NetForum uses cookies to ensure that we give you the best experience on our website. If you continue to use the site, we'll assume that you are happy to receive these cookies on the NetForum website. Read about our cookies.
NetForum Community
Learn. Share. Optimize.
Log in | Sign up now | Submit content | Contact
Go to similar content

Resolution recovery in the Ingenuity TF PET/CT

White Paper
Philips NM Clinical Science Philips Healthcare • USA

Manoj Narayanan, PhD; Amy Perkins, PhD

Resolution recovery with the PSF algorithm implemented in the Ingenuity PET/CT corrects for partial volume effects in PET images. In this paper, the implementation and usage of PSF will be discussed. The benefits of applying PSF for improved visualization and quantification of structures will be illustrated with examples of simulations, phantom and clinical patient results.

Partial volume effect
PET quantification has greatly improved with the ability of iterative reconstruction algorithms to correct for the effects of attenuation and scatter. In addition, corrections for partial-volume effects (PVE) due to finite scanner resolution are requisite for further improvements in PET quantification. Note that PET spatial resolution is affected by a number of factors: the finite positron range which is radioisotope dependent, non-collinearity of the annihilation photons, depth-of -interaction effects in the crystal, crystal/detector size and reconstruction parameters (voxel dimensions, post-filters etc.). Since PET scanner resolution is spatially variant, it is typically measured by imaging point sources at various locations within the scanner. These measurements then represent the point-spread function (PSF) of the scanner in all three dimensions (x, y and z).

The finite spatial resolution results in a spillover of activity into neighboring regions wherein smaller structures appear dimmer in contrast to larger ones with the same activity concentration. This effect can be explained by a simple simulation example. Consider a phantom that consists of different sized spheres (10, 13, 17, 22, 28 and 37 mm diameter) in a warm background with a contrast to background ratio of 4:1 as shown in Figure 1. Spatial blurring is simulated by convolving (filtering) the object with a Gaussian shaped PSF of 6 mm FWHM (full-width half maximum) in x, y and z dimensions. 
Figure 1 Schematic representation of spatial blurring in a PET scanner with finite (limited) spatial resolution. The true object shown on the left is convolved with a Gaussian point-spread function representing the system resolution, resulting in the blurred image shown on the right. Here X represents convolution.
Figure 1
Schematic representation of spatial blurring in a PET scanner with finite (limited) spatial resolution. The true object shown on the left is convolved with a Gaussian point-spread function representing the system resolution, resulting in the blurred image shown on the right. Here X represents convolution.

A comparison of profiles through the smallest spheres (10 and 13 mm diameter) clearly demonstrate the effect of partial volume where there is spillover of activity from the hot spheres into the neighboring background voxels as shown in Figure 2. This smearing results in lower measured activity concentrations in small objects indicating that they are more affected by partial volume effects.

Alternatively, the largest sphere (37 mm in diameter) is minimally affected by partial volume as seen in Figure 3. A good rule of thumb is that partial volume does not affect objects whose dimensions are greater than 4 times the system resolution. In this simple experiment, given a system resolution of 6 mm FWHM, the maximum activity concentration in the 37 mm diameter sphere is close to the true value of 4 as expected (see Figure 3).
Figure 2 Comparison of profiles of the true object (blue) and
the blurred object (red) for the 10 mm and 13 mm diameter
spheres which have been blurred with a Gaussian PSF of 6 mm
FWHM. The profiles clearly indicate that the maximum activity
concentration value is lower than the true value of 4.Figure 3 Comparison of profiles of the true object (blue) and
the blurred object (red) for the largest 37 mm diameter sphere
which has been blurred with a Gaussian PSF of 6 mm FWHM.
The profiles indicate that the activity concentration (maximum) is
close to the true value of 4 demonstrating that larger objects are
minimally affected by resolution effects.
Figure 2
Figure 3
Comparison of profiles of the true object (blue) and the blurred object (red) for the 10 mm and 13 mm diameter spheres which have been blurred with a Gaussian PSF of 6 mm FWHM. The profiles clearly indicate that the maximum activity concentration value is lower than the true value of 4.
Comparison of profiles of the true object (blue) and the blurred object (red) for the largest 37 mm diameter sphere which has been blurred with a Gaussian PSF of 6 mm FWHM. The profiles indicate that the activity concentration (maximum) is close to the true value of 4 demonstrating that larger objects are minimally affected by resolution effects.

Since the PET scanner resolution reduces the true activity concentrations of smaller objects, resolution recovery techniques can be used in principle to recover the true activity. However, in the presence of noise, resolution recovery methods may amplify noise. Therefore, the application of resolution recovery methods to noisy PET images is ultimately a trade-off between improving the resolution of the images while reducing the effects of noise as shown in the following sections.

Richardson-Lucy (RL) algorithm for resolution recovery
The R-L algorithm1,2 is a widely used maximum likelihood (ML) algorithm used in image restoration. The algorithm incorporates Poisson statistics and is very similar in form to the maximum-likelihood expectation maximization (MLEM) image reconstruction algorithm used in PET. The main short coming of the ML approach is that in the presence of noise, the iterates try to fit the data as closely as possible. In order to avoid over-fitting to the noise, some form of noise-regularization is required. Using the method of sieves proposed by Snyder3, a regularized version of R-L algorithm can be described as shown in the equation 1 below

where ƒˆ is the original undistorted image, g is the distorted noisy image, h is the system PSF, s is the sievekernel used for regularization, X is the convolution operator, while * is the correlation operator. Note that the regularization (sieve) kernel s is specified in terms of mm FWHM.

The effective implementation of the R-L algorithm requires an accurate measurement of the scanner point-spread-function. This was accomplished for the Ingenuity TF PET/CT scanner by careful measurement of point-sources at many points in the field-of-view. Curve fitting using analytical functions were used to model these spatially varying kernel measurements in the implementation of the iterative scheme described in equation 1 above.

Note that in the Ingenuity TF PET/CT, resolution recovery is referred to as PSF and in the remainder of this document, PSF is the nomenclature we will use for resolution recovery. The PSF operation has two parameters that may be modified by the user: a) number of iterations and b) the amount of regularization (in mm FWHM of the Gaussian sieve kernel).

Simulation studies
The benefits of resolution recovery as well as optimal parameter selection for the regularized R-L algorithm can be illustrated using simple simulated objects. Hot spheres were simulated within a warm-background in a 20 cm diameter long cylinder (length of 180 cm) with a contrast to background ratio of 4:1. Three different sized hot-spheres (10, 22 and 37 mm) were simulated to cover the range of blurring that one might see typically. Assuming a system resolution of 6 mm, these objects were blurred with a 3D Gaussian function with a FWHM of 6 mm as shown in Figure 4.
Figure 4 Schematic describing the
simulation details where different sized
spheres (10, 22 and 37 mm diameter)
are embedded in a 20 cm diameter,
long cylinder (length of 180 cm) with
a contrast to background ratio of 4:1.
The original object is convolved with
a 3D Gaussian filter with a FWHM of
6 mm to simulate the blurring that takes
place during PET acquisition. Here
X represents convolution.
Figure 4
Schematic describing the simulation details where different sized spheres (10, 22 and 37 mm diameter) are embedded in a 20 cm diameter, long cylinder (length of 180 cm) with a contrast to background ratio of 4:1. The original object is convolved with a 3D Gaussian filter with a FWHM of 6 mm to simulate the blurring that takes place during PET acquisition. Here X represents convolution.

Resolution recovery in the presence of noise

Noise was added to the images4 and effectiveness of PSF was assessed both with and without regularization. With no regularization, Figure 5 indicates that the error (Euclidean distance between the true object and the reconstructed image) as well as the background noise (calculated as the ratio of standard-deviation divided by mean) tends to increase as the number of iterations increase. For all three objects (10 mm, 22 mm and 37 mm diameter spheres), 1 or 2 iterations are sufficient to get the benefits of resolution recovery, with higher iterations only increasing the noise in the final image.
Figure 5 Comparison of images a) without PSF, b) after 1 iteration of PSF, c) the quantification
error as a function of iteration number and d) the increase in background noise as a function of
iteration. Rows 1 to 3 represent the results for the three different sized objects (10 mm, 22 mm
and 37 mm diameter spheres). Note that 2 iterations are sufficient for adequate resolution
recovery for all three objects.
Figure 5
Comparison of images a) without PSF, b) after 1 iteration of PSF, c) the quantification error as a function of iteration number and d) the increase in background noise as a function of iteration. Rows 1 to 3 represent the results for the three different sized objects (10 mm, 22 mm and 37 mm diameter spheres). Note that 2 iterations are sufficient for adequate resolution recovery for all three objects.

Next, the effect of including PSF regularization to control the noise was testing by increasing the amount of regularization (0, 6, 10, 15 and 20 mm FWHM) while keeping the number of iterations fixed at 1. Figure 6 shows the resulting images and line profiles for each of these regularization values that are being evaluated. In particular, when regularization values are chosen to be much greater than the actual scanner resolution (15 and 20 mm for example), it not only creates visible edge-enhancement artifacts, but also results in overestimation of activity in the hot spheres. On the other-hand, a regularization value of 6.0 provides a small improvement in image resolution without overestimating the activity. This suggests that it is preferable to choose a sieve-kernel that is similar to the system resolution.

A closer look at Equation 1 is useful for understanding the contribution of the regularization to the final image. From Equation 1, we note that both the system PSF (h) and regularization (s) kernels are convolved together. This means that the addition of the regularization term is equivalent to specifying an overall system resolution where the measured system PSF and the regularization terms add in quadrature i.e., √PSF2+Regularization2. Since we assume a system PSF of 6 mm for this simulation, regularization values of 6, 10, 15 and 20 mm are equivalent to effective system resolutions of 8.5, 11.7, 16.2 and 20.8 mm, respectively. At large regularization values of 15 and 20 mm that are significantly greater than the system resolution, the R-L algorithm tries to over-correct, resulting in edge enhancements at object boundaries and overestimation of the activity concentrations.

PSF default settings
As these simulations illustrate, the PSF parameters (PSF Iterations and PSF Regularization) must be chosen judiciously. Typically, 1 to 2 iterations are sufficient. Similarly, choosing a PSF regularization of 6-8 mm appears adequate to provide images with good resolution without excessive noise amplification or quantification errors. Using these guidelines, the whole body, body-detail, neck and brain protocols have the following PSF default parameters:
  • PSF Iterations = 1 
  • PSF Regularization = 6 mm 

The following sections summarize results from phantom and clinical patient studies providing guidance as well as highlighting the pitfalls with incorrectly chosen PSF parameters.


Effect of voxel size on resolution recovery
Based on the discussions so far, it is reasonable to assume that voxel size can play a role in reducing the effect of partial volume and improve resolution recovery. Six different sized spheres (0.5, 1, 2, 4, 8 and 16 ml) within the Jaszczak phantom were filled with different activities (SUV of 29, 21, 13.5, 9, 4.5 and 3, respectively) and scanned using a single bed position for 90 seconds. The images were reconstructed as described below using both 2 mm and 4 mm voxel dimensions:

  • 4 mm Whole-body with no PSF 
  • 4 mm Whole-body with PSF Iterations = 1 and PSF Regularization = 6 mm 
  • 2 mm Whole-body with no PSF 
  • 2 mm Whole-body with PSF Iterations = 1 and PSF Regularization = 6 mm 

Average SUV values were computed from each sphere using a region-of-interest (ROI) drawn using a threshold of ≥80% of the SUV-max in that sphere. SUV-mean values for the different reconstructions are tabulated in Table 1 along with the % error (100[SUV-mean –SUVtruth]/ SUVtruth). Figure 7 compares the corresponding images for each of these reconstruction methods.

Table 1 indicates that applying PSF to the 2 mm reconstructions provide the best quantitative results for the smaller sized spheres (0.5 and 1 ml spheres) i.e., SUV-mean values that are a lot closer to the true SUV values. Alternately, when PSF is applied to the 4 mm reconstructions there is an expected increase in the SUV-mean values in both the 0.5 and 1 ml spheres in comparison to the 4 mm reconstruction with no PSF. However, even the addition of PSF to the 4 mm reconstructions still provide SUV-mean values that are considerably lower than the corresponding 2 mm reconstructions for both the 0.5 and 1 ml spheres. On the other hand, both the 4 mm and 2 mm reconstructions with PSF provide better quantitative estimates for the larger 8 and 16 ml spheres since the spillover effect due to partial-volume is lower in these cases. However, the 2 mm reconstructions with PSF still provide the best overall performance in terms of quantitation over the different sized spheres considered in this experiment. This is also qualitatively evident from Figure 7 where the 2 mm reconstruction with PSF provides images with the best contrast resolution.
Table 1 Tabulation of SUV-mean values and the % error in parenthesis for 6 different sized
spheres (0.5, 1, 2, 4, 8, and 16 ml), each filled with a different activity concentration (SUV
values of 3, 4.5, 9, 13.5, 21 and 29, respectively). Results indicate that the 2 mm reconstruction
with PSF (WB-2 mm PSF 1 iters , Reg = 6) converges closer to the true SUV value than the
corresponding 4 mm reconstructions with PSF (WB-4 mm PSF 1 iters, Reg=6).
Table 1
Tabulation of SUV-mean values and the % error in parenthesis for 6 different sized spheres (0.5, 1, 2, 4, 8, and 16 ml), each filled with a different activity concentration (SUV values of 3, 4.5, 9, 13.5, 21 and 29, respectively). Results indicate that the 2 mm reconstruction with PSF (WB-2 mm PSF 1 iters , Reg = 6) converges closer to the true SUV value than the corresponding 4 mm reconstructions with PSF (WB-4 mm PSF 1 iters, Reg=6).
Figure 7 Reconstructed images of
the Jaszczak phantom with 6 fi llable
spheres (0.5, 1, 2, 4, 8, and 16 ml) .
The images from left to right represent
a) 4 mm Whole-Body with no PSF
applied, b) 4 mm Whole-body with PSF
iteration = 1 and PSF Regularization
= 6 mm, c) 2 mm Whole-body with
no PSF applied and d) 2 mm Whole-body
with PSF iteration = 1 and PSF
Regularization = 6 mm.
Figure 7
Reconstructed images of the Jaszczak phantom with 6 fi llable spheres (0.5, 1, 2, 4, 8, and 16 ml) . The images from left to right represent a) 4 mm Whole-Body with no PSF applied, b) 4 mm Whole-body with PSF iteration = 1 and PSF Regularization = 6 mm, c) 2 mm Whole-body with no PSF applied and d) 2 mm Whole-body with PSF iteration = 1 and PSF Regularization = 6 mm.

This indicates that it is preferable to use a finer sampling with 2 mm reconstructions (rather than the 4 mm reconstructions) with PSF in phantom studies if quantitative accuracy is the objective. Note that for the same scan duration, when going from a 4 mm reconstruction to a 2 mm reconstruction, we need to account for a decrease in counts/voxel (by a factor of 8) in the 2 mm reconstructions. This additional noise could be compensated by appropriately increasing the acquisition time per bed position.

ACR phantom results
The ACR phantom was imaged on the Ingenuity TF PET/CT at the Hospital of the University of Pennsylvania. One section of the phantom has 4 hot cylinders: diameters = 8 mm, 12 mm, 16 mm, and 25 mm, as well as 3 cold 25 mm diameter regions: water, air, and Teflon. The hot cylinders were filled with the standard ACR 2.5:1 contrast ratio5. The data were reconstructed with and without PSF applied. The number of PSF iterations and the regularization parameter were varied, but the TOFlistmode reconstruction parameters were kept at the default settings.

The effect of applying PSF is demonstrated in Figure 8 which shows transverse slices of 2 mm reconstructions of the ACR phantom without PSF and with PSF applied with different iteration and regularization settings. The edges of the cylinders appear sharper in the PSF images with little or no regularization as compared to the image without PSF. There is a noticeable edge artifact when a large regularization of 20 mm FWHM is used. Also, the counts in the hot spheres with a large regularization of 20 mm are over enhanced.

The quantitative results of the noise and the SUV-mean in the background are listed in Table 2. The noise metric was calculated as the standard deviation of the counts divided by the mean counts in a large central background region. As expected, the noise level increases with increased number of PSF iterations in the absence of any regularization (see PSF reconstruction with 3 iterations and no regularization in Table 2). However, the addition of regularization results in lowering the noise-level. Please note that the SUV-mean in the background remains at the correct value of 1.0 (no bias) for the cases of no-PSF, PSF with 3 iterations and the recommended PSF setting of 1 iteration and 6 mm regularization. However, choosing a large regularization value of 20 mm FWHM with 3 iterations of PSF increases the number of counts in this background region increasing the background SUV-mean to 1.2. This indicates that excessively large regularization values results in the unwanted effect of redistributing counts within the image.
Figure 8 Transverse slices of the resulting 2 mm voxel ACR phantom images with no PSF, 3
iterations of PSF, 1 iteration of PSF and regularization with a 6 mm FWHM, and 3 iterations
of PSF and regularization with a 20 mm FWHM. Data courtesy of the Hospital of the
University of Pennsylvania.
Figure 8
Transverse slices of the resulting 2 mm voxel ACR phantom images with no PSF, 3 iterations of PSF, 1 iteration of PSF and regularization with a 6 mm FWHM, and 3 iterations of PSF and regularization with a 20 mm FWHM. Data courtesy of the Hospital of the University of Pennsylvania.
Table 2 Noise and SUV-mean in the background for the images shown in Figure 8.
Table 2
Noise and SUV-mean in the background for the images shown in Figure 8.

Following the ACR procedure5, the SUV-max value for each hot cylinder was recorded and the results are shown as a function of cylinder diameter in Figure 9. The choice of a single iteration of PSF and a regularization of 6 mm FWHM results in an 8 – 12% increase in contrast over no PSF. Increasing the number of PSF iterations to 3 (with no regularization) does not provide any additional recovery in the SUV-max values. Choosing a large regularization of 20 mm FWHM overestimates the SUV-max in the hot cylinders, as seen visually in Figure 8.
Figure 9 Measured SUV-max as a function of cylinder size for
the 2 mm whole-body reconstructions with and without PSF.
Note that the expected SUV in the hot spheres is 2.5.
Figure 9
Measured SUV-max as a function of cylinder size for the 2 mm whole-body reconstructions with and without PSF. Note that the expected SUV in the hot spheres is 2.5.

As noted in the previous section (Effect of Voxel size on Resolution Recovery), reconstructing with a smaller voxel size of 2 mm in contrast to a larger voxel size of 4 mm can result in improved contrast recovery and quantifi cation accuracy. To illustrate this effect, Figure 10 shows a comparison of the cold rod section of the ACR phantom (diameters = 4.8 mm, 6.4 mm, 7.9 mm, 9.5 mm, 11.1 mm, and 12.7 mm, confi gured in a pie-shape pattern) for the following three reconstructions: a) 4 mm with no PSF, b) 2 mm with no PSF and c) 2 mm with the default PSF setting of 1 iteration and 6 mm regularization. Figure 10 shows that even without PSF, the 2 mm reconstruction (middle) shows a clear improvement in resolving the 6.4 mm rods over the 4 mm reconstruction (left). A further improvement in the rod contrast and edge delineation is seen with the 2 mm reconstructions with default PSF settings (right). In addition, the SUV-max values in the hot cylinders (see Table 3) indicates that the 2 mm reconstruction with default PSF parameters yields about a 12% increase in SUV-max for the larger hot cylinders compared to the standard 4 mm reconstruction without PSF and this improvement increases to about 20% for the smaller cylinders. 
Figure 10 Transverse slices of the ACR phantom pie sector section for a standard 4 mm
reconstruction (left), a 2 mm reconstruction (middle), and a 2 mm reconstruction with the
default PSF settings (right). Data courtesy of the Hospital of the University of Pennsylvania.
Figure 10
Transverse slices of the ACR phantom pie sector section for a standard 4 mm reconstruction (left), a 2 mm reconstruction (middle), and a 2 mm reconstruction with the default PSF settings (right). Data courtesy of the Hospital of the University of Pennsylvania.

These results indicate that the recommended PSF settings of 1 iteration and 6 mm regularization can provide reconstructions with improved resolution without excessive amplification of the noise. Choosing large regularization parameters (20 mm in this example) can introduce artifacts in the final image. Additionally, smaller 2 mm voxel reconstructions with the default PSF parameters can provide qualitative and quantitative improvements over the standard 4 mm reconstructions. However, it is recommended that acquisition times be appropriately increased for 2 mm voxel based acquisitions to account for the corresponding decrease in counts/voxel.

NEMA image quality phantom results
The NEMA image quality phantom was prepared with a contrast to background ratio of 4:1 and scanned. The images were reconstructed using a 2 mm voxel size with the default PSF settings of 1 iteration and regularization of 6 mm as well as without PSF. The percent contrast as well as the percent background variability for each of the 6 spheres are tabulated in Table 4. Results indicated that we obtain improved contrast when PSF is applied. For example, we see the % contrast improve from 38.8 % to 48.8% for the 10 mm sphere. We also see a small increase in the background variability when PSF is applied, which is expected. However, this represents an acceptable compromise between a slight enhancement in noise in return for the improved contrast that PSF provides.
Table 3 The ACR-prescribed SUV-max values for the hot cylinders for 3
different reconstructions: standard 4 mm reconstruction, 2 mm reconstruction
without PSF, and 2 mm default PSF reconstruction.
Table 3
The ACR-prescribed SUV-max values for the hot cylinders for 3 different reconstructions: standard 4 mm reconstruction, 2 mm reconstruction without PSF, and 2 mm default PSF reconstruction.
Table 4 Comparison of % contrast and % background variability for the NEMA Image quality phantom
with and without PSF. Default PSF parameters of 1 iteration and regularization of 6 mm were used in the
PSF reconstruction. Note the improved % contrast with the PSF reconstructions which comes with a slight
increase in % background variability or noise.
Table 4
Comparison of % contrast and % background variability for the NEMA Image quality phantom with and without PSF. Default PSF parameters of 1 iteration and regularization of 6 mm were used in the PSF reconstruction. Note the improved % contrast with the PSF reconstructions which comes with a slight increase in % background variability or noise.

PSF evaluation with patient data

Whole body data
Five whole body patient data from the Hospital of the University of Pennsylvania were selected and 2 mm reconstructions were performed with and without PSF applied. PSF was applied using the default settings of 1 iteration and a regularization parameter of 6 mm FWHM. Figure 11 illustrates the improvement in visualizing small lesions with PSF.
Figure 11 Illustration of a patient 2 mm whole-body study with (a) PSF and
b) without PSF. Note the improved delineation of the lesion (arrow) in the
images with PSF. Patient data courtesy of the Hospital of the University of
Pennsylvania.
Figure 11
Illustration of a patient 2 mm whole-body study with (a) PSF and b) without PSF. Note the improved delineation of the lesion (arrow) in the images with PSF. Patient data courtesy of the Hospital of the University of Pennsylvania.

For quantitative analysis, three ROIs were drawn for each patient study: a region in the lesion and larger regions in the lung and the liver. The same ROI set was used for the images with and without PSF for the same patient data to evaluate the impact of PSF. The percent change in SUV-mean with PSF applied when compared to the case without PSF is shown in Figure 12. The average increase in the lesion SUV-mean across the five patients is 18%. The sizes of these lesions are in the range 12-20 mm. The increase in lesion SUV-mean is consistent with the SUV increase observed for the phantom measurements.

The resulting percentage SUV-mean change in larger regions such as the lung and liver are not expected to change significantly with the addition of PSF, similar to the consistency of background SUV-mean values in the ACR phantom results. The results are shown in Figure 13. In these larger structures the percent change in SUV-mean are < ±5%, and the average value across all five patients is approximately 0, as expected.

The percent increase in the noise level with PSF was measured in the liver and lung and is shown in Figure 14. The average increase in the noise over the five patients is 14%, consistent with the percent increase in noise for the ACR phantom.
Figure 12 The percent change in lesion SUV-mean with PSF applied when
compared to the case without PSF. The average value across all 5 patients is
shown in green.Figure 13 The percent change in SUV-mean in the lung and liver.
Figure 12
Figure 13
The percent change in lesion SUV-mean with PSF applied when compared to the case without PSF. The average value across all 5 patients is shown in green.
The percent change in SUV-mean in the lung and liver.

Brain data
Five brain patient data from the Hospital of the University of Pennsylvania were selected and reconstructions were performed with and without the default PSF applied. Figure 15 illustrates the improved resolution in the cortical structures with PSF.
Figure 14 The percent change in noise with PSF applied in the lung and liver.
The average value across all 5 patients is represented by the lines.
Figure 14
The percent change in noise with PSF applied in the lung and liver. The average value across all 5 patients is represented by the lines.

ROIs were drawn for each patient study: a region in the striatum, a region in the white matter, and a larger region in the cerebellum. The same ROI set was used for the images with and without PSF for the same patient data for a direct evaluation of the impact of PSF. The percent changes in SUV-mean of the smaller structures are shown in Figure 16. The average percent increase in the striatum SUV-mean with PSF across the 5 patients is 6%. The white matter is colder than the surrounding hot brain areas and the average percent change in white matter SUV-mean with PSF is 10% lower than without PSF. These structures are narrow but extended and, as expected, exhibit less of an increase in SUV than the more sphere-like lesions in the previous whole-body study.
Figure 15 Illustration of a patient brain study with (a) PSF and b) without PSF.
Note the improved delineation of small structures such as the striatum (arrow)
in the images with PSF. Patient data courtesy of the Hospital of the University
of Pennsylvania.Figure 16 The percent change in striatum (upper) and white matter (lower)
SUV-mean with PSF applied when compared to the case without PSF. The
average value across all 5 patients is shown in green.
Figure 15
Figure 16
Illustration of a patient brain study with (a) PSF and b) without PSF. Note the improved delineation of small structures such as the striatum (arrow) in the images with PSF. Patient data courtesy of the Hospital of the University of Pennsylvania.
The percent change in striatum (upper) and white matter (lower) SUV-mean with PSF applied when compared to the case without PSF. The average value across all 5 patients is shown in green.

Figure 17 shows the results from the larger cerebellum region, taken as a surrogate for a background measurement. The resulting percent changes in SUVmean in the cerebellum are < ±5%, and the average value is approximately 0, as expected. The percent change in noise in the large cerebellum region with PSF applied is 5% for the cerebellum, as some increase in noise is consistent with applying PSF.
Figure 17 The percent change in SUV-mean in the large region of the
cerebellum (upper), and the percent change in noise in the cerebellum with
PSF applied (lower). The average value across all 5 patients is represented
by the blue line.
Figure 17
The percent change in SUV-mean in the large region of the cerebellum (upper), and the percent change in noise in the cerebellum with PSF applied (lower). The average value across all 5 patients is represented by the blue line.

Conclusions
The PSF algorithm in the Ingenuity TF PET/CT corrects for PVE in PET images. Simulations, phantom and clinical patient results show good improvement in image resolution as well as quantification. Since PVE results in lowering the measured activity of objects (particularly objects that are much smaller than the scanner resolution), the application of PSF resulted in a corresponding increase in the SUV’s, which is expected,. Phantom and clinical patient test results indicate that PSF needs to be implemented very carefully, since the choice of parameters can significantly influence the quantitative accuracy. The PSF method implemented in the Ingenuity PET/CT has two user adjustable parameters: number of iterations and a regularization factor. Test results indicate that one to two PSF iterations is an appropriate choice. Similarly, choosing PSF regularization values comparable to the scanner resolution (for example 6-8 mm) provide images with good resolution without excessive noise amplification or quantification errors. Using the suggested PSF default settings of 1 iteration and regularization of 6 mm, clinical and phantom results are consistent with one another, providing better visual delineation of structures, increase (decrease) in SUV for hot (colder) small structures with only a moderate increase in background noise levels. Finally, if quantitative accuracy is paramount, then using a combination of the smaller 2 mm voxels with PSF provides superior results to the corresponding reconstructions that use the larger 4 mm voxels.

References
  1. W. H. Richardson, “Bayesian-based iterative method of image restoration.” J. of Opt. Soc. Of Amer., 1972. 
  2. L. B. Lucy, “An iterative technique for the rectification of observed images.” The Astron. Jour., 1974. 
  3. D. Snyder, “The use of sieves to stabilize images produced with the EM algorithm for emission tomography.” IEEE Nul. Sci, 1985. 
  4. Logan J, Fowler JS, Volkow ND, Ding YS, Wang G-J, Alexoff DL. A strategy for removing the bias in the graphical analysis method. J. Cereb. Blood Flow Metab. 2001; 21:307-320. 
  5. “PET Phantom instructions for evaluation of PET image quality.” ACR Nuclear Medicine Accreditation Program-PET Module, Feb 22, 2010.


This content has been made possible by NetForum Community.
Share this on: Share your link in twitter Share your link in facebook Share your link on LinkedIn Print Rate this article: Log in to vote

 
Rating:
Votes:
4
Views:
9407
Added:
Nov 4, 2013

Rate this:
Log in to vote
 

White Paper
Ingenuity TF PET/CT
Body, brain, Head, iterative reconstruction, lesion, liver, lungs, Neck, Neuro, phantom, PSF algorithm, whole body
 

Clinical News
Best Practices
Case Studies
Publications and Abstracts
White Papers
Web seminars and Presentations
ExamCards
Protocols
Application Tips and FAQ
Training
Try an Application
Business News
Case Studies
White Papers
Web Seminars and Presentations
Utilization Services
Contributing Professionals
Contributing Institutions
Become a Contributor