|
Results
Both
metastases are visible as a single hot area when the 128x128
image matrix is used for PET image reconstruction, but they
can not be delineated as separate lesions (Fig. 3a). The OSWLS
method provides a slightly lower image quality regarding the
delineation of the lesion in comparison to the other reconstruction
procedures when the 256x256 matrix is used (Fig. 3a).
Fig
3a
click = zoom
The effect
of attenuation correction was evaluated using the correction
data provided by the PET system as well as an iteratively
reconstructed attenuation map, based on the originally acquired
transmission data (Fig.3b). We noted for all four reconstruction
methods artefacts in areas with high attenuation, which limit
both the qualitative and quantitative assessment of the image.
Fig
3b
click = zoom
Furthermore,
the detectability of the liver metastasis was limited using
the system based attenuation data, especially for the OSISRA
and OSWLS method (Fig: 4b, upper row). In contrast, the image
quality was significantly improved for all reconstruction
procedures when the iteratively reconstructed attenuation
map was used for the correction of the emission data (Fig.
4b, lower row).
Fig 4b
click
= zoom
The visual
evaluation of the four different reconstruction algorithms
demonstrates that the two metastases are visualized clearly
as separate lesions even after three iteration steps for all
reconstruction methods when an image matrix of 256*256 pixel
and four subsets were used (Fig. 4a). However, the image quality
was rapidly deteriorating when the number of iteration steps
was increased without applying the median root prior correction
to the data (Fig. 4a). The two metastases were not detectable
when 12 iterations and OSEM were used for reconstruction (Fig.
4a, lower row, left). The image quality was slightly better
for the three other reconstruction algorithms (Fig. 4a, lower
row).
Fig 4a
click
= zoom
In contrast,
the use of the median root prior correction with mrp=0.3 resulted
in an acceptable image quality for all four reconstruction
methods when at least 3 iteration steps were used for reconstruction
(Fig. 4b). The visual analysis demonstrated no further improvement
when more than 3-6 iteration steps were used for image reconstruction,
provided
that the mrp is applied to the data. While the two small metastases
were noted as separate lesions for OSEM, OSISRA, and OSSAGE
when twelve iteration steps were used, the image quality was
slightly lower for the OSWLS method regarding the delineation
of the lesions (Fig. 4b, lower right).
The total
number of iterations was changed from 1-32 for the four reconstruction
methods and the SUV as well as the noise of the data were
calculated for the metastasis and the normal liver parenchyma
(Tab. 1a-d). The number of iterations was set to six and the
256x256 matrix was used for reconstruction. Using the SUV
of the metastasis with 32 iterations for reference, 90 % of
this value was achieved within 10 iterations for OSEM, OSISRA,
and OSSAGE without using the median root prior correction
(Tab. 1a). The convergence was best for OSWLS, the 90 % value
was achieved within 5 iterations. All methods demonstrated
a fast convergence within 2-3 iteration steps for the normal
liver parenchyma (Tab. 1a). Using a 1 % difference to the
SUV with 32 iterations, 18-24 iterations must be used for
all reconstruction methods (Tab. 1a). However, no real plateau
phase was achieved for the metastasis when the median root
prior correction was not used. We noted oscillating SUV for
the low uptake area and the values did not achieve a constant
level (Tab. 1a).
Table 1a
click
= zoom
In contrast,
the convergence was improved for the metastasis as well as
for the normal liver parenchyma when the median root prior
correction was applied and now the 90 % value was achieved
within 3-5 iterations (Tab. 1b). Even for a 1 % difference
to the SUV with 32 iterations, only 10-12 iterations are required
(Tab. 1b).
Table 1b
click
= zoom
While
the results are comparable for OSEM, OSISRA, and OSSAGE, we
noted unreliable results for the OSWLS method when more than
12 iterations and four subsets were used for reconstruction.
A plateau phase was observed for OSEM, OSISRA, and OSSAGE
for the SUV of the metastasis and the liver parenchyma. The
noise was calculated form the standard deviation and the mean
SUV (Tab. 1c,d). The data show a rapid increase of the noise
with higher numbers of iterations when the median root prior
correction was not used (Tab. 1c).
Table 1c
click
= zoom
In contrast,
we noted an improvement by a factor of 2-3 when the median
root prior correction was applied to the data (Tab. 1d). The
best noise reduction was observed for OSWLS (Tab. 1d)
Table 1d
click =
zoom
The effect
of different values for the mrp correction was assessed using
6 iterations and four subsets for all methods (Fig. 5a). Generally,
the mean SUV in the liver parenchyma was not dependent on
the mrp value when OSEM, OSISRA, or OSSAGE were used, while
a constant decrease of the SUV was observed for the metastasis
(Fig. 5). The difference of the SUV with mrp=0.3 and mrp=0.8
was less than 10 % for OSEM, OSISRA, and OSSAGE, but 22.8
% for OSWLS (Fig. 5a).
5a
click =
zoom
The use
of ordered subsets may be helpful to decrease the overall
reconstruction time. However, we noted a dependency of both
mean SUV and noise on the number of subsets for the metastasis
as well as the normal liver parenchyma (Fig. 5b,c). The OSWLS
provided unstable results when more than 4 subsets were used.
When ordered subsets are used for reconstruction, we can keep
the product "number of iteration steps * number of subsets"
constant. Interestingly, we noted an increase of the mean
SUV for the liver metastasis when the number of iteration
steps was low and the number of subsets was increased (Fig.
5d). Furthermore, the noise was generally higher for both
the metastasis and the normal liver parenchyma when the number
of subsets was increased and less iterations were used (Fig.
5e). In contrast to OSEM, OSISRA, and OSSAGE, the OSWLS method
showed only for 8 subsets and 3 iteration steps a major increase
of the noise (Fig. 5e).
|
fig.
5d
|
 |
|
|
Benchmarks
were performed on different PC systems running with Windows
2000 professional for the four reconstruction methods (Tab.
2).
Table 2
click =zoom
The memory
usage, which was 85 MB for the 256*256 matrix, requires at
least 128 MB memory. Smaller systems may be considered if
the 128*128 matrix is primarily used for reconstruction. The
time per iteration was significantly longer for OSSAGE due
to the pixel update procedure. In general, a typical dynamic
FDG study, including 23 frames for 60 minutes (1449 slices)
is reconstructed within less than 3 hours when a Pentium III
with 700 MHz is used. The software supports semi-parallel
processing, so several PET patient studies may be processed
on several PC systems simultaneously. Besides the use of the
reconstruction program on Windows systems, the reconstruction
program was also compiled on a Linux system (SuSE 6.2, SuSE
GmbH, Nürnberg, Germany) as well as on a system running
BeOS (V4.5, Be Inc., Menlo Park, CA, USA) using the standard
compiler program provided with the system software. In general,
the time per iteration was longer as compared to the Windows
2000 professional system on the same Pentium III with 600
Mhz (OSEM: Linux: 5.6 sec/iteration, BeOS: 6.5 sec/iteration
Windows 2000 professional:
1.35 sec/iteration).
|
