IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 12, NO.

4,DECEMBER 1993 653

Reducing Negativity Artifacts in

Emission Tomography: Post-Processing
Filtered Backprojection Solutions
Finbarr 0’Sullivan, Yudianto Pawitan, and David Haynor

Abstruct- The problem of negative artifacts in emission to- reconstruction methods. Relative to x-ray computed tomogra-
mography reconstructions computed by filtered backprojection phy (CT), emission tomography resolution is limited by low
(FBP)is of practical concern particularly in low count studies. count rates, particularly in dynamic studies [15], and thus the
Statistical reconstruction methods based on maximum likelihood
(ML) are automatically constrained to be non-negative but their statistical aspects of the reconstruction process are potentially
excessive computational overhead (orders of magnitude greater quite important [ 161. Indeed, over the past decade a substantial
than FBP) has limited their use in operational settings. Mo- literature has grown up around this topic. A key theme in this
tivated by the statistical character of the negativity artifact, literature is the incorporation of likelihood principles into the
we develop a simple post-processing technique that iteratively reconstruction process. The motivation for this goes back to
adjusts negative values by cancellation with positive values in
a surrounding local neighborhood. The compute time of this Rockmore and Mackovski [22], who suggested that since raw
approach is roughly equivalent to two applications of FBP. data in emission tomography are well modeled by Poisson
The approach was evaluated by numerical simulation in one- statistics [22], [25], [27], the method of maximum likeli-
and two-dimensional settings. In two dimensions, the source hood could potentially yield a more efficient reconstruction
distributions included the Hoffman, the Shepp and Vardi, and than FBP. Maximum likelihood (ML) was made feasible by
a digitized version of the Jaszczak cold spheres phantoms. Our
studies compared smoothed versions of FBP, the post-processed Shepp and Vardi [23], who proposed an iterative expectation-
FBP, and ML implemented by the expectation-maximization maximization (EM) algorithm for its implementation. Building
algorithm. The root mean square (RMS) error between the true on this, a rich collection of refinements and developments of
and estimated source distribution was used to evaluate perfor- ML reconstruction methods have appeared [20], [24], [26].
mance; in two dimensions, additional region-of-interest-based A whole range of more elaborate Bayesian reconstruction
measures of reconstruction accuracy were also employed. In
making comparisons between the different methods, the amount methods which also make essential use of likelihood have
of smoothing applied to each reconstruction method was adapted also been proposed [4], [5], [8], [9]. Unfortunately, the com-
to minimize the RMS error-this was found to be critical. putational burden of ML and related Bayesian reconstruction
After adjusting for this effect, the average RMS error for FBP methods has severely limited the ability to quantify in practical
was typically between 13% and 20% higher than ML; the engineering terms the magnitude of the gains associated with
modified FPB method was between 2% and 5% greater than ML.
Similar results were found for the region-of-interesterror. These the use of likelihood-based reconstruction methods [5], [7],
experiments suggest that it may be possible to achieve much of the [20]. This tums out to be an important issue because a
gain associated with ML reconstruction in emission tomography simple ML reconstruction typically requires at least two orders
by simpler computational means. of magnitude more computing time than FBP. Thus, in an
operational setting, it is relevant to know if the efficiency gain
I. INTRODUCTION achieved by ML is worth the additional compute time [7].
ML reconstructions are constrained to be positive and as
T HE STANDARD reconstruction methodology for emis-
sion tomography is an algorithm known as filtered back- a result show reduced reconstruction artifacts in background
projection (FBP) [16]. This algorithm or variations of it, such regions [20], [26], [27]. Motivated by this we wondered if neg-
as confidence-weighted filtered backprojection [IO], [25], is ativity artifacts in FBP could be reduced by some form of post-
used in almost all tomographs in operation today. However, processing. Here we propose a simple computational scheme
because of the increasing interest in quantitation [15], there based on an iterative local cancellation technique. This method
has been considerable interest in more-sophisticated statistical is evaluated by a set of numerical simulation studies with one-
and two-dimensional source distributions. The results indicate
Manuscript received June 20, 1991; revised January 25, 1993. This work
was supported by the National Institutes of Health under CA-42593 and CA- that it may be possible to realize much of the gain associated
42045. The associate editor responsible for coordinating the review of this with ML by simpler computational means. The paper is orga-
paper and recommending its publication was Dr. I. Sezan. nized as follows: Section I1 reviews the reconstructionproblem
F. O’Sullivan is with the Departments of Statistics and Biostatistics,
University of Washington, Seattle, WA 98195. and defines FBP, ML, and the post-processing technique for
Y. Pawitan is with the Department of Statistics, University College Dublin, FBP. Section 111 presents numerical simulation studies to
Belfield, Dublin 4, Ireland. compare the alternative reconstruction methods. We conclude
D. Haynor is with the Department of Radiology, University of Washington,
Seattle, WA 98195. in Section IV by offering some possible interpretationsas well
IEEE Log Number 9213405. as some potential limitations of this work.
0278-0062/93$03.00 0 1993 IEEE
654 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 12, NO. 4, DECEMBER 1993

11. PROBLEM FORMULATION

AND Least Squares and Filtered Backprojection: The least-
RECONSTRUCTION “ I Q U E S squares solution, z ( l S ) minimizes
, the residual sum of squares
The reconstruction problem of emission tomography can be between the data vector (y) and its expected value under the
formulated as the inversion of a linear integral operator based Poisson model (KA). This solution is given by
on Poisson-distributed data [20], [27]: the observable data .(Is) = (K’K)-lK’y. (3)
is a realization of an inhomogeneous Poisson process whose
intensity is related to the target radiotracer source distribution In the case that K corresponds to the Radon transform and
by a linear integral equation. The reconstruction problem is the discretization is fine enough, the matrix K’K is Toeplitz
to estimate the source distribution given a realization of the with the ith eigenvalue proportional to i for i = 1,2, . . . , S.
Poisson process. In both positron emission tomography (PET) This observation is the basis of an algorithm known as filtered
and single photon emission computed tomography (SPECT), backprojection (FBP) for computing approximate least-squares
the integral operator is modeled as a Radon transform, with solutions-Natterer [16] shows that the application of K’ to
potential modifications to account for attenuation, scatter, and the data is backprojection and the multiplication by (K’K)-’,
time-of-flight measurements [ 11, [2], [20], [25]. Here we focus which is equivalent to convolution with the ramp filter, is the
on a non-time-of-flight PET machine without consideration of Jiltering step. More generally, the confidence-weightedfiltered
attenuation or scatter effects, c . f . [24], [27]. In practice the backprojection method used with time-of-flight PET machines
reconstruction problem is discretized. Let there be S pixels in also implements the method of least squares. In practice,
the imaging domain and T pixels in the observation domain. FBP is the standard reconstruction algorithm used in emission
The desired source distribution is specified by a vector X = tomography; the connection to least squares is noteworthy
(As; s = 1,2, . . . , S) and the mean value of the observation even though the accuracy of this connection formally depends
process is a vector p = ( p t ;t = 1 , 2 , . . . ,T). In this discrete on the degree of discretization.
setting, A and p are related by a linear equation The main advantage of FBP is its the computational ef-
ficiency. After backprojection, filtering is carried out using
p=KKX (1) fast Fourier transform techniques. Even more efficient convo-
lution backprojection implementations, combine the filtering
where K is a T x S matrix representing the discretized linear
integral operator. The data consist of a vector of binned counts and backprojection steps [16]. By way of comparison, each
iteration of the EM algorithm used to compute the ML
yt for t = 1,2, . . . T which are modeled as independent
Poisson variates with rate parameters given by the pt’s. reconstruction, involves a projection (multiplication by K)
and backprojection (multiplication by K’) operation, so a
single EM iteration is roughly equivalent to two applications
A. Raw Reconstructions
of FBP. Since the most efficient implementations of the EM
Following Vardi et al. [27] and Geman et al. [5], we will algorithm (with appropriate initialization) typically require
present reconstruction methods in their basic form first without between 100-1OOO iterations to produce satisfactory estimates
any smoothness considerations. Subsequently we will discuss [5], [20], EM reconstructions are at least two orders of
techniques used to impose smoothing constraints by filtering magnitude slower than FBP.
these raw reconstructions. Post-Processed Filtered Backprojection: The source distri-
Maximum Likelihood: This method naturally follows from bution in emission tomography cannot be negative. Thus an
the assumed Poisson structure. The raw maximum likelihood obvious shortcoming of least squares is that it does not impose
solution maximizes the likelihood or log-likelihoodof the data. positivity constraints on the solution. By contrast, provided one
Up to a constant, the log-likelihood is given by starts with an everywhere positive guess, the EM algorithm
T guarantees that maximum likelihood solutions always remain
l(Yl4 =Z{Yt W K N -
t=l
mJ non-negative. Directly modifying least squares to impose posi-
tivity leads to a quadratic programming problem, but currently
available algorithms [6] for solving quadratic programs cannot
where Ki is the tth row of K. In the case that all the elements
take advantage of the special structure of K, which is the
of K are positive Vardi et al. [27] showed that the expectation- key to the computational efficiency of FBP. An alternative
maximization (EM) algorithm could be used to evaluate the
approach would be to apply some type of post-processing
ML solution. The basic EM iteration involves two steps
scheme to reduce negativity artifacts in FBP solutions. In
corresponding to the conditional expectation and subsequent
order to motivate such a scheme we note that the negativity
maximization of the conditional log-likelihood. When these
artifact in FBP solutions has the following structure: In the
two steps are combined the iteration (zold + Pew) is
neighborhood of a pixel with a large negative value, there
Z , ” e ~= p l d . Ytktslk., typically are a number of surroundingpixels where the solution
(2) is relatively large and positive. Theoretically this can be ex-
8 ZFldktr
plained by the fact that the ramp filter used in FBP accentuates
where kt, is the (t,s)’-th element of the matrix K and high frequencies and consequently tends to create negative
k., = Et kt,. This iteration is performed a specified number spatial autocorrelation when applied to a white noise process
of times and the ML estimate, dml), is set to the final value [3]. Our post-processing algorithm attempts to take advantage
of Pew. of this negative spatial autocorrelationby iteratively applying a
O'SULLIVAN er al.: REDUCING NEGATIVITY ARTIFACTS IN EMISSION TOMOGRAPHY 655

local cancellation rule which distributes the negative artifact at [27], [29] as well as methods that directly incorporate spatial
a pixel over the surrounding neighborhood, while preserving smoothing constraints into the EM iteration [9], [26], [24].
the total number of counts in the reconstructed image. The We will make some comments on these latter methods in the
definition of the algorithm is as follows: discussion.
Initialize:
X c 2 s ; c r i t t 00; to1 = .0001 ; 111. NUMERICALCOMPARISON OF
maxit = 100 ; iter = 0 ; w t 2 THE RECONSTRUCTION METHODS
While ( crit < to1 & iter 5 maxit ) {
iter +- iter + 1 A. One-Dimensional Experiments
Scan the image, As for s = 1,2, . . . S The reconstruction methods were first examined in a one-
if (A, < 0 ) { dimensional setting where computational simplicity allowed
a. Find the most positive value (A,+) us to study a large number of simulations. We considered a
of X in a w-neighborhood of pixel s model problem in which the spectral characteristics of K were
+
b. Adjustments: X, + min(X, A,+, 0) ; chosen to match those of the parallel-beam Radon transform
+
A+, + max(X, A,+, 0) } arising in conventional emission tomography. For a > 0 and
crit = max,:x <olX,I
T an even integer, let the vector IC(") be defined as the inverse
mean,::,,,(x,) 1 discrete Fourier transform of the periodic sequence
X jp = +
v = 1,2, . . . ,T / 2 1
End
,US+)

{ v-"
(T + 2 - v>-" v = T / 2 + 2,. . . ,T.
The neighborhood in step (a) is defined as the set of all By direct computation it can be seen that k(") is real, T -
pixels which are within a distance w of s. In the case that no periodic and positive. We let K be the T x T Toeplitz matrix
positive values are found, the neighborhood is expanded by a whose rows are periodic shifts of k("). The one-dimensional
factor of two until the condition is met. Typically, we find that
inverse problem is to estimate the source distribution X (a
the magnitude of the maximum negative artifact in the FBP vector of dimension T) given a realization from a Poisson
solution is reduced by at least two orders of magnitude by process whose mean is p = KX. By definition, K'K has
the algorithm. The computational cost appears to be roughly a spectral decomposition in terms of the Fourier transform
comparable to an additional one or two backprojection steps. with eigenvalues given by [k?'I2 for v = 1 , 2 , . . . ,T. The
Thus the computation of the post-processed FBP solution, parameter Q controls the degree of difficulty problem-a larger
which we will denote FBP-p, retains much of the efficiency cr will make the inversion of K more unstable numerically and
of FBP. statistically. The parallel beam Radon transform arising in PET
simulation can be modeled with Q = 1/2 (see Theorem 1.6 of
Natterer [ 161). Projection and backprojection operations with
B. Spatial Smoothing K can be achieved by periodic convolution with k("), which
All of the above reconstructions must be smoothed in some is readily computed by the EFT. Thus the EM computation is
manner in order to obtain acceptable solutions. There are a greatly simplified in this case.
wide variety of possible schemes including regularization, the A piece-wise constant phantom source distribution (A) was
method of sieves, and low-pass filtering. In regularization (as used in the simulations. A graph of the phantom as well as
well as more general Bayesian methods) the objective function, its projection ( p = KX) for Q = 0.75 is shown in Fig. l(a).
residual sum of squares or likelihood, is modified by adding a Most of the activity in the phantom lies in the middle third
term which measures the plausibility (e.g. smoothness) of the of the field of view, similar to operational tomographs. The
solution. This gives rise to EM, and MAP techniques [5], [9], discretization is set at T = 216. The bars in the phantom
[24]. The method of sieves achieves smoothness by restricting range in width from 13 pixels down to 4 pixels. An additional
the range of the optimization in least squares or maximum bar with diminished intensity representing a PET calibration
likelihood to a subset of smooth functions on the parameter vial is placed off to the side of the phantom. We examined
space [26], [20]. three cases: a = 0.50,0.75, and 1.00. A method described
The most commonly used approach with filtered backpro- in Ripley [21] is used to simulate pseudo-random Poisson
jection low-pass-filters the raw FBP solution by convolution deviates. The phantom was scaled so that the expected total
with an appropriate smoothing kernel. A radially symmetric counts N = C t p t per pixel ranged (linearly on a log2
Gaussian convolution kernel is typically used [lo]. The full- scale) over 9 discrete values between 5 and 200. For each
width-half-maximum of the Gaussian kernel, controls the N, a total of 50 data sets were generated and analyzed using
amount of smoothing; the larger this parameter, the more three reconstruction methods; least squares, post-processed
smooth the result. We use this simple scheme in our numerical least squares, and maximum likelihood. The post-processed
simulations. This approach to smoothing ML solutions was least-squares procedure was the same as described in Section
suggested by Vardi [28]. Some alternative techniques, which 11, except the neighborhood for local cancellation is now
have been proposed in the context of maximum likelihood one-dimensional. 3000 iterations were allowed in the EM
reconstructions,include early stopping of the EM iteration [5], algorithm, and the iteration was initialized with a uniform
656 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 12, NO. 4, DECEMBER 1993

TABLE I
IMPROVEMENT
RELATIVE IN RMS ERRORS
IN I-D SIMULATION:
MEAN(s.D.)
h-matrix a = 0.50 a = 0.75 a = 1.00
100x(LS-ML)/ML 1.3 (1.8) 1.4 (1.9) 1.7 (2.1)
lOOx(LS+-MULS+ -3.9 (1.3) -1.4 (1.6) 1.9 (2.9)

TABLE I1
RATIOS
OF SMCOTHING
PARAMETERS KERNEL):
(WIDTHOF GAUSSIAN
MEAN(s.D.)
K matrix a = 0.50 a = 0.75 a = 1.00
MULS 0.75 (0.07) 0.67 (0.07) 0.59 (0.12)
ML/LS+ 1.06 (0.14) 0.93 (0.09) 0.80 (0.15)

0 50 100 150 200

P
1000 5000 10000 50000
N
(b)
Fig. 1. I-D simulation results. (a) The source distribution ( f ; dotted line)
and g = Sf (solid line) for a = 0.75 (see Section 111-A). (b) The RMS
error characteristics as a function of count rate for the different methods.
The symbols 1, p, and m refer to least squares (LS), modified least squares
(LS-P), and maximum likelihood (ML), respectively. The differences between
the reconstruction methods are not dramatic-LS is very competitive with ML.
The RMS error rate as a function of expected counts (N) drops more rapidly
as the degree of blurring (a) is reduced.
guess. Typically the EM iteration converged in fewer than
loo0 iterations. Raw reconstructions were smoothed by convo-
lution with a Gaussian kernel whose full-width-half-maximum
was adapted to minimize the root-mean-square (RMS) error
between the reconstruction (A) and the true phantom (A).'
average over 50 simulations) as a function of the total expected
counts N . Here there do not appear to be any dramatic
differences between any of the reconstruction methods. Table I
summarizes the results by averaging across the different values
of N . Overall, the RMS error for least squares is between 1%
and 2% greater than for maximum likelihood. The differences
between post-processed least squares and maximum likelihood
are of a similar order. Surprisingly, the post-processed least
squares is even somewhat better than maximum likelihood
when (I: = 0.5; this corresponds to the case where the inversion
of the operator K is most stable.
A comparison of the optimized smoothing parameter for the
different reconstruction methods is shown in Table 11. Least
squares consistently requires a greater amount of smooth-
ing than maximum likelihood; but the amount of smoothing
required for post-processed least squares and maximum like-
lihood are very similar.
B. Two-Dimensional Experiments
The reconstruction methods were next compared in a more
realistic two-dimensional setting. Simulated data sets were
generated from a tomograph whose parameters were set in
accordance with the specifications given in Hoffman e? al.
[ 1 I]; an imaging domain with 128 x 128 pixels of dimension
2.1 mm, an observation domain with 128 x 160 distance-angle
bins of size 2.1 mm x 7r/160 rad. and a transverse dector
resolution of 4.5". The observational model [25] is
p = KA = R(g * A) (4)
where R is a discrete approximation to the line integral
radon transform and g is a spherically symmetric Gaussian
distribution whose full-width-half-maximum is 4.5 mm.
Three phantom source distributions, shown in Fig. 3, were
considered. The first phantom is the most realistic anatomi-

-4-
In Fig. 1, the RMS error is scaled by the standard deviation
of the phantom ( s.d. = with the mean x
of A). Figure 2 shows a sample reconstructions with total
expected counts N = 4320 and (I: = 0.75. There appears
to be very little difference between the modified least squares
and the maximum likelihood results. Figure l(b) shows the
behavior of the RMS error (each point on the graph is an
'While this can be done in a simulation setting, a data-dependent method
is needed for application in practice. Such a method is proposed in Pawitan
and O'Sullivan [18].
cally; it is a slice number 8 from the Hoffman et al. [ 111 brain
phantom set. The second phantom comes from Vardi et al.
[27]; it is a more abstract representation of a brain slice, but,
relative to the Hoffman phantom, has a greater range of gray-
scales. The third phantom is a digitized version of the Jaszczak
phantom which is commonly used for quality monitoring with
tomographs [ 141. Data sets were generated from each of these
phantoms using the same approach as outlined for the one-
dimensional experiments. The total expected counts N were
allowed to vary over 9 distinct values (equi-spaced on a log,
O'SULLIVAN et al.: REDUCING NEGATIVITY ARTIFACTS IN EMISSION TOMOGRAPHY 657

2. Sample Reconstructions
(a=0.75, N = 4320)

Errors
U?

LS WPl ML

Fig. 2. A sample I-D reconstruction for a = 0.75. The columns correspond to LS, LS-P, and ML (2000 EM iterations). The
amount of smoothing applied to obtain each reconstruction on the top row was adjusted to minimize the RMS error. Differences
between the true source distribution (dashed line on the top row) and the reconstructed estimate (solid line on the top row) are
shown on the second row. LS-P and ML are qualitatively very similar.

scale) between N = lo4 and N = lo6 counts. This range of
counts is comparable to what is typically seen on individual
scans in dynamic F-18 deoxyglucose brain study [19]. Re-
constructions were computed using FBP, post-processed FBP,
and ML, even though our simulation model is admittedly
quite an idealization. The maximum number of EM iterations
used in the ML reconstructions was 3000 and the initial
guess was uniform. As in the one-dimensional experiments,
raw reconstructions were smoothed by convolution with a
Gaussian kemel whose full-width-half-maximum was adapted
to minimize the RMS error between the reconstruction and the
true phantom.
Figure 4 shows a set of sample reconstructions obtained
with datasets from each phantom. The most striking source
of disagreement between the reconstructions arises in the
background region where the FBP reconstruction suffers by
not being constrained to be positive. This artifact appears to
be eliminated by post-processing. Qualitatively, there appears
to be little difference between the post-processed FBP and ML.
Figure 5 shows profiles through the reconstructions in Fig. 4.
These profiles were chosen to pass through interesting sections
of each phantom. In these graphs the differences between the
reconstruction methods are only strongly evident on back-
ground (see the Hoffman phantom), where FBP has negative
artifacts. In general, both FBP-p and ML reconstructions are
very similar-a exception is the Hoffman with N = lo6,
where the post-processed FBP reconstruction has a spurious
positive artifact on background (6-8 o'clock), which in tum
leads to some negative bias in the positive part of the phantom.
The overall reconstruction was measured using the RMS
error as well as with a more elaborate region-of-interest
(ROI) based measurement. For the ROI error, regions Rj for
j = 1 , 2 , . . . , J were defined on each phantom as follows:
white and gray matter for the Hoffman phantom (J = 2); the
six homogeneous elliptical regions as well as the compliment
of those structures over the head region for the Shepp-Vudi
phantom ( J = 7); and the six cold spheres in the Jaszczak
phantom (J = 6). Note that none of the regions of interest
considered include background. The ROI error was defined
as an average relative error over the regions in the phantom.
Formally, let A ( j ) and A ( j ) be the average value over region
Rj in the true phantom and reconstructed image respectively,
I

658 IEEE TRANSACTIONS ON MEDICAL IMAGING. VOL. 12, NO. 4, DECEMBER 1993

(a) Hoffman (b) Shepp-Vardi (c) Jaszczak

F
F
h q F F
U F

I M I M I

10-4 5'10"4 lO"5 5'10"5 10% 10"4 5'10"4 10%

1.21 F
F F F F
F F
F b 6 " F F
M
I'
M
F
F M F
t F
iA
M F 04 F F
F M
M F n
E I F F F
M hl M Y
F

Dl
M
F F M :,
M
M M M

0.J . M 0.4
1014 5'10'4 10*5 5'10% 10%
1014 SlO"4 10% 5'1015 10"6 1014 5*10"4 10% 5'1W5 1W6

Fig. 3. 2-D simulation results. The top panel shows the three source distributions: (a) Hoffman brain phantom, (b) Shepp-Vardi
phantom, and (c) digitized Jaszczak phantom. The middle panel shows the RMS error as a function of expected counts for the three
reconstruction methods; the symbols F, P, and M refer to filtered backprojection (FBP), modified filtered backprojection (FBP-p),
and maximum likelihood (ML), respectively. The bottom panel shows similar results for ROI error rates (see Section III-B).

then we use error for the post-processed FBP is 42% greater than the
ML error. We believe that the most likely explanation for
this is that at high counts the ML reconstructions, which
use K directly, are beginning to tackle some of the effect
of detector blurring (the Gaussian convolution term in (4);
The scaling of the difference between A ( j ) and i ( j )is to ensure the Hoffman phantom has many involutions so the difference
that each region receives equal weight; the factor .1/S is between the blurred phantom (g * A) and the true is more
important for the Jaszczak because the average intensity of pronounced than in the other two phantoms. On the other
each of the regions of interest is zero in that case. hand, the FBP reconstructions only account for the Radon
The behavior of the RMS and ROI errors as a function transform component of K, i.e. R in (4).Table I11 summarizes
of total expected counts is shown in Fig. 3(a) and (b). relative increases in the RMS and ROI errors averaged across
The RMS errors are more stable than the ROI error with the different values of N . For the RMS error, the FBP is 13%
the Jaszczak phantom ROI errors being particularly variable. to 20% greater than ML while the post-processed FBP is 2%
The differences between FBP and ML reconstructions are and 5% greater than ML. The corresponding relative increases
consistently greater for the RMS error. The reason is because for the ROI error range from 13% to 26% for FBP and from
this error includes background while the ROI error does 3% to 15% for the post-processed FBP reconstruction.
not. For both error assessments, the post-processed FBP is
seen to improve on FBP; indeed, for the Shepp-Vardi and IV. DISCUSSION
Jaszczak phantoms, post-processed FBP is very close to ML The tomograph simulation model used in Section 111 has
over the entire range of counts. The greatest discrepancy various idealizations which are obviously not realized in
between the post-processedFBP and ML occurs at high counts practice. For example, the model does not include the effects
with the Hoffman phantom; here, in the extreme, the ROI of scatter or random coincidences [l], [20]. In view of this,
OSULLIVAN er al.: REDUCING NEGATIVITY ARTIFACTS IN EMISSION TOMOGRAPHY 659

4(a) Hoffman Braln Phantom

(N=10"6)

(N=10"5)

4(b) Shew-Vardi Brain Phantom

(N-IO"6)

(N=10"5)

Fig. 4. Sample 2-D reconstructions with expected counts N = lo6 and N = lo5. There are 3 sets of plots, (a), (b), and (c),
with six images on each: (a) Hoffman brain phantom, (b) Shepp-Vardi phantom, and (c) digitized Jaszczak phantom. Note that the
color scales on each image range from the maximum to the minimum value for that image. Maximum likelihood (ML) and modified
filtered backprojection (FBP-p) achieve a dramatic reduction in the background artifact seen in filtered backprojection (FBP).

we would not expect the absolute reconstruction accuracy performance of the different reconstruction methods would
found by the various methods to be achieved on an operational change. It would be valuable to develop a calibration between
tomograph. Nevertheless, we see no reason why the relative the counts used in the simulation model and physical counts
 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 12, NO. 4, DECEMBER 1993

4(c) Digibzed Jaszczak Phantom

(N=1O"6)

(N=i0"5)

Fig. 4 (continued).

TABLE III the examples we see much lower rates of estimation because
2-DIhENSIONAL SIMULATIONS
RECONSTRUCTION RESULTS IN
the reconstruction problem is inherently a non-parametric
Phantom Hoffman Shepp-Vardi Jaszczak function estimation problem [ 121. In non-parametric function
Relative Improvement in RMS Error: mean (s.d.) estimation, both bias and variance contribute to the error which
IOOX(FBP-ML)IML 13.0 (4.4) 20.4 (3.4) 20.0 (4.0)
leads to modified rates of estimation. Some further aspects of
lOOX(FBP-p-ML)IML 4.9 (4.2) 3.1 (1.4) 2.3 (1.0)
these rates of estimation will be develop in a forthcoming
Relative Improvement in ROI Error: mean (s.d.) report [ 171.
IOOX(FBP-ML)/ML 25.5 (12.6) 15.0 (4.3) 12.5 (4.6) An important confounding factor in comparisons of re-
lOOx(FBP-p-ML)/ML 14.7 (13.0) 3.0 (2.4) 4.8 (1.8) construction methodologies is the choice of the smoothing
parameter. A single value for a smoothing parameter will
in an actual tomograph. In terms of reconstruction quality, not be appropriate for both maximum likelihood and filtered
a physical count of N = lo6 events probably corresponds to backprojection reconstructions. In the work reported here, we
many fewer counts in the simulation model. Perhaps motivated overcome this problem by adaptively choosing the smoothing
by this, recent work by Politte and Snyder [20], which used a parameter that is best in a mean-square error-sense for each
simulation model similar to ours, focused on experiments with method. This was easy for us because we knew the true
N = lo5 counts. With this in mind, we wonder if the gains image. In practice, it would be desirable to have adaptive
seen by maximum likelihood at high counts on the Hoffman data-dependent rules for selecting good smoothing parameters.
phantom could be realized in a practical setting. In the future We have begun to investigate the use of unbiased risk-
we plan to conduct some physical phantom experiments to estimation techniques for this purpose [ 181. EM algorithms that
explore the reconstruction methods in a fully realistic setting. directly incorporate some form of smoothing [20], [24] into the
The behavior of the reconstruction errors as a function of iteration have the potential to dramatically reduce the compute
counts is highly structured. Figures 1 and 3 show that the RMS times associated with ML reconstructions. The magnitude of
error is remarkably well modeled by an equation of the form this reduction is not well documented in the literature-method
of sieves algorithms are still reported to use between 100
R M S ( N ) C . N-' (5) and 500 iterations [20]-but it is clear that any reduction in
where the constants C and T depend on the phantom and on K. compute time depends on the amount of smoothing used [24].
In statistical terms T is known as the rate of estimation of the Since in practice reconstructions are computed for a range of
reconstruction method. The best possible rate of estimation is amounts of smoothness before selecting the most reasonable
0.5, but this is typically only achievable in parametric estima- result, any estimate on the compute-time reductions associated
tion problems where error is dominated by variance [ 131. In all with these methods would need to factor in this effect.
OSULLIVAN er al.: REDUCING NEGATIVITY ARTIFACTS IN EMISSION TOMOGRAPHY 661

5(a) Hoffman Phantom

L L L L-

N=10"6

5(b) Shepp-Vardi Phantom

N=10"6

FBP FBP P

(b)

Fig. 5 . Line-plot quantitation of sample 2-D reconstructions shown in Fig. 4. As in Fig. 4 there are 3 sets of plots, (a), (b), and (c),
that correspond to the different phantoms. In each case the reconstructions are examined along a line through the phantom (see the
top panel). The results are shown for N = lo6 and N = lo5 on the middle and bottom panels. In these graphs the reconstructions
(solid line) and source distributions (dashed line) are normalized so that the maximum value for the source distribution is unity-this
allows direct comparisons between the N = l o 6 and N = l o 5 results.

We have proposed a simple ad hoc approach to reducing be a variety of other schemes that would behave just as well or
negativity artifacts in filtered backprojection reconstructions. better, e.g. the negative artifact could be distributed in some
While the technique seems to perform quite well, there may altemative manner in the local neighborhood or the image
662
5(c) Jaszczak Phantom
N=10”6
N=10“5
FBP
Fig. 5 (continued)
could be scanned in some alternative manner. In any event,
most of the benefit of maximum likelihood reconstruction for
emission tomography appears to derive from its imposition
of positivity constraints, and the method of post-processing
filtered backprojection to reduce negativity artifacts appears
to be quite promising.
V. CONCLUSION
We have shown through numerical simulation that maxi-
mum likelihood achieves better reconstruction characteristics
than standard filtered backprojection.A local cancellation post-
processing algorithm applied to the filtered backprojection
solution substantially reduces the negativity artifact and re-
covers much of the gain associated with maximum likelihood.
This is of practical importance because the computation of
maximum likelihood by the EM algorithm requires orders-of-
magnitude-more compute time than the post-processed filtered
backprojection reconstruction.
REFERENCES
B. Bendriem, R. Soussaline, R. Campagnolo, B. Verrey, P. Wajnberg,
and A. Syrota, “A technique for correction of scattered radiation is a
PET system using TOF information,” J. Compur. Assisr. Tomogr., vol.
10, pp. 287-295, 1986.
M. Bergstrom, L. Eriksson, C. Bohm, G. Blomqvist, and J. Litton,
“Correction for scattered radiation in a ring dector positron camera by
integral transformation of the projections,” J. Compur. Tomogr., vol. 7,
pp. 42-50, 1983.
D. Brillinger, Time Series: Dura Analysis and Theory. San Francisco:
Holden Day, 1981.
S. Geman and D. E. McClure, “Bayesian image analysis: Application
to single photon emission tomography,” in Proc. Statistical Computing
Section, American Statistical Association, pp. 12-18, 1985.
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 12, NO. 4, DECEMBER 1993
iBP Y
(c )
S. Geman, K. M. Manbeck, and D. E. McClure, “Design and imple-
mentation of a statistical method for emission computed tomography,”
Reports in Partem Analysis, No. 154, Div. of Applied Mathematics,
Brown Univ., 1991.
P. E. Gill, W. Murray, and M. H. Wright, Pracrical Optimizarion. New
York: Academic Press, 1981.
D. R. Gilland, B. M. Tsui, C. E. Metz, R. J. Jaszczak, and J. R.
Perry, “An evaluation of maximum likelihood-expectation maximization
reconstruction for SPECT by ROC analysis,” J. Nucl. Med., vol. 33, pp.
451-457, 1992.
P. J. Green, “Bayesian reconstruction from emission tomography data
using a modified EM algorithm,” IEEE Trans. Medical Imaging, vol. 9,
pp. 84-93, 1990.
E. Levitan and G. T. Herman, “A maximum a posteriori probability ex-
pectation maximization algorithm for image reconstruction in emission
tomography,” IEEE Trans. Medical Imaging, vol. 6, pp. 185-192, 1987.
B. D. Hoffman, D. C. Ficke, T. J. Holmes, D. Politte, and M. M.
Ter-Pogossian, “Image reconstruction of data from SUPER PE= 1:
A first generation time-of-flight positron emission tomograph,” IEEE
Trans. Nucl. Sci., vol. 33, pp. 428434, 1986.
E. J. Hoffman, P. D. Cutler, W. M. Digby, and J. C. Mazziotta, “3-
D phantom to simulate cerebral blood flow and metabolic images for
PET,” IEEE Trans. Nucl. Sci., vol. 37, pp. 61-20, 1990.
I. Johnstone and B. W. Silver”, “Speed of Estimation in Positron
Emission Tomography,” Ann. Sratisr., pp. 616-620, 1990.
E. L. Lehmann, Theorey of poinr esrimurion. New York: Wiley, 1983.
T. K. Lewellen, A. B. Bice, R. L. Harrison, M. D. Pencke, and J.
M. Link, “Performance measurements of the SP3OOOILJWtime-of-flight
positron emission tomograph,” IEEE Trans. Nucl. Sci., vol. 35, pp.
665-669, 1988.
J. C. Mazziotta, M. E. Phelps, and H. R. Schelbert, Posirron Emission
Tomography and Aurorudiography. New York: Raven, 1986.
F. Natterer, Marhemarics of Compurerized Tomography. New York:
Wiley, 1986.
F. O’Sullivan, “A study of least squares and maximum likelihood for
image reconstruction in positron emission tomography,” (under review,
Ann. Statist, 1993).
Y. Pawitan and F. O’Sullivan, Data-dependent bandwidth selection
for emission computed tomography reconstruction, IEEE Trans. Med.
Imaging, vol. 12, pp. 167-172, June 1993.
M. E. Phelps, S . C. Huang, E. J. Hoffman, C. Selin, L. Sokoloff, and D.
E. Kuhl, “Tomographic measurement of local cerebral glucose metabolic
OSULLIVAN et al.: REDUCING NEGATIVITY ARTIFACTS IN EMISSION TOMOGRAPHY
rate in humans with [F-18]2-Fluoro-2-deoxy-D-glucose: Validation of
method,” Ann. Neurol., vol. 6, pp. 371-388, 1979.
[20] D.Politte and D.Snyder, “Corrections for accidental coincidences and
attenuation in ML image reconstruction,” IEEE Trans. Medical Imaging,
vol. 10, p. 82, 1991.
[21] B. Ripley, Srochastic Simulation. New York: Wiley, 1987.
[22] A. Rockmore and A. Makovski, “A maximum likelihood approach to
emission image reconstruction from projections,” IEEE Trans. Nuclear
Science, vol. 23, pp. 1428-1432, 1976.
[23] L. A. Shepp and Y. Vardi, “Maximum likelihood reconstruction in
positron emission tomography,” IEEE Trans. Medical Imaging, pp. 113-
122, 1982.
[24] B. W. Silverman, M. C. Jones, J. D. Wilson, and D. Nychka, “A
smoothed em approach to indirect estimation problems, with particular
reference to stereology and emission tomography,” J. Royal Statist.
SOC.(B), vol. 52, pp. 271-324, 1990.
663
[25] D.L. Snyder, L. J. Thomas, and M. M. Ter-Pogossian, “A mathematical
model for positron emission tomography system having time of flight
measurement,” IEEE Trans. Nucl. Sci., vol. 28, pp. 3575-3581, 1981.
[26] D. L. Snyder and M.I. Miller, “The use of sieves to stabilize images
produced with the EM algorithm for emmission tomography,” IEEE
Trans. Nucl. Sci., vol. 32, pp. 3864-3872, 1985.
[27] Y. Vardi, L. A. Shepp, and L. A. Kaufman, “Statistical model for
positron emission tomography,” J. Amer. Statis. Assoc., vol. 80, pp.
8-37, 1985.
[28] Y. Vardi, “Comments on ‘A smothed EM approach to indirect esti-
mation problems, with particular reference to stereology and emission
tomography [Silverman e?al.]”’ , J . Royal Statist. Soc. (B), vol. 52, pp.
271-324, 1990.
[29] E. Veklerov and J. Llacer, “The feasibility of images reconstructed with
the method of sieves,” IEEE Trans. Nucl. Sci., vol. 37, pp. 835-841,
1990.