ORIGINAL ARTICLE

Annals of Nuclear Medicine Vol. 19, No. 2, 75-82, 2005

Characterization of ordered-subsets expectation maximization

with 3D post-reconstruction Gauss filtering and comparison
with filtered backprojection in 99mTcSPECT

Marco BRAMBILLA,*Barbara CANNILLO,*Marco DOMINIETrO,*

Lucia LEVA,** Chiara SECCO* and Eugenio INGLZSE**

*Health Physics Department, **Nuclear Medicine Department,

Azienda Ospedaliera "Maggiore della Carith," Novara, Italy

Purpose: To characterize ordered-subset expectation maximization algorithm with a fixed 3D

Gauss post-reconstruction filtering (OSEM) in 99mTc SPECT as for noise, contrast and spatial
resolution with varying number of subset and iteration and to compare OSEM with an optimized
set of parameters, with filtered backprojection (FBP) with filter parameters typical of brain and
myocardial SPECT, both with and without Chang's method of attenuation correction (AC).
Methods: SPECT images of a Jaszczak phantom with cold rod inserts, hot and cold spheres and
capillary line sources were acquired. Different background activity concentrations of the phantom
were simulated as well as different lesion-to-background activity ratios. OSEM reconstructions
were halted after 5, 10 and 15 iterations using 4, 8 and 16 subsets. Results: The effect of subset and
iteration number over noise is additive: thus, it is possible to define an EM-equivalent iteration
number that indicates the product between the subset and the iteration numbers. Noise increases
linearly with increasing EM-equivalent iteration number. For each level of nominal contrast, the
measured contrast after OSEM shows a little increase with increasing iteration number and saturates
after 80 EM-equivalent iterations. The application of AC leads to diminished contrast values both
in FBP and OSEM. The contrast of cold lesions after OSEM increases with increasing number of
EM-equivalent iteration number: after 80 iterations the contrast values with OSEM overtake the
ones obtained with FBP; contrast values diminished as background concentration raised. Resolu-
tion values did not change with increasing EM-equivalent iteration number and were higher than
those obtained with FBP. Conclusion: The major findings of the present work are the demonstration
of additivity of subset and iteration in OSEM over noise, with the possibility of defining an EM
equivalent iteration number, and the superiority of OSEM with respect to FBP in terms of spatial
resolution.

Key words: iterative reconstruction, ordered-subsets expectation maximization, filtered

backprojection, SPECT

~TRODUCTION practical alternative to filtered backprojection (FBP). Maxi-

mum-likelihood expectation maximization (MLEM) al-
ITERATIVE RECONSTRUCTIONALGORITHMSare finding gorithm has been widely investigated, so that its noise and
widespread use in commercial software because of in- convergence properties are well known. 1-4 In particular
crease in computer power and are becoming a clinically MLEM has improved noise properties over FBP and
eliminates streaking artifacts, which are common in
Received July 12, 2004, revision accepted October 25, 2004. noisy FBP studies. Unfortunately, convergence requires
For reprint contact: Marco Brambilla, M.D., Servizio di too many iterations to be practical for use in clinical rou-
Fisica Sanitaria, A.O. "Maggiore della Carit~", C.so Mazzini tine. 3
18-28 I00 Novara, ITALY. Ordered-subset expectation maximization (OSEM)
E-mail: [email protected] algorithm, 5 which is related to but much faster than

Vol. 19, No. 2, 2005 Original Article 75

MLEM, has become the dominant iterative reconstruc-
tion procedure in SPECT. Both MLEM and OSEM have
the ability to model Poisson noise and the exact photon
transport. Even scatter and distance dependent camera
response can be modeled in detail resulting in truly
quantitative images and in better noise contrast trade-
off. 6-1~Nevertheless, as pointed out by Lalush et al. II to
date the only way to prove that these algorithms are accu-
rate and effective is through careful empiric studies.
The performances of OSEM and FBP have been
compared in a number of experimental and clinical
studies 12-16 with a variety of reconstruction parameters
employed with OSEM, including number of subset,
number of iteration and the type and amount of post-
Fig. 1 Jaszczak phantom with cold rod inserts, hot and cold
reconstruction smoothing to reduce noise with increasing
spheres.
number of iterations: Blocklet et al. 12 compared FBP and
OSEM (8 subsets and 2 iterations without post filtering) in
bone SPECT reconstruction; Case et al. 13 compared cylinder with an assortment of inserts. The inserts in-
attenuation correction techniques in bone SPECT of cluded six solid and six void spheres (diameters: 9.5, 12.7,
the spine using OSEM (12 subsets and 3 iterations, 3D 15.9, 19.1, 25.4, 31.8 mm), a cold rod insert (diameters:
Wiener filter); Kauppinen et al.14 compared FBP and 4.8, 6.4, 7.9, 9.5, 11.1, 12.7 mm) and capillary line sources
OSEM (6 subsets and 4-12 iterations, Butterworth post (inner diameter: < 1 mm). In order to simulate different
filtering) in brain perfusion SPECT with scatter and non background activity concentrations the phantom was filled
uniform attenuation correction; Vanhove et al. 15 imple- with uniform 99mTcsolution of 37 kBq/ml and 12.3 kBq/
mented OSEM algorithm for data acquired with a pinhole ml, hereafter referred as high and low background. To
collimator in phantom studies: using OSEM different introduce some variability, 5 hot lesions were created
iterations were used in combination with 1-32 subsets. (filling the void spheres) with lesion-to-background ac-
Wells et al. 16compared FBP and OSEM (8 subsets and 1 tivity ratios of 2:1, 5:1,8:1 and 3:1,30:1 in the case of high
iteration, 3D Gauss) for small-lesion detection and local- and low background activity concentration, respectively.
ization in 67Ga SPECT. Hereafter, they will be referred to as nominal contrast
The choice of which parameters to employ in OSEM C,om, where Cnom = 29 [(30 - 1)/1] and 2 [(3 - 1)/1]
algorithm during image reconstruction is not entirely represent the low background activity, while Cnom= 7, 4
obvious. The problem is intrinsically multifactorial: OSEM and 1 (following the same notation) represent the high
results depend not only on subset and iteration number background activity. It must be acknowledged that there
but also on pixel size and amount of post reconstruction are relative uncertainties in these assays, which we be-
filtering. According to OSEM theory, in noiseless situa- lieve are on the order of +5%. The hot sphere contrast
tion the effect of subset and iteration number should be ratio depends on the relative accuracy of these assays and
additive over noise. It has to be demonstrated that the thus will exhibit corresponding uncertainties on the order
same holds true in noisy situations. In the present work of_+10%.
OSEM performances were characterized with respect to
subset and iteration number, having fixed the pixei size SPECT imaging
and the amount of post reconstruction filtering. OSEM Image acquisition was performed with a Siemens E-Cam
performances were expressed in terms of noise, resolution dual-head scintillation camera equipped with a low en-
and contrast in simulated hot and cold lesions. ergy high resolution collimator (Siemens Medical Sys-
The OSEM algorithm in 99mTcSPECT with an opti- tems, Inc., Hoffman Estates, I!!., USA). SPECT acquisi-
mized set of parameters, was compared with filtered tions were performed using a 180~ non circular orbit for
backprojection (FBP) with filter parameters typical of each detector, with 128 projection angles, 128 • 128
brain and myocardial SPECT studies, both with and matrix size, zoom = 1 and pixel size of 4.75 mm. A
without Chang's method of attenuation correction (AC). symmetrical 15% wide energy window for the acquisition
was centered at 140 keV. The phantom studies were
MATERIALS AND METHODS acquired for 30 sec per projection angle. A total of 8
tomographic acquisitions were performed: 5 with hot
Phantom study sphere inserts with different lesion-to-background activ-
The phantom used was a Jaszczak phantom (Deluxe ECT ity ratios and background activity; 2 with cold sphere
Data Spectrum Corporation, Chapel Hill, USA), depicted inserts and different background activity; 1 with capillary
in Figure 1. The phantom was a standardized diameter line sources imbedded in water to assess tomographic

76 Marco Brambilla, Barbara Cannillo, Marco Dominietto, et al Annals of Nuclear Medicine

resolution with scatter. Acquisitions were taken within center of the slice. Noise was assessed as percentage root
two hours of phantom preparation. mean square % rms = 100 • or~M,being M and crthe mean
and the standard deviation of pixel counts in the region of
Image processing interest.
The scintillation camera was connected to a Siemens E- Contrast: For each lesion in a single reconstructed
Soft acquisition and processing computer (Siemens Medi- slice, the signal S was defined as the average pixel value
cal Systems, Inc., Hoffman Estates, I11., USA). The re- in a region of interest, centered over the lesion and with a
construction algorithms considered in this study were diameter of about half the lesion diameter. For the 10 mm
FBP and OSEM. The images reconstructed with FBP diameter lesion, a ROI of two pixels covering the entire
were pre filtered with 2 dimensional Butterworth filter 17 area of the lesion was employed. The background B was
of order 5 and a cutoff of 0.7 and 0.5 Nyquist (0.74 and defined as the average pixel value in a circular ROI
0.53 cycles/cm) in the case of high and low count density (diameter = 50 mm) positioned in the uniform region of
of images. The first choice is typical of brain studies 14 the phantom.
while the second is typical of that used in clinical 99mTc- Contrast C was defined as:
sestamibi myocardial SPECT studies.18-2~ The standard
JB - SI
first-order Chang attenuation correction 21 was applied to C-
B
reconstructed slices. Briefly, an attenuation map, based on
patient (or phantom) boundary determination and an Quantitative assessment of contrast was obtained from
approximate constant linear attenuation coefficient of the slices having at least 2 • 106 counts statistics. Measured
medium/1, is generated and applied to the reconstruction contrast values were plotted as the percentage of the
transverse slice. This method will only work well when/1 nominal contrast value versus EM-equivalent iteration
is, in fact, approximately constant (e.g. brain or abdomen). number and versus lesion diameter, with and without AC.
The linear attenuation correction was set to/.t = 0.11 cm-1. Spatial resolution with scatter: One transverse slice,
Iterative reconstruction program OSEM (Siemens 2 pixels in thickness, was reconstructed through the center
Medical Systems, Inc., Hoffman Estates, I11., USA) in- of the phantom using reconstruction techniques detailed
stalled in the E-Soft system was used for iterative recon- in Table 1. Two additional transverse slices of the same
struction. The images reconstructed with OSEM were thickness were also reconstructed and centered about +20
filtered with a symmetric 3-D Gaussian function 22 having mm from the center along the axis of rotation. The three
a full width at half maximum of 1 pixel (4.75 mm). Iter- reconstructed points in each of the three reconstructed
ative reconstruction of these images was halted after 5, 10 slices were analyzed individually with a square region of
and 15 iterations using 4, 8 and 16 subsets. Each combi- interest. Each region of interest was centered on the
nation of subsets and iterations was applied to the same set maximum count pixel. The size of the square region of
of noisy projection data. interest was at least four times the anticipated FWHM of
the count profile to be analyzed. For each point in the
Data analysis images, the FWHM in X and Y was determined. The
Data were analyzed according to the guidelines for the average FWHM radial value of the six radial measure-
quality controls of SPECT systems, elaborated by the ments on the three slice images of the two peripheral
Italian Association of Nuclear Medicine, 23 as follows: sources was calculated. Likewise, the average tangential
Noise: One transverse slice was reconstructed through FWHM value of the six tangential measurements on the
the center of the uniform portion of the phantom using images of the two peripheral sources was calculated. Also
reconstruction techniques detailed in Table 1. Two addi- the average of the six measurements (three in the X
tional transverse slices were also reconstructed and cen- direction and three in the Y direction) for the three images
tered about + 20 mm from the center along the axis of of the center source was calculated. Measured resolution
rotation. Quantitative noise assessment was obtained from values were plotted as FWHM (mm) values versus itera-
slices having at least 2 • 106 counts statistics. A circular tion number.
region of interest of 50 mm diameter was positioned in the

Table 1 ANOVA table for the factorial model referring to the influence of subset and iteration number on % rms noise.
Background activity concentration = 37 kBq/ml
Source of variation Degrees of freedom Sum of squares Mean square Variance ratio (F) P value
Subset (A) 2 187.53 93.76 37.46 ,~ 0.0001
Iteration (B) 2 108.42 54.21 21.67 < 0.0001
AB 4 8.46 2.11 0.85 0.515
Error 18 45.06 2.50

Vol. 19, No. 2, 2005 Original Article 77

 Table 2 ANOVA table for the factorial model referring to the influence of subset and iteration number on % rms noise.
Background activity concentration = 12.3 kBq/ml
Source of variation Degrees of freedom Sum of squares Mean square Variance ratio (F) P value
Subset (A) 2 108.88 54.44 82.00 < 0.0001
Iteration (B) 2 57.23 28.62 43.11 < 0.0001
AB 4 10.21 2.55 3.96 0.043
Error 9 5.98 0.64

20 i i i i

15

"8
Z 10

Fig. 2 Noise (% rms) with varying number of subset and r 2 = 0.982

iteration in OSEM. S = subset number; IT = iteration number. I I , I I
50 100 150 200 250

EM-equivalent iterations
Statistical analysis It
Continuous variables were expressed as mean and stan-
20 I I i I
dard deviation.
Noise analysis in iterative reconstruction with varying
numbers of subsets and iterations was performed by 15

means of a two factor factorial A N O V A , non repeated

measures, balanced model. 24 The factorial model allows
the simultaneous assessment of the influence of factors
over the dependent variable, also keeping into account the
interaction between factors. The two factors were subsets
(three levels of 4, 8, 16 subsets) and iterations (three levels z
O 5

of 5, 10, 15 iterations). The dependent variable was noise.

r z = 0.986
There were two and three identical experimental phantom
0 I I I I
data with low and high background concentration, respec- 50 i00 150 200 250
tively. EM --equivalent iterations
The influence of different reconstruction strategies
b
(OSEM, OSEM with AC, FBP, FBP with AC) over "hot"
lesion contrast was assessed by single factor repeated Fig. 3 a: Linear regression with 95% confidence intervals
between % rms noise and EM-equivalent iteration number, in a
measures A N O V A for each level of lesion-to-background
low background activity concentration, b: Linear regression
activity ratios. The statistical significance of differences
with 95% confidence intervals between % rms noise and EM-
in contrast between different methods of reconstruction equivalent iteration number, in a high background activity
was assessed by Fisher's protected least significant differ- concentration.
ence (PLSD) method. Statistical significance was set at
p=0.01.
Table 1 and 2 are reported A N O V A tables for the two
RESULTS factor factorial model referring to the influence of subset
and iteration number on % rms noise, with low and high
Noise increases with increasing number of subsets and background.
iteration, as shown in Figure 2. This increase is statisti- Since no interaction is present, the EM-equivalent
cally significant (p < 104). No significant interaction was iteration number may be considered as the independent
evidenced between subsets and iterations (p = 0.04 and p variable in a linear regression study, being the % rms noise
= 0.53 for low and high background). That is, the effect of the dependent variable. The regression lines together with
subset and iteration number over noise is additive. In r z correlation coefficients are shown in Figures 3a and 3b

78 Marco Brambilla, Barbara Cannillo, Marco Dominietto, et al Annals of Nuclear Medicine

Fig. 4 The measured contrast in a hot lesion of 19.1 mm
diameter, for different Cnorn, versus EM-equivalent iteration Fig. 5 Measured contrast in hot lesion with C,,om= 7 versus
number. FBP values correspond to iteration zero. lesion diameters, with four different reconstruction strategies
(OSEM: 8 subset; 10 iteration; FBP: Butterworth order 5; Fc =
0.7 Nyquist).

Table 3 Post-hoc comparison among reconstruction methods means of measured contrast values for different Cnom
Comparison Cnom= 29 Cnom= 7 Cnom= 4 Cnom= 2 Cnom = 1
FBP vs. FBP-AC 1.9 14.9" 4.6 6.2* 13.0"
FBP vs. OSEM -6.3* -5.3 -8.0* -9* -6.0*
FBP vs. OSEM-AC -6.1 * -0.7 -2.1 -0.3 8.3*
FBP-AC vs. OSEM -8.2* -20.2* -12.7" -15.2" -19"
FBP-AC vs. OSEM-AC -7.9* -15.6* -6.6* -6.5* --4.7
OSEM vs. OSEM-AC 0.3 4.6 5.9* 8.7* 14.3"
Fisher PLSD 3.3 7.5 4.8 3.0 6.0
*Statistically significant differences

for low and high background, respectively. The slope of on Chore, as expected: lesions with a diameter of 9.5 mm
the two regression lines is nearly identical (0.047 vs. or less were detectable with Cnom> 8; only lesions with a
0.050), while the noise at zero iterations is different (7.5% diameter > 12.7 mm were detectable with 5 _<Cnom< 8;
vs. 4.7%) for low and high background. The correlation is only lesions with a diameter _> 19.1 mm were detectable
almost perfect with high r 2 values. with 2 < Cnom< 5.
In Figure 4 measured contrast in a "hot" lesion of 19.1 Repeated measures ANOVA analysis demonstrated
mm diameter, expressed as a percentage of Chore, is the presence of significant differences (p < l0 -4) among
reported for different levels of Cnomand with respect to the reconstruction methods with respect to measured contrast
EM-equivalent iteration number. Corresponding to a zero values, for each level of Chore. Post-hoc comparisons
abscissa is reported the contrast measured with FBP. For between reconstruction method are provided in Table 3,
each level of Chore, the measured contrast after OSEM for each level of Cnom.For each treatment comparison the
shows a little increase with increasing number of itera- following information is provided: difference between
tions. Contrast enhancement stops after 80 EM-equiva- the contrast means and Fisher's PLSD test. If the Fisher's
lent iteration, approximately. A further increase in itera- test is significant an asterisk appears by the comparison
tions brings no additional gain in contrast. At this level of value. Table 3 indicates that the application of Chang's
EM-equivalent iteration number, the maximum gain in attenuation correction leads to diminished contrast values
contrast (16.4%) between OSEM and FBP is obtained for both in FBP and OSEM. The smaller values of contrast are
Cnom= 29. There is no clear dependence of contrast from obtained with FBP-AC, intermediate values are obtained
Cnom. with FBP and OSEM-AC; highest values are obtained
In Figure 5 measured contrast for each lesion diameter with OSEM. It is worth noting that for some level of Cnom,
(with Cnom= 7) is reported with four reconstruction statistical significance might not be attained due to the
strategies: FBP, FBP with AC, OSEM and OSEM with small statistical power of the tests, due in turn to the small
AC. The numbers of subsets and iterations in OSEM were sample size.
fixed at 8 and 10, respectively (in the region of contrast In Figure 6 the measured contrast in a "cold" lesion of
saturation). A similar trend was observed when the re- 19.1 mm diameter, expressed as a percentage of nominal
maining Cnomwere examined. contrast, is reported for two different levels of back-
The minimum lesion diameter visually detectable was ground activity concentration and with respect to the EM-
independent of the reconstruction method but dependent equivalent iteration number. Corresponding to a zero

Vol. 19, No. 2, 2005 Original Article 79

Fig. 6 Measured contrast in a cold lesion of 19.1 mm diameter
versus EM-equivalent iteration number. FBP values correspond
to iteration zero.

Fig. 7 Measured contrast in cold lesions with a high background Fig. 8 a: Spatial resolution with scatter versus EM-equivalent
activity concentration versus lesion diameters, with four differ- iteration number, b: Spatial resolution with scatter in OSEM and
ent reconstruction strategies (OSEM: 8 subset; 10 iteration; FBP with ramp filter and Butterworth (order 5, cutoff 0.7
FBP: Butterworth order 5; Fc = 0.7 Nyquist). Nyquist) filter.

abscissa is reported the contrast measured with FBP. For with respect to subset and iteration number, having fixed
each level of background, the measured contrast after the pixel size and the amount of post reconstruction
OSEM shows an increase with increasing EM-equivalent filtering. The pixel size was set according to the one used
iteration number. Approximately after 80 iterations the in our laboratory for brain studies. The type of post recon-
contrast values with OSEM surpass the corresponding struction filtering (Gaussian 3D) is the only available
values obtained with FBP. Contrast values diminish as choice in the software employed. The FWHM of I pixel
background concentration increases. was chosen on the basis of qualitative judgments per-
In Figure 7 measured contrast for high background con- formed over clinical images reconstructed both with OSEM
centration was reported with four reconstruction strate- and FBP, as the one that provided images of comparable
gies: FBP, FBP with AC, OSEM and OSEM with AC. The texture and quality with respect to conventional FBP.
numbers of subsets and iterations in OSEM were fixed at In studies reconstructed using EM, in the absence of
8 and 10. These figures indicate that contrast values regularization, as iteration number increases, noise in-
overlap with FBP and OSEM and the application of AC creases. 25 The same is true for studies reconstructed with
leads to a slight increase in contrast both with FBP and OSEM and a fixed amount of 3D post-reconstruction
OSEM. A similar trend was observed for low background Gauss filtering. According to OSEM theory, in the noise-
concentration. less situation the effect of subset and iteration number
In Figure 8a the tomographic spatial resolution with should be additive over noise (e.g. 8 subsets and 10 itera-
scatter, expressed as axial, radial and tangential FWHM tions produce the same amount of noise as 5 subset and
in mm, is reported with respect to the EM-equivalent it- 16 iterations). The present study demonstrated that the
eration number. Resolution values did not change with same is true also with noise and different background
increasing number of subsets and iterations. In Figure 8b levels.
the values of spatial resolution obtained with OSEM are The linear relationship between noise and subset times
compared with the values obtained with FBP, with a ramp iteration number leads to a predictable (at least in the
filter and with a high resolution filter used in clinical range explored) and accurate characterization of this
practice. increase: % rms increases at a rate of 1% every 20 itera-
tions approximately, both with low and high activity back-
DISCUSSION ground. Noise equivalence between FBP and OSEM is
reached at 23 and 55 iterations with low and high back-
In this work OSEM performances were characterized ground. Above these thresholds noise is higher in OSEM

80 Marco Brambilla, Barbara Cannillo, Marco Dominietto, et al Annals of Nuclear Medicine

than in FBP. used that cannot reflect many of the subtle features of real
Sphere contrast in simulated "hot" lesions saturates clinical images. The analyses are based on computer
after about 80 iterations in OSEM reconstruction. The measurements of quantities without interaction with hu-
finding of contrast saturation at increasing number of man observers. Thus, the detailed quantitative results and
iterations after Gaussian regularization was previously conclusions are likely correct for the specific phantom
reported by Wallis et al. 26 for ML and a number of imaging conditions, but the reader may wonder whether
iterative-weighted methods with and without attenuation the conclusions are applicable to the real word of clinical
correction. This is a typical behavior of OSEM and does images. Moreover, the problem of algorithm selection is
not arise from a termination of contrast increasing but intrinsically multi parametric and a simple recipe that
from noise increasing; the real behavior of OSEM is holds for every situation cannot be devised. The optimal
covered by the 3D Gauss filter and the number 80 is due set of parameters depends on clinical task. For instance,
to the filter, not to iterative reconstruction. The contrast at if we are looking for hot lesions, then OSEM with the
80 iterations is higher in OSEM than in FBP for every suggested threshold of 80 EM equivalent iteration and
level of lesion-to-background activity ratio, both with without AC correction will lead to an increase in contrast,
high and low background. The contrast gain using OSEM a better spatial resolution and only a slight increase of
ranges from 6.7% to 16.4% for different Cnom. noise that, in a study with a high count statistic, may be
The use of first order Chang's AC leads to diminished disregarded. On the other hand, if the study count statistic
contrast values both in FBP and OSEM; again contrast is low the increase of noise could be harmful and then we
values for OSEM-AC were higher than the ones obtained could better switch to FBP or OSEM with a smaller
with FBP-AC. Kauppinen et a1.14 reported a contrast gain number of iterations to reach noise equivalence (for
ranging from 7.1% to 12.5%, using a brain phantom with instance at 23 EM equivalent iterations in low background
a gray-to-white matter activity ratio of 4:1, reconstructed studies). In this case we still have better resolution with
with OSEM + variable 3D Gauss (72 iterations) and FBP OSEM, but contrast for hot lesions will be worse than with
(Butterworth: order 6, cutoff = 0.6 cm-1), both with FBP. If we are looking for cold lesions, the equivalence of
attenuation correction. contrast between OSEM and FBP is reached at 80 EM
This situation is different when taking into account equivalent iterations, but noise is higher for OSEM than
sphere contrast in "cold" lesions: contrast increases in for FBP; the application of AC in this context will provide
OSEM reconstruction as iteration number increases, better contrast both for OSEM and FBP.
without reaching saturation. The equivalence between
contrast in OSEM and FBP is reached at 80 iterations. REFERENCES
Contrast values are dependent on background activity
concentration, being smaller with high background. More- 1. Barret HH, Wilson DW, Tsui BMW. Noise properties of the
over, the application of AC leads to a slight increase in EM algorithm: part I. Theory. Phys MedBio11994; 39: 833-
contrast both with OSEM and FBP. 846.
The tomographic spatial resolution with scatter ex- 2. Liow J-S, Strother SC. Practical tradeoffs between noise,
pressed as axial, radial and tangential FWHM did not vary quantitation and number of iterations for maximum likeli-
with increasing EM-equivalent iteration number. This hood-based reconstruction. IEEE Trans Med Imaging 1991;
10: 563-571.
finding was expected since resolution mainly depends on
3. Miller TR, Wallis JW. Clinically important characteristics
the collimator characteristics. The values obtained are
of maximum-likelihood reconstruction. J Nucl Med 1992;
significantly lower than the ones obtained with FBP with 33: 1678-1684.
a high-resolution filter of clinical use. Also FBP with 4. Snyder DL, Miller MI, Thomas LJ Jr, Politte DG. Noise and
ramp filter shows FWHM values slightly higher than edge artifacts in maximum likelihood reconstruction for
OSEM. emission tomography. IEEE Trans Med Imaging 1987; 6:
It must be acknowledged that conclusions are strongly 228-238.
affected by the adopted 3D Gauss filter of OSEM; a 5. Hudson HM, Larkin RS. Accelerated image reconstruction
different filter would likely lead to other conclusions, at using ordered subsets of projection data. IEEE Trans Med
least from a quantitative point of view: without the 3D Imaging 1994; 13: 601-609.
Gauss filter, OSEM optimization should be found more 6. Yokoi T, Shinohara H, Onishi H. Performance evaluation
of OSEM reconstruction algorithm incorporating three-
rapidly than after 80 iterations. Moreover OSEM should
dimensional distance-dependent resolution compensation
have better spatial resolution and contrast than FBP, both
for brain SPECT: a simulation study. Ann Nucl Med 2002;
in hot and cold lesions, but with notable increase of noise 16: 11-18.
with increasing EM-equivalent iteration number. 7. Lau YH, Hutton BF, Beekman FJ. Choice of collimator for
To devise the best performing image reconstruction cardiac SPET when resolution compensation is included in
algorithm with an optimal set of parameters in clinical use iterative reconstruction. Eur J Nucl Med 2001 ; 28: 39--47.
is an issue that cannot be addressed fully by the present 8. Hutton BF, Baccarne V. Efficient scatter modelling for
work. First of all, this is because a geometric phantom is incorporation in maximum likelihood reconstruction. Eur J

Vol. 19, No. 2, 2005 Original Article 81

 NuclMed 1998; 25: 1658-1665. 18. lflikar I, Koutelou M, Mahmarian JJ, Verani MS. Simulta-
9. Hutton BF, Lau YH. Application of distance-dependent neous perfusion tomography and radionuclide angiography
resolution compensation and post-reconstruction filtering during dobutamine stress. J Nucl Med 1996; 37: 1306-
for myocardial SPECT. Phys Med Bio143: 1679-1693. 1310.
10. Kamphuis C, Beekman FJ, van Rijk PP, Viergever MA. 19. Kiat H, Maddhai J, Roy LT, et al. Comparison of technetium
Dual matrix ordered subsets reconstruction for accelerated 99m methyl isobutyl isonitrile ant thallium 201 for evalua-
3D scatter compensation in single-photon emission tomog- tion of coronary artery disease by planar and tomographic
raphy. Eur J Nucl Med 1998; 25: 8-18. methods. Am HeartJ 1989; 117:1-11.
11. Lalush DS, Tsui BMW. Performance of ordered-subset 20. Taillefer R, DePuey EG, Udelson JE, Bellar GA, Latour Y,
reconstruction algorithms under conditions of extreme at- Reeves F. Comparative diagnostic accuracy of TI-201 and
tenuation and truncation in myocardial SPECT, JNucl Med Tc-99m sestamibi SPECT imaging (perfusion and ECG-
2000; 41: 737-744. gated SPECT) in detecting coronary artery disease in women.
12. Blocklet D, Seret A, Popa N, Schoutens A. Maximum- JAm Coil Cardiol 1997; 29: 69-77.
likelihood reconstruction with ordered subsets in bone 21. Chang LT. A method for attenuation correction in radionu-
SPECT. J Nucl Med 1999; 40:1978-1984. clide computed tomography. IEEE Trans Nucl Sci 1978; 25:
13. Case JA, Licho R, King MA, Weaver JP. Bone SPECT of 638-643.
the spine: a comparison of attenuation correction tech- 22. Beekman FJ, Viergever MA. Evaluation of fully 3D itera-
niques. J Nucl Med 1999; 40:604--613. tive scatter compensation and post-reconstruction filtering
14. Kauppinen T, Koskinen MO, Alenius S, Vanninen E, Kuikka in SPECT. In: Grangeat P, Amans J-L, eds. Three-dimen-
JT. Improvement of brain perfusion SPET using iterative sional image reconstruction in radiology and nuclear medi-
reconstruction with scatter and non-uniform attenuation cine. Dordrecht Boston London; Kluwer Academic Pub-
correction. Eur J Nucl Med 2000; 27:1380-1386. lishers, 1996: 163-175.
15. Vanhove C, Defrise M, Franken PR, Everaert H, Deconinck 23. Pedroli G, Crespi A, De Agostini A, Farinelli C, Moretti R,
F, Bossuyt A. Interest of the ordered subsets expectation et al. The quality control of SPET systems: results of an
maximization (OS-EM) algorithm in pinhole single-photon inter-laboratory comparison study in Italy. J Nucl Med and
emission tomography reconstruction: a phantom study. Eur All Sci 1989; 33: 396--400.
J Nucl Med 2000; 27: 140-146. 24. Winer BJ. Statistical principles in experimental design.
16. Wells GR, King MA, Simkin PH, Judy PF, Brill AB, New York; McGraw-Hill, 1971.
Gifford HC, et al. Comparing filtered backprojection and 25. Hutton BF, Hudson HM, Beekman FJ. A clinical perspec-
ordered subsets expectation maximization for small-lesions tive of accelerated statistical reconstruction. Eur J Nucl
detection and localization in 67Ga SPECT. J Nucl Med Med 1997; 24: 797-808.
2000; 41: 1391-1399. 26. Wallis JW, Miller TR. Rapidly converging iterative recon-
17. Gonzalez RC, Wood RE. Digital Image Processing. Read- struction algorithms in single-photon emission computed
ing, MA; Addison Wesley Publishing, 1992: 208-209. tomography. JNucl Med 1993; 34: 1793-1800.

82 Marco Brambilla, Barbara Cannillo, Marco Dominietto, et al Annals of Nuclear Medicine