a 512 X 512
graphic representation, it looks exactly the same as the image. The difference is due to the computation of 1- and
image reconstructed by conventional backprojections with 2-D FFT and other miscellaneous computations.
linear interpolation. Table I shows a comparison of re-
V . DISCUSSION
construction errors. (The error is evaluated as the rms dis-
tance from the phantom.) F3 gives the best accuracy. We Many reports exist on CT reconstruction with FFT
used highly aliased projection data, which is produced by [lo]-[13]. In fact, some methods are even more compu-
direct sampling of the phantom without any prefiltering tationally efficient than the procedure presented in this pa-

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
82 IEEE TRANSACTIONS ON IMAGE PROCESSING. VOL I , NO I . JANUARY 1 Y Y 1

per [ 131. However, these methods are based on the con-


cept related to the sampling theorem that states that the
bandlimited signal can be reconstructed from the fre-
quency components up to folding frequency. Although
they generally work well, in practice, the assumption con-
cerning the bandlimiting conditions makes these proce-
dures somewhat vulnerable to aliased data and/or reduces
the resolving power in comparison with the CBP method.
In contrast, the proposed method is intended to mimic
the backprojection operation of the CBP method in the
frequency domain. The frequency response of the linear
interpolation function of conventional backprojection has
a first zero at the sampling frequency, i.e., twice the fold-
ing frequency. To achieve these characteristics, the DFT
of the convolved projection is extended up to the sampling
frequency by replication using complex folding. As a re-
sult, the proposed method becomes less susceptible to
aliased projection data while maintaining high resolving
power.
Fig. 7(a) and (b) shows the overall frequency charac- -e: ' -2I ' -1I ' ; ' ; ' ; ' ;
teristics and spatial response of the method. The singular Bin distance ( r / A r )
characteristic of the deblur filter is realized precisely by (b)
the linear convolution, and its repetitive images beyond Fig. 7. Overall frequency characteristics. (a) Overall frequency response.
the sampling frequency are filtered out by low-pass filter (b) Overall spatial response.
characteristics of interpolation. Note that the interpola-
tion does not have any response beyond sampling fre-
quency in constrast to the conventional CBP method. Ac- reported work is optimal in the sense that a minimum
tually, the results for the proposed method converge with number of distributions is required to attain the required
the results of the CBP method if we extend the Fourier precision. The method proposed in this paper is not op-
series expansion further with frequency response FL. timal in the same sense. It requires slightly more points
However, the reconstruction error does not improve, be- of distribution for the same precision because the Gaus-
cause the frequency components beyond the sampling fre- sian distribution never reaches zero, thus we need to cal-
quency do not contribute to the reconstruction at all. culate the value until it becomes negligible. For instance,
On the other hand, better reconstruction precision is for the relative precision of the Kaiser-Bessel func-
obtained by adjusting the frequency Characteristics below tion requires a 6 x 6 distribution and the Gaussian func-
the sampling frequency; for instance, the use of a half- tion requires a 7 x 7 distribution; for lop6, 12 x 12 and
band filter F2 or F3 gives better precision. This is the rea- 14 x 14 distributions are required, respectively. The dif-
son why the higher order interpolation function has been ference corresponds to a 36% increase in the grid points
introduced in the conventional CBP method. With the in the two dimensions.
proposed method, arbitrary frequency characteristics can The sampled values of the distribution must be gener-
be obtained without increasing computational complex- ated each time depending on the central position of the
ity. distribution. The Gaussian function has an advantage in
In the proposed procedure, prior to backprojection, the the practical sense, because it can be generated by a re-
projection data are processed by a 2Nr-point DFT and cursive procedure (see Appendix B), whereas the Kaiser-
IDFT to realize the linear convolution, and then its Bessel function requires interpolation of precomputed ta-
N,-point DFT is calculated. This back and forth operation bles. Our implementation of the procedure requires only
between spatial and frequency domains may seem redun- two multiplications per grid point whereas the reported
dant. However, in the strict sense, the deblur filter is only work requires seven multiplications. Their procedure does
realizable as a linear convolution because of its singular not utilize the fact that the distribution is separable in two
frequency characteristics at the origin; therefore, it is cru- directions; however, even with an improvement in that
cial to go back to the spatial domain once and abandon respect, their procedure requires more computations on
the wrapped portion of the convolved projection and then the whole. The published research has not reported the
calculate the DFT. actual computation time; however, we can infer it might
For the latter half of the method, a similar procedure be even slower than the conventional CBP method be-
for the generation of the arbitrary frequency sinusoidal cause of the significant increase in the amount of data dur-
function has been reported with the prolate spheroidal ing the filtering operation. With the Gaussian function,
function or Kaiser-Bessel function [ 121. This previously the optimization procedure is explicit, and we can evalu-

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
TABEI AND UEDA: BACKPROJECTION BY UNSAMPLED FOURIER SERIES EXPANSION 83

ate the necessary condition immediately from the required By replacing the infinite integral on the right-hand side
precision. of (16) with the infinite sum using Poisson’s sum formula,
The use of the Gaussian function has a further advan- we obtain the following [ 141:
tage. One can modulate the amplitude and frequency of
the synthesized signal because the procedure maintains the
natural characteristics of the Gaussian distribution [4].
This is quite useful in the analysis of many physical prob-
lems, and will be discussed in detail in a future publica-
tion.

VI. CONCLUSION
In this paper, we described the use of upsampled Four-
exp ( j 2 a g) .
ier series to simulate the CBP method in the frequency Here we express the discrete coordinates of x and U as
domain, and we also discussed its fast evaluation with an i x / 2 N x and k,, respectively, where 2N, is length of the
interpolated FFT. Although many reports exist on suc- 1-D (inverse) DFT. This equation states that by replacing
cessful CT reconstruction with the FFT technique, in the the integral of the Fourier transform with the summation,
strict sense, the accuracy of reconstructed images has not we obtain the wrapped form of the transform as a result.
been equivalent to that of the CBP method. The proposed Substituting t, = p , / 2 N , and dividing both sides of (17)
method improves the computational efficiency by an order by exp ( - rP(i,/2N,)2), we obtain the following:
of magnitude without loss of precision.
With recent advances in hardware, the computational
consideration has become less important. However, there
has remained a long-standing question: why is the CT re-
exp ( j 2 a i , t X ) + c
m

l x = - m ,# 0
exp (-rp 6,i + 1;))

construction by FFT not equivalent to that of the CBP


* exp ( j 2 a ( i x + 2N,l,)t,)
method? This paper gives one answer to that question.
Besides its computational efficiency, this procedure
gives versatility to the CT reconstruction. One aspect of
this versatility is the arbitrary choice of the frequency
characteristics of the interpolation. This provides an
adaptive filtering capability, depending on the signal-to-
noise ratio (SNR) of the projection data. Furthermore, this
exp ( j 2 r 5)
procedure may be directly applicable to the fast compu-
where G $ is defined as
tation of back-propagation algorithms.

A
APPENDIX
In the following, we will discuss the derivative of (1 1).
For the sake of simplicity, we will start the discussion (19)
with a 1-D case, and extend it to the 2-D case afterwards. For the convenience of the following discussion, we limit
The Fourier transform of a Gaussian distribution is also the range of i, as follows:
a Gaussian distribution [ 141. By shifting the center of the
distribution to p , in the frequency domain, the next rela- -N, Ii, 5 N, - 1. (20)
tion holds: Then the absolute value of the aliasing term, which ap-
pears as the summation on the left-hand side of (18), can
be evaluated as follows:

where x and U are the variables in the spatial and fre-


quency domain, respectively. is a positive constant that We neglect the terms other than the two major ones be-
is proportional to the square of the width of the Gaussian cause they have smaller values by several orders of mag-
distribution in the frequency domain. nitude.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
84 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. I . NO. I . JANUARY 1992

Now, we denote the desired precision for the result as


E.Then EA must satisfy the following in the interval given
by ( 9 ) (see Fig. 8).

E! takes the largest value at the both ends of the interval.


Therefore, E! should satisfy
Fig. 8 . Truncation of Gaussian distribution.
E!(-+) IE

and we obtain
2
p 2 - - log ( E ) . (24)
a

In practice, it is not possible to evaluate the infinite


Gaussian distribution of ( 1 9 ) . Typically, we wish to ap-
proximate it by calculating only NI points around its cen-
ter and by truncating the others. Fig. 8 illustrates the trun-
cation. The error resulting from this truncation is given
-112 -114 0 114 112
by
ixl(2Nx)
Fig. 9 . Aliasing and truncation errors.

* (sum at points outside of N I ) . (25) and takes its largest values at the extremes of this interval
(see Fig. 9 ) . Therefore, it follows that:
This is due to the fact that the DFT is bounded by the
absolute sum of the input data. The total sum at the out- E:(-:) IE . (29)
side is dominated by the values of the first two points just
outside the NI points. With a typical value of 0,the next For simplicity, we neglect p'" in the denominator of ( 2 7 ) ,
adjacent values are less than lo-* of the inner neighbors and substitute ( 2 4 ) and ( 2 7 ) into ( 2 9 ) , to obtain
and the ratio becomes even larger as it gets further from
the center. The distance d from the center of the Gaussian
distribution to the first truncation is given by
In practical implementation, it is advantageous to choose
N a value for /3 which is as large as possible but which does
- r1 d ~ - N
+I 1 . not increase N, to the next integer value, because such an
2 2
increase will allow some margin for error without increas-
The largest error occurs when the truncation is most as- ing the amount of necessary computation.
symetrical. When it is symetrical, i.e., when d is bal- The roundoff error of the arithmetic should also be taken
anced at both sides, the error sum is much smaller due to into consideration. The roundoff error E : is given by
the convergent nature of the Gaussian distribution. Thus
the largest value of E: is evaluated as follows:

where E , is a representation error of floating point numbers


because

E: also has to satisfy the following in the interval given


by (9):

E, ( I): ;
exp j217 - = E,.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
~

TABEI AND UEDA: BACKPROJECTION BY UNSAMPLED FOURIER SERIES EXPANSION 85

The typical values of E , is 2 - 2 4 for the IEEE single pre-


cision floating point number and the magnitude of E : is
usually much smaller than Ef,or E f . However, E: im-
poses a certain limitation on the achievable precision E .
The roundoff also takes the largest value at both sides of As a result, the values of /3 and NI should be chosen with
(9). Substituting (24) into (31) gives the same limit as for the 1-D case.
The roundoff error in the 2-D case is given in a some-
El (- i)2 E-(’/’) * E,. (33) what different way as follows:

Therefore:
E 2 E-(l/8) . E,. (34)
Then. hence:

E 2 €;I9. (35) This is because, due to the random nature of the roundoff
As a result, the achievable precision is limited to 8 / 9 of error, the errors given by the implementation of (36) and
the arithmetic bit length. (37) are different. (Note that (36) requires fewer numbers
Now we can discuss the 2-D case, using the results of of operations than (37); however, (36) cannot be used for
the above discussion. The 2-D sinusoidal function with the calculation if there is more than one Gaussian distri-
frequencies t, and ty is calculated by bution.) However, in practice, the roundoff will not yield
significantly different results from the simple sum of (31)
exp ( j27r(ixt, + i,t,)) unless we choose an excessively large 0.(The reason for
the near equivalency of results is that since we evaluate
E ; as the absolute sum of the input to the 2-D DFT in
= exp (j27rixt,) exp ( j 2ai, ty) analysis, the practical error becomes much smaller by the
2Nr - I
statistical effect of different signs of many roundoff com-
exP (TO ($)2) k,c =0 Gk(k, - 2Nxtx,P , 2N,)
ponents. of the 2-D distribution.) We found by numerical
experiment that the condition given by (35) is also suffi-
cient in a 2-D case.

B
APPENDIX

* G,!& - 2Nyty,P , 2NJ * exp (j27r $) (36)


Fig. 10 shows a C subroutine which generates a Gaus-
sian distribution by second-order recursion. The subrou-
tine starts the recursion from the center of the distribution
and works out to either side. (Otherwise, although it may
not be a serious problem for CT reconstruction, the sec-
ond-order recursion produces a fairly large numerical er-
ror, which is proportional to the square of the number of
steps [9].) The 2-D distribution is generated as the prod-
uct of two 1-D distributions in orthogonal directions. Fig.
1 l(a) and (b) show the error distribution of the generated
(37) complex exponential function using this subroutine (the
folding part was omitted from the subroutine) followed by
a 64 x 64 2-D FFT. (We generate a complex exponential,
The aliasing error E: and truncation error E : in the inter- rather than a real sinusoid, because the error from real
vals in (9) and (10) are given by the following: sinusoids has many zeros and is difficult to show in graph-
ical representation. The error from real sinusoids is
bounded by the error of complex exponentials.) The am-
plitude of the synthesized complex exponential is 1, and
the error is evaluated as the absolute value of the differ-
ence with the complex exponential function calculated by
* (1 + EA(&)) - 1 double precision arithmetic. The frequencies are t, =
14.O1/2Nx and ty = 2 1 . 5 / 2 N y for both cases. and NI
are 4.65, 7 , and 9 . 3 , 14, respectively. The error distri-
bution is not affected significantly by the choice of fre-
quencies.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
86 IEEE TRANSACTIONS ON IMAGE PROCESSING. VOL. I . NO. I . JANUARY 1992

#include <stdio.h>
#include <math.h>
#define real float
#define PI 3.141592653589793238
typedef struct cmplx
( real re; real im; ) complex;
void gaus2d( gw, mx, my, beta, sx, sy. q )
I* This subroutine adds one Gaussian distribution to the two
dimensioal array "gw[]" on each call. The array should be
cleared externally before the first call.
To form the image, we process the resulting array "gw" with
two dimensional real inverse FFT, rearrange the area (swap
first half and latter half in each dimension), and divide
by Gaussian function (See text for detail).
gw[] : Two dimensional complex array which holds the
spectrum of the image.
mx, my : Size of "gw[]". Physical size of "gw(]" will be
'lmx+(my/2+1)" when spectrum is folded for real
valued image.
beta : Squared width of Gaussian function in spectral domain.
sx, sy : Coordinate of the center of the Gaussian function.
q : Complex amplitude of the Gaussian function.

"mx" , "my" and "beta" are supposed to be identical on


each call.
*I
complex g w [ l . q;
real beta, sx, sy;
int mx, my;
i
real vx[20], vy[20], vy-re, vy-im;
real dx, dy. wd, el-m, el-p, e2;
int nt, jx, jy, kxO, kyO, kx, ky, kx-f;
nt = ceil( 1.5 * beta ) ; I* "nt" : number of distributions. *I
wd = -
PI 1 beta;
e2 = exp( 2.0 * wd ) ;
/ * These three variables are identical * I
I* on each call. *I
kxO = ceil( sx - 0.5 nt ) ; I * (kx0,kyO) designates the lower left * I
kyO = ceil( sy - 0.5 nt ) ; I * indices of mapping area. *I
dx = ( kxO + nt 1 2 ) - sx; / * -0.5 <= dx. dy < 0.5, "nt" : odd. * I
dy = ( kyO + nt 1 2 ) - sy; I * 0.0 <= dx, dy < 1.0, "nt" .' even. * I
if( ( kxo = kxo % mx ) < 0 ) kxO += mx; /* 0 <= kxO <= mx-1 */
if( ( kyO = kyO % my ) < 0 ) kyO += my; I* 0 <= kyO <= my-1 *I
I * Generate two 1-D Gaussian distributions * I
vx[ nt 1 2 ] = exp( wd * ( dx * dx + dy * dy ) ) 1 beta;
vy[ nt I 2 J = 1.0;
el-m = exp( 2.0 * wd * ( 0.5 - dx ) ) ;
el-p = e2 I el-m;
for( jx = nt 1 2 - 1; jx >= 0; jx -- ) ( / * downward recursion * I
vx[ jx ] = vx[ jx + 1 ] * el-m;
el-m = el-m * e2;
1
for( jx = nt 1 2 + 1; jx < nt; jx ++ ) ( / * upward recursion * I
vx[ jx ] = vx[ jx
el-p = el-p * e2;
1 ] el-p; -
>
el-m = exp( 2.0 * wd * ( 0.5 - dy ) ) ;
el-p = e2 1 el-m;
for( jy = nt I 2 - 1; jy >= 0; jy -- ) ( / + downward recursion * /
vy[ jy 1 = vy[ jy + 1 1 e1-m; *
el-m = el-m e2;
>
for( jy = nt 1 2 + 1; jy < nt; jy ++ ) ( / * upward recursion * I
vy[ jy 1 = vy[ ju
el-p = el-p * e2;
1 1 * e1-p; -
}

/ * Map 2-D Gaussian distribution on "gw[l" * I


ky = kyO;
for( jy = 0; jy < nt; jy ++ ) (
v-re = q.re * vy[ jy I;
vy-im = q.im * vy( jy 1;
if( ky <= my 1 2 ) ( / * check folding * I
kx = kxO;
for( jx = 0 ; jx < nt: jx ++ ) {
qw[ mx ky + kx ].re += vy re * vx[ jx I;
g w [ mx * ky + kx ].im += vyIim * vx[ jx 1 ;
if( ( ++ kx ) == mx ) kx = 0;
)

} else ( / * delete following if folding is not required * I


if( kxo kx-f = 0;
== 0 )
else kx-f = mx - kxO;
for( jx = 0 ; jx < nt: jx ++ ) (
gw[ mx 4 ( my -
ky ) + kx f ].re += vy re * vx[ jx I;
gw[ mx * ( my - ky ) + kxIf ].im -= vyIim * vx[ jx I ;
if( ( -- kx-f ) < 0 ) kx-f = mx - 1;
)

if( ( ++ ky ) == my ) ky = 0;
1
)

Fig. 10. A C subroutine to generate Gaussian distribution.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.
TABEl A N D U E D A : B A C K P R O J E C T I O N BY U N S A M P L E D F O U R I E R S E R I E S E X P A N S I O N 87

[6] R. E. Crochiere and L. R. Rabiner, Multirare Digital Signal Pro-


cessing. Englewood Cliffs, NJ: Prentice-Hall, 1983.
[7] H. W. Schussler and P. Steffen, “A hybrid system for reconstruction
of a smooth function from its samples,” Circuits, Syst. Signal Pro-
cessing, vol. 3, pp. 295-314, 1984.
[8] G . T. Herman, S . W. Rowland, and M . Yau, “A comparative study
of the use of linear and modified cubic spline interpolation for image
reconstruction,” IEEE Trans. Nucl. Sci., vol. NS-26, pp. 2879-2894,
1979.
[9] L. R. Rabiner, R. W. Schafer, and C. M. Rader, “The chirp
z-transform algorithm,” IEEE Trans. Audio Electroacoust, vol.
AU-17, pp. 86-92, 1969.
[IO] R. M. Mersereau and A. V. Oppenheim, “Digital reconstructions of
multidimensional signals from their projections,” Proc. IEEE, vol.
62, pp. 1319-1338. 1974.
[ I I] H. Stark, J. W. Woods, 1. Paul, and R. Hingorani, “An investigation
-N, -3
2
0 -
Nr
2 Nz
of computerized tomography by direct Fourier inversion and optimum
interpolation,” IEEE Trans. Biomed. Eng., vol. BME-28, pp. 496-
2, 505, 1981.
[ 121 J. P. O’Sullivan, “A fast Sinc function gridding algorithm for Fourier
inversion in computer tomography,” IEEE Trans. Med. Imaging, vol.
MI-4, pp. 200-207, 1985.
Yy [I31 P. Edholm, G . T. Herman, and D. A. Roberts, “Image reconstruc-
tion from linograms: Implementation and evaluation,” IEEE Trans.
Med. Imaging, vol. 7, pp. 239-246, 1988.
4
- [I41 S. Papoulis, Systems and Transforms with Applications in Optics.
2 New York: McGraw-hill, 1968.

i, 0
M a k o t o Tabei was born in Osaka, Japan, on Feb-
ruary 12, 1956. He received the B . E . degree in
_ -N Y information engineering from the University of
2 Electro-Communications, Tokyo, Japan, in 1979,
and the M.E. degree in information processing
from the Tokyo Institute of Technology, Tokyo,
- NY Japan, in 1981.
-N, --N*
2
, -
Ns He worked at the Japan Atomic Energy Re-
2 Nz search Institute in Ibaraki, Japan, in 1981. Since
1982, he has been a Research Associate at the Pre-
cision and Intelligence Laboratory, Tokyo Insti-
(b)
tute of Technology. During 1989-1990 he was a Visiting Scientist at the
Fig. 11. Error contour map of synthesized complex exponential. (a) 0 = Research Laboratory of Electronics, Massachusetts Institute of Technol-
4.65, N , = 7. ( b ) p = 9.3, N , = 14. ogy, MA. His research interests include numerical analysis of wave phe-
nomena and signal processing in acoustical applications.

ACKNOWLEDGMENT
The authors would like to thank the members of DSPG, Mitsuhiro U e d a was born in Tokyo, Japan, on
MIT, for their assistance and encouragement. January 13. 1942 He received the B.E. degree in
instrument engineering from Keio University,
Yokohama, Japan, in 1964, the M.E. degree i n
REFERENCES electrical engineering and the Dr.Eng degree in
control engineering from Tokyo Institute of Tech-
[ I ] A C Kak, “Computerized tomography with x-ray, emission, and nology, Tokyo, Japan in 1966 and 1969, respec-
ultrasound sources,” Proc. IEEE, vol 67, pp 1245-1272, 1979 tively
[21 R M. Lewitt, “Reconstruction algorithms Transform methods,” From 1969 to 1972 he was a Research Associ-
Proc. IEEE, vol 71, pp. 390-408, 1983 ate and in 1972 he became a Associate Professor
131 A . K. Jain, Fundamentals of Digital Image Processing Englewood of the Precision and Intelligence Laboratory at the
Cliffs, NJ: Prentice-Hall, 1989 Tokyo Institute of Technology He is currently a Professor at the Interna-
141 M Tabei and M Ueda, “FFT multi-frequency synthesizer,” in Proc. tional Cooperation Center for Science and Technology, Tokyo Institute of
ICASSP-88, Apr 1988, pp 1431-1434 Technology His current research interests are in the areas of theoretical
[5] L A Shepp and B F. Logan, “The Fourier reconstruction of a head analysis of wave propagation. the inversion problem of scattering, and sig-
section,” IEEE Trans Nucl. Sci , vol NS-21, pp 21-42, 1974 nal processing.

Authorized licensed use limited to: Proof and Experimental Establishment. Downloaded on December 26, 2009 at 10:41 from IEEE Xplore. Restrictions apply.

You might also like