Professional Documents
Culture Documents
Stark1981 PDF
Stark1981 PDF
C. E. Shannon and W. Weaver, The Mathematical Theory of Com- [31] J. D. Markel and A. H. Gray, Jr., Linear Prediction of Speech.
munication. Chicago, IL: Univ. Illinois Press, 1949. New York: Springer-Verlag, 1976.
M. S . Bartlett, An Introduction to StochasticProcesses. Cam- [32] Y.Yaglom, An Introduction to the Theory of Stationary Ran-
bridge, England: Cambridge Univ. Press, 1966. dom Functions. Englewood Cliffs, NJ: Prentice-Hall, 1962.
R. G. Gallagher, Information Theory and Reliable Communica- [33] A. Papoulis, SignalAnalysis. New York: McGraw-W, 1977.
tion. New York: Wiley, 1968. [34] D.C. Youla, “The flatechoestimator: A newtunable high-
E. T. Jaynes, “Prior probabilities,” IEEE Trans. Syst. Sci. Cybern., resolution spectral estimator,” Polytech. Inst. New York, Tech.
v01. SSC-4, pp. 227-241,1968. Note 3, Contract F30602-78-C-0048.
W. M. Elsasser, “On quantum measurements and the role of the
uncertaintyrelationsin statistical mechanics,” Phys.Rev., vol.
52, pp. 987-999, NOV. 1937.
E. T. Jaynes, “Information theory and statistical mechanics I,”
fhys. Rev., vol. 108, pp. 171-190,1957.
I. J. Good, “Maximum entropy for hypothesis formulation, es- John E. Shore (M’72-SM’Sl) was born in
pecially formultidimensionalcontingencytables,” Ann.Math. England on September 2, 1946. Hereceived
Statist., vol. 34, pp. 911-934, 1963. the B.S. degree in physics from Yale University,
A. Hobson and B. Cheng, “A comparison of the Shannonand New Haven, CT, in 1968, and the Ph.D. degree
Kullback information measures,” J. Statist. Phys., vol. 7, no. 4, in theoretical physics fromthe University of
pp. 301-310, 1973. Maryland, CollegePark, in 1974.
I. J. Good, Probability and the Weighing of Evidence. London: In 1968 he joined the Naval ResearchLab-
Griffen, 1950. oratory, Washington, DC, where he is currently
R. W. Johnson, “Determining probability distributions by maxi- Head of the Computer Sciences Section in the
mum entropy and minimum crossentropy,” Proc. APL 79 (ACM Computer Science and Branch of the Informa-
0-89791-005), pp. 24-29, May 1979. tion Technology. His research interests include
L. Geronimus, Orthogonal Polynomials. New York: Consultants :mation theory, queuing theory, computer system
signal processing, info1
Bureau, 1961. performance modeling:,and softwareengineering.
Absrract-Direct Fourier transform (FT) reconstruction of images inconversion. Although, in practice, approximations to exact in-
computerized tomography (CT) is not widely used because of the dif- terpolation may be required for reasons of economy and
ficulty of precisely interpolating from polar to Cartesiansamples.In
speed,the underlying theory furnishes both a guide and a
this paper, an exact interpolation scheme is proposed which, in prac-
standard against which other schemes can be compared.
tice, can be approached with arbitrary accuracyusing well-conditioned
algorithms. Several features of the direct FT method are discussed. A The organization of the paper is as follows. First, in Section
method that allows angular band limiting of the data beforeprocessing
I1 we review the theory that underlies the direct Fourier tech-
-to avoid angular aliasing artifacts in the reconstmcted image-is dis-
niques in CT. Next, in Section I11we discuss the theory be-
cussedand experimentally verified. The experimental results demon-
strate the feasibilityof direct FT reconstruction of CT data. hind the exact interpolation formulas. In Section IV we dis-
cuss extension of the results of Section I11 to a more general
class of functions. In Sections V and VI, sources of computa-
I. INTRODUCTION tional error and methods to ameliorate the effects of these
N this paper we present the results of an investigation into errors aregiven. Finally, some preliminary experimental re-
I computerized tomography (CT) image reconstruction using sults are furnished in Section VII.
direct Fourier techniques. The investigation is based on some
new analytical results [ l ], [2] which under certain conditions 11. BASIC THEORY
provide exact interpolation formulas for polar to Cartesian CT reconstruction by direct Fourier techniques i s based on
the projection theorem (as is the convolution back-projection
Manuscript received November 7, 1979; revised September 22, 1980. method). The theorem states that the Fourier transform of a
This work was supported by the General Electric Corporate Research projection is a center cross section of the Fourier transform of
Division and the NationalScience Foundation.
The authors are with the Department of Electrical and Systems En- the image. Thus, if p ( x , y ) represents the absorbtivity func-
gineering, Rensselaer Polytechnic Institute, Troy,NY 12181. tion of the object with respect to a fixed coordinate system
0096-3518/8l/0400-0237$00.75OIEEE
238 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-29, NO. 2 , APRIL 1981
x-y and /.to (2,9)represents the object in a coordinate system circularsampling theorem (CST). The theorem can be stated
2-9 rotated from x-y by an angle 4, then the projection of the as follows.
object at view angle 4 is defined as Theorem: Let x ( t ) be periodic with highest frequency
f,, = K/T where T is the period and K is a positive integer.
P @ G ) 6 J/.t@(P,9) du^ (1) If the intersample period T, satisfies
L
p ( x , y ) =JJWM,(u, v) e j 2 n ( u x + v y ) dudv. (4) Next we observe that if the object is space limited to 2 A , its
-00 projection will also be limited in extent to 2 A and therefore
theFT of the projection P$(u) canbe reconstructed from
Perhaps the most serious difficulty with the above approach
samples uniformly spaced Au < 1/2A apart. The ideal sam-
is that M ( p , @)is not known everywhere but only on a finite
pling reconstruction is
discrete set of points { p i , @ k } . The problem then becomes one
of interpolating from the known values at the polar points to m
the values required over a rectangular Cartesian grid which al- P@(U) = P$ - sinc [ 2(u~- n / 2 ~ ) ] .
n=-m ( 2 2 (10)
lows the approximate realization of (4) via an inverse 2-D FFT
routine. For certain classes of functions, one can develop ex- Now using (3), (7), and (10) we obtain the final result:
act interpolation formulas forthis problem. We discuss this
matter below.
On
+7T
present although its effect can be ameliorated by choosing N Thus, the approximate required number of views to achieve a
large enough. In addition, the finite width of the X-ray beam certain angular resolution is related to the size of the object
produces a smoothed projection function which is somewhat and the highest spatial frequency required in the reconstruc-
akin to passing the projection function through a low-pass fil- tion. This result agrees with what is given in the literature [3].
ter. Hence, in practice, the amount of aliasing can be made Another consideration in choosing K (or N ) relates to the
tolerably small without great difficulty. polar-to-Cartesian interpolation problem. Assume for simplic-
ity that a nearest-neighbor type interpolation is used. Then if
B. Inadequate Sampling in Radial Frequency the number of samples N in a centralslice of the Fourier plane
Inadequate sampling in radial frequency can, of course ,pro- is very large, the bandwidth B = NAu/2 will be large and the
duce additional aliasingif the sampling rate is less than the “spokes” in the frequency plane willgo far out. If K is too
Nyquist rate of 2 A points/cycle. However, assuming that this small, nearest-neighbor interpolation at high radial frequencies
is not a problem, we must still consider twoother effects. could cause significant errors since the “nearest neighbors”
First, if the Fourier samples extend out to a radius B and are will in factbequitedistant.Thus,fora fixed K , one may
zerobeyond,the effect is the sameasif an ideal low-pass want to decrease the bandwidth, i.e., decrease N to reduce in-
truncationfunction were overlayed onan infinite array of terpolation-related artifacts. An alternative to decreasing the
samples. The effective bandwidth of the reconstructed image bandwidth when the number of views is considerably less than
is NAu/2, where Au is the frequency interval spacing and N is required by (31) or (32),is to employ the low-pass angular fil-
the number of samples along a central slice in the Fourier tering of Section N to eliminate possible artifacts. This will
plane. Clearly, if N is too small, a low-pass filtered image is result in an effective reduction in angular bandwidth or resolu-
reconstructed. Second, inadequate sampling in radial fre- tion, while leavingradial bandwidth and resolution intact.
quency can also produce interpolationartifacts. However,
these are more difficult to discuss since they depend on the VI. NEAR-OPTIMAL INTERPOLATION FILTERS
nature of the interpolation, the number of views and the maxi- A. Truncated Cardinal Function Expansion
mum angular frequency K .
In the computer implementation of the optimal interpola-
C. Insufficient Number of Views tion formula (1 1 ) it is necessary to truncate the interpolation
For many nonisotropic or off-centered objects, the theoret- functions about their largest values. Thus, practically, (1 1) is
ical value of K may be very large or infinite. However, it is replaced by
always possible to determine a K value that will result in very
good reconstructions. This value is based on the radius of sup-
port A and the approximate isotropic bandwidth B. In deter- n=np-L, k=k@-L@
mining a suitable value of K , several factors need to be con-
sidered. First consider an object of extent 2 A . The maximum
Cartesian grid spacing in frequency is Au = 1/(2A). If B is an
approximate bandwidth for the image, then B is the approxi-
mate radius of the nonzero portion of the function on the where the nearest neighbors [2Ap.] 4 n, and [(K + 1 ) @/TI
A
Fourier plane and the maximum angular separation is obtained = k@.
from If L , = LG = 0, the estimate reduces to the nearest neighbor
interpolation. For all values of L , and LG, the interpolation is
Au = BAG an FIR filtering of the polar samples. Since the filter coeffi-
or cients are truncated cardinal functions, it is possible to alter
. .
these values to optimize over the interpolation error.
Belowwegive an analysis of the case L , = LG = 0 corre-
sponding to nearest neighbor interpolation. This is followed
Now A @is given by by a discussion of the use of optimal FIR interpolators to im-
2n n prove on theapproximation (33).
A$ = - (30)
number of samples number of views * B. Nearest Neighbor Interpolation: Approximate Analysis
Also, the smallest allowable number of views that will recon- An exact computation of nearest neighbor (NN) noise in the
struct an angular frequency K is K + 1 . For large K , K + 1 =K reconstructed image appears to be a difficult task. However,
so that A@N n/K = 2/N from (29). some idea of the nature of this noise can be obtained by con-
Hence , sidering the following “dual” problem. We shall consider the
1-Dcase only since our primary objective is to obtain a quali-
tative understanding of the process. The dual problem con-
siders the noise resulting from a jitter introducedin an equally
= 2nAB. spaced Fourier sequence from which the image is recon-
242 IEEE
TRANSACTIONS
ON
ACOUSTICS,
SPEECH,
AND
SIGNAL
PROCESSING,
VOL. ASSP-29, NO. 2 , APRIL 1981
structed. In the actual problem, the jitter is introducedby From (37), wesee that the noise in the reconstructed image
associating the FT values at unequally spaced points with can be reduced by increasing h (i.e., increase the sampling rate
those at equally spaced points. which brings the neighbors closer to the point and thereby re-
Consider thenthe following problem. Suppose that F(u) duces the error). When the slope F’(k) is constant, i.e., inde-
represents the Fourier transform of a space-limited function pendent of k , then the mean-squared error of the frequency-
of support 2A. In particular we want to reconstruct what plane noise is independent of k. Hence, the noise is white and
we shall call the ideal image using the IDFT of the sequence R ( m - 1 ) is zero everywhere over a cycle except when m = 1.
3 ( k ) & F [ k / 2 h A ] where h > 1. Now assume that instead This is easily verifiedfrom (39).
of obtaining the sequence F [ k / 2 A ] we obtain the sequence In image reconstruction by direct Fourier methods, we have
F [ k / 2 h At Ek] where Ek is a uniform random variable (RV) foundthat NN-interpolation produces both high-frequency
with probability density function (pdf) noise (indicating a constant slope component in the Fourier
transform) and relatively large artifacts indicated strongly
:1
1 - AU
- < Ek <-Au correlated components of noise.
P(Ek) 2
= ’ 2 (34)
otherwise C. FIR Interpolators
Here we consider the approximation to (11) that results
and Au 2 1/2hA.
from replacing the cardinal functions in (33) by optimal FIR
+ Ek), where pk & 2ek h A .
NOW define 3 (k t p k ) & F ( k / 2 h A
interpolators [4]. In place of the continuous ( p , 4) in (33),
Note that pk = 2EkhA is a uniform RV such that I p k I < $. If
the FIR interpolators will produce the output on an interpo-
h is large enough, then to a first approximation
lated polar grid that is, say, I times as dense as the one on
F(Uk t Ek) E F(Uk) t EkFr(Uk) (3 5) which A4 is sampled. One then resorts to nearest neighbor
interpolation on this finer grid to actually obtain the required
or
Cartesian samples.
3(k t p k ) 3(k) -I- 3‘ ( k ) (36 ) Iffor I-foldinterpolation we have theFIR interpolating
2hA
functions
where F’(Uk) is the first derivative of F(u) evaluated at Uk =
k / 2 h A and 3’(k) 4 F’(k/2hA).
Now define the sequence { f(m)} given by
N- 1 then (33) is modified to become
f ( m )= 3 ( k )ej(2n/N)mk m = 0,. . . ,N - 1
k=O
n ( m )= Pk F’(k) ej(Zn/N)mk
- m = 0, . ,N - 1
k=O 2hA
where
(37)
be the noise due to the random displacement of the required
points fromthe given points in the frequency domain. The
correlation function of the noise is and
-
N-1N-1 p k pi
R(m - o = n(m>n*(l>= 4h2A2
k=O i = O
Through the use of optimum Chebyshev interpolators, the
error in the interpolated samples can be reduced over that in
We assume that (33). The Gibbs phenomena associated with thetruncation
of the cardinal functions is eliminated. Computationally, the
calculation of the sinc and (T functions isreplaced by table
look-up of the appropriate fdter coefficients. A disadvantage
(0, otherwise. to using the optimal interpolators is that one must ultimately
resort to nearest neighbor interpolationonthe fine output
Then
grid. This last step can cause the accuracy advantage to shift
back to thetruncated Cardinal expansion if I is not large
enough. To reduce this effect one can
1) design the FIR filters to interpolate by a larger factor;
2 ) design one or both of the low-pass filers in the interpo-
lator equation (40) to band limit below the associated Nyquist
frequencies to introduce an additional smoothing between
samples.
STARK et al.: RECONSTRUCTION IN COMPUTER TOMOGRAPHY 243
r
Such a comparative study is recommended for further investi-
gation.
M.CONCLUSIONS
We have addressed the problem of direct Fourier transform
reconstruction of images in computerized tomography and
have derived an exact interpolation algorithm to convert dis-
crete polar data to Cartesian data. In practice, exact interpola-
tion can be ap.proached with arbitrary accuracy provided that
the projection data are first angularly band limited. A practical
method for achieving the required band limiting was described
and experimentally verified on images reconstructed by the
direct Fourier method. Image reconstruction using the direct
Fourier technique was then compared with reconstruction by
FCBP.Theimageswere found to be of comparable quality
proving thatthe direct Fourier technique-requiring only
O(N2 log N ) computations-is an effective and efficient image
reconstruction procedure.
REFERENCES
Indraneel Paul (S’78) was born in Calcutta, In- 2-D recursive filter design, image modeling and estimation, and com-
dia, on June 11, 1956. He received his B.Tech. puted tomography.
(Hons.)degreein electrical engineering from
the Indian Institute of Technology, Kharagpur,
in 1978 and the M.S. degree in electrical engi-
neering from Rensselaer Polytechnic Institute,
Troy, NY, in 1979.
He is at present working towards his doctor-
ate at R.P.I. His research activities and interests Rajesh Hingorani, photograph and biography not available at the time
include information and rate-distortion theory, of publication.
A6smcr-This paper presents new methods for thedesign of optimal digital ones. On theotherhand,themethods in the time
and suboptimal recursive digital fiiters in the time domain. The pro- domain are based on the approximation of the given desired
posed methods are based on the approximation of the desired impulse
response of a digital fiiter. Since the duration of impulse response of a
impulse response [3] - [ 9 ] .
recursive digital fiiter is not fiiite, itis desirable to use an infiiite inter- This paper presents new methods for the design of optimal
val performance index from the point of stability. First, a method for and suboptimal recursive digital filters in the time domain.
the design of an optimal recursive digital filter is presented. The opti- First, a method for the design of an optimal recursive digital
mal solution is obtained by minimizing the given infinite interval per- filter is presented. The recursive digital filter is expressed by
formanceindex. Multivariable optimizationtechniques such as the
Fletcher-Powell method can be applied to obtain the optimal solution.
a state equation, andan infinite interval quadratic performance
However,when the order of a filter becomeshigh, it is difficult to index is minimized. Since the impulse response of a recursive
obtain the optimal solution. Next, suboptimal recursive digital filters digital filter does not vanish in a finite interval, it is desirable
with parallel and cascade structures areproposed. When these methods to use an infinite interval performance index from the point
are used, it becomes easy to design ahigh-ordersuboptimal filter. of stability. Although the preceding methodscannot be
Various numerical examples areshown to illustrate the results indetail.
applied to the infinite interval performance index, the optimal
parameters of a recursive digital filter can be obtainedby using
I. INTRODUCTION the proposed method. It is important to note that the for-
T HE role of digital filters in the area of digital signal pro- mulas become simpler when the interval of the performance
cessing has become more important with the development index is infinite as shown in this paper. Nevertheless, when
of digital techniques. It is known that digital fiters can be the order of a filter becomes high, it is difficult to obtain the
designed in the frequency and time domains [l], [ 2 ] . The optimal solution. Next, two methods for the design of sub-
most usual methods in the frequency domain are based on the optimal recursive digital filters are presented. When these
change of variables from Laplace transform to z-transform. methodsare used, high-order suboptimal recursive digital
These are the straight transformations from analog filters to fdterswith parallel and cascade structures can be designed.
If two mth-order fiters are connected in parallel or in cascade,
Manuscript received May 15, 1979;revised September 30, 1980.
The author is with the Department of Electronic Engineering, Chiba then the suboptimal value of the performance index is always
University, Chiba, Japan. smaller than the optimalvalue of one mth-order filter.