Improved Geometric Fidelity of Radiotherapy Target Volumes

1
Improved Geometric Fidelity of Radiotherapy
Target Volumes in 2D and 3D PET Based on Point
Spread Functions
S.L. Breen and H. Keller Department of
Radiation Physics, Radiation Medicine Program,
Princess Margaret Hospital and Department of
Radiation Oncology, University of Toronto,
Toronto, Canada
Purpose / Objectives Positron emission
tomography (PET), which has shown benefit in
diagnosis and staging of many cancers, may be
useful in delineating target volumes. However,
it is difficult to accurately determine the
spatial extent of target volumes on PET images
because of (i) imperfect contrast recovery and
(ii) partial volume effects in small volumes.
Deficiencies in contrast recovery complicate the
setting of a reproducible window and level for
delineating areas of high radiopharmaceutical
uptake. For the precise delivery of
radiotherapy, target volumes must be delineated
with geometric fidelity to match the anatomic
volumes of increased radiotracer uptake. The
question of appropriate threshold levels for
contouring PET images confounds the use of PET in
radiotherapy treatment planning. Phantom studies
and computer simulations were performed to
determine appropriate image thresholds that
produce accurate target volumes.
Results Activity Profiles Acquired
Images Figure 2 shows the activity profile
through the largest sphere. The number of voxels
Ns(T) above a particular threshold T is defined
as the shaded area. An image possessing high
geometric fidelity will have Ns(T) equal to the
number of voxels in the sphere in the phantom
that is, the volume above the threshold will
equal the sphere volume. Image Threshold Figure
3 plots the number of pixels above threshold for
all spheres, acquisition modes, and activity
ratios for the simulated data. The intersection
of the horizontal and curved lines is the
threshold which predicts the correct volume of
each sphere. This data is summarised in Figure
4, where the intersection points are plotted for
all acquisitions and spheres.
The data on Figure 4 are for all acquisitions,
and for the simulated data for the 41 and 81
acquisitions. It is seen that the simulated data
predicts the appropriate thresholds for all
sphere sizes in the 3D acquisition reasonably
well. The 2D data is not so well modeled this
may be due to a higher level of statistical noise
in the 2D images. The 2D thresholds in the
acquired images do not follow the smooth trend of
the 3D case. This may be due to image noise and
reconstruction technique.
Figure 1 IEC phantom for NEMA image quality
test. The four smallest spheres (diameters of
10, 13, 17, and 22 mm) contain 18F in a
concentration either four or eight times
background. The central circle represents a
cylinder of density 0.3 g cm-3 with no activity.
On the right, an axial image through the spheres
is shown for an activity ratio of 81 for a 2D
acquisition.
Simulated Images of Phantom A model of the true
activity distribution in the IEC phantom was
created in MatLab. The inherent blurring of the
PET scanner was simulated by convolving this
activity distribution with a point spread
function which was modeled as an isotropic three
dimensional Gaussian, whose full width at half
maximum (6.8 mm) was obtained from the spatial
resolution measured at 10 cm radius. The
simulated activity distribution, convolved with
the point spread function, is referred to as the
simulated image. Comparison of Simulated and
Acquired Images The activity profile through the
hot spheres was compared for the simulated and
acquired images for both activity ratios and both
acquisition techniques. For each threshold, the
diameter of each sphere was calculated. The
acquired PET images were upsampled to an axial
image size of 512 x 512 pixels.
Methods Image acquisition and reconstruction The
IEC image quality phantom1 was filled with 18F
solution as described in Figure 1. PET-CT images
were acquired in both 2D and 3D mode for both
activity ratios with a hybrid scanner (Discovery
ST, GE HealthCare, Milwaukee, USA). CT data were
used for attenuation correction, and
reconstruction was done with OSEM and FORE
iterative techniques (Table 1). PET images of
the activity per unit volume were transferred to
a MatLab program for analysis.
Figure 4 Nominal thresholds, Ns(T) for all
data, and for the simulated images.
Conclusions The point-spread function technique
models PET image acquisition reasonably well,
particularly for 3D data. The PSF model confirms
the difficulty of thresholding PET images, and
indicates that it is unsuitable to have a single
threshold for volume delineation throughout an
image because the threshold depends on the volume
of the high signal region, and the relative
activity compared to background. A predictive
model for PET thresholds reduces the need for a
series of measurements to calibrate the system
for the range of situations encountered
clinically Future work aims to improve the model
by incorporating noise and improving contrast
recovery.

• Reference
• NEMA NU2 Performance Measurements of Positron
  Emission Tomographs. NEMA Standards Publication
  NU 2-2001. National Electrical Manufacturers
  Association, Rosslyn, USA. 2001

Table 1 Reconstruction parameters for 2D and 3D
acquisition. Randoms and scatter corrections
were applied. Images were reconstructed on 256 ?
256 matrices.
Figure 2 Nominal threshold defined on the
profile through a hot sphere. The shaded area
corresponds to NS(T).