CHAPTER 3 : MATERIALS AND METHODS
3.1 GEMS DISCOVERY 690 PET-CT SYSTEM
The PET tomograph used in this investigation was the GEMS Discovery 690 PET-CT system, which combines a lutetium-yttrium orthosilicate (LYSO) PET tomograph with a 64- slice CT scanner. The system is shown in Figure 3.1.
3.1.1 Hardware
Figure 3.2 illustrates the crystal arrangement of the Discovery 690 PET system. Each individual LYSO crystal has dimensions 6.3mm (axial) x 4.2mm (transaxial) x 25mm (radial). 54 crystals in a 9x6 arrangement form a detection block. The dimensions of this square block of crystals are marginally increased by the use of reflective material between crystals to prevent optical spillover. The block of crystals is optically coupled to a four-anode PMT (Figure 3.2 (B)) to complete a detector block.
Figure 3.2: GEMS Discovery 690 crystal arrangement
The PET tomograph consists of four rings of 64 detector blocks (or 24 rings of 576 crystals). This creates a crystal ring diameter of 810mm (Figure 3.2 (C)), which provides an imaging FOV of 700mm [167]. The axial FOV created by the 24 crystal rings is 157mm (Figure 3.2 (D)). The PET tomograph has 256 detector blocks in total, containing 13,824 crystals. It operates only in 3D acquisition mode: all 24 crystal rings are able to form LORs with each other.
The Discovery 690 is equipped with a powerful processing system (IBM BladeCentre), designed to accelerate image reconstruction and data processing.
3.1.2 Reconstruction Software
3.1.2.1 OSEM Algorithm
The standard reconstruction method used by the Discovery 690 is a fully 3D OSEM reconstruction, known as VUE Point HD (VPHD). This algorithm includes corrections for scatter, randoms, and attenuation inside the iterative loop [36]. The VPHD algorithm uses a system matrix that accounts for the system’s geometry (block-based crystal distribution and detector curvature), normalisation and dead time.
3.1.2.2 TOF and PSF
TOF data can be included in the reconstruction process by selecting the VUE Point FX (VPFX) algorithm. This applies timing information to each correction step within the iterative loop [47]. The Discovery 690 uses a timing kernel of 650ps in the reconstruction process [168]. Applying this to Equation 1.2 translates to a positional uncertainty of 9.75cm. The reconstruction can also include PSF modelling by selecting the SharpIR option. This PSF model was developed by measuring the detector response to a point source placed at discrete locations throughout the radial and axial directions in the FOV [49], [89]. The detector response was then incorporated into the system matrix used by the OSEM reconstruction algorithm.
Table 3.1 summarises the four reconstruction methods used in this thesis. The abbreviations in the right-hand column are used throughout this thesis.
GEMS Notation Description Thesis Notation
VPHD OSEM Algorithm HD
VPHD-S OSEM plus SharpIR PSF
VPFX OSEM plus TOF TOF
VPFX-S OSEM plus TOF and SharpIR PSF+TOF
Table 3.1: Thesis notation for reconstruction methods
3.1.2.3 Effective Iterations
The OSEM algorithm allows the user to select the number of subsets and iterations used for reconstruction. The maximum permitted number of subsets is 48, whilst the maximum permitted iterations is 200; the maximum number of effective iterations is therefore 9,600.
In this thesis, the number of subsets was fixed at 18, while the number of iterations was varied. This follows the methodology used by Bettinardi et al [19] in their assessment of the Discovery 690’s performance. Most experiments in this thesis used between 18 and 540 effective iterations. This range of effective iterations was chosen to provide a reasonable spread of results and should demonstrate convergence under most circumstances (this goes beyond the range used by Bettinardi et al, who stopped at 360 iterations). The remainder of this thesis describes reconstructions in terms of the effective iterations employed, instead of subsets and iterations.
3.1.2.4 Voxel sizes
The Discovery 690 has unevenly sampled projections due to the system geometry (block- based crystal distribution and detector curvature). The OSEM algorithm uses this projection data to reconstruct an image volume consisting of cuboid voxels. Distance driven projectors use the known detector boundaries and the position of voxels within the FOV to determine a detector’s contribution to a particular voxel in the forward projection process, and a voxel’s contribution to a particular detector during the back projection process [169].
The dimensions of the reconstructed voxels depend on both selected matrix size and transaxial FOV. Three matrix sizes can be selected for the transaxial images: 128*128, 192*192 and 256*256. The maximum possible transaxial FOV is 700mm, which is used for whole-body imaging. This can be reduced, e.g. for paediatrics or brain imaging, which in turn reduces the voxel dimensions in the x & y planes. The minimum possible FOV is 64mm; selectable voxel sizes therefore range from 0.25mm up to 5.47mm. The axial FOV is fixed at 157mm, and the axial sampling (z-axis voxel size) is fixed at 3.34mm. This produces 47 transaxial image slices for a single frame acquisition and cannot be altered by the user.
3.1.2.5 Post-Reconstruction Filtering
The Discovery 690 has two separate filter options that can be applied to the OSEM reconstructed data: one that is applied transaxially (x & y axes) and one that is applied axially (z-axis).
The transaxial filter is a two-dimensional Gaussian filter, which is defined by selecting the filter’s FWHM in millimetres. As the filter is Gaussian, the standard deviation, σ, can be calculated for a given FWHM as follows:
The GEMS filter design is truncated to ± 4σ, and therefore the full width of the filter is 8σ. As the truncation of the filter is minimal, it is assumed that the filter closely approximates a Gaussian curve in the frequency domain. The two-dimensional filter is implemented as a one-dimensional filter over the transaxial image rows and a one-dimensional filter over the transaxial image columns (Dr Charles Stearns, personal correspondence, 21st July 2014).
The z-axis filter is a 3-point average filter with four possible weight settings, as shown in Table 3.2. This filter is applied to the corresponding voxels in three contiguous transaxial image slices. The ‘heavier’ the filter weighting, the more smoothing is applied between slices.
Z-Axis Filter Filter Coefficients
None n/a
Light [1 – 6 – 1] ÷ 8
Standard [1 – 4 – 1] ÷ 6
Heavy [1 – 2 – 1] ÷ 4
Table 3.2: GEMS Discovery 690 z-axis filter weights
3.1.2.6 Vendor Suggested Reconstruction Parameters
The following combination of reconstruction parameters was suggested by GEMS for whole-body 18F-FDG imaging upon the system’s installation:
• Both TOF and PSF applied (VPFX-S algorithm)
• OSEM reconstruction with 18 subsets and 3 iterations (54 effective iterations) • 700mm FOV with 192*192 matrix (3.65mm voxel size in transaxial plane) • Post-reconstruction Gaussian filter with 4mm FWHM
• Standard z-axis filter
The time taken to reconstruct a single frame acquisition using this combination of parameters is 1 minute and 46 seconds.
3.1.3 Quality Control and System Calibration
The Discovery 690 is subject to a robust quality control program to ensure consistent quantitative accuracy of reconstructed images over the system’s life cycle. Detector performance is assessed daily by uniformly exposing the crystals to a long-lived 68Ge rod
source and comparing measurements against expected values. Weekly image quality tests are performed by acquiring, reconstructing and analysing images of a uniform phantom. WCC is performed on a quarterly basis, as advised by GEMS, to ensure accurate quantification in the reconstructed images. Annual quality control tests involve repeating the
acceptance tests performed at system installation and comparing with baseline values to ensure there is no degradation in system performance over time.
All experiments undertaken during this study were performed after verifying all appropriate quality assurance (QA) tests had been performed successfully, registration between corresponding PET and CT image volumes was satisfactory and WCC was up-to-date.