Scanner geometry
Using MRI images from 22 patients, we have determined the average abdominal dimensions at the axial place where the prostate is located. We found a wide size and patient thickness of 36 cm and 22 cm, respectively (see Fig. 1 left). Considering these values, the prostate-dedicated PET ring design, named ProsPET, has been designed with an aperture of about 410 mm in diameter. The ProsPET ring includes 24 detector blocks separated by a thin gap of just 4 mm. Figure 1 right shows a sketch of the ring geometry and detector positions.
System description
Each one of the 24 detector modules of the ProsPET ring is based on a single monolithic LYSO scintillation crystal of 50 × 50 × 15 mm3. All faces have been polished and the four lateral ones (50 × 15 mm2) black painted, in order to reduce undesired internal reflections. The entrance face of the scintillation block (50 × 50 mm2) was covered with a retroreflector layer that bounces back the scintillation light to the emission point preserving the light distribution profile [16, 17] while increasing the light collection efficiency. The exit face of the scintillator was coupled by means of optical grease (BC630, Saint Gobain) to a photosensor array of 12 × 12 SiPMs. In particular, we made use of silicon photomultiplier (SiPM) arrays of the type C-Series (SensL, now On-Semi) with 3 × 3 mm2 active area, 4.2 mm pitch, and 35 microns cell size [18]. A custom readout electronics based on passive components reduces the 144 SiPM signals to only 24. In particular, the 12 SiPM signals of each row and column of the photosensor array were summed and pre-amplified before transferring to the data acquisition system (DAQ). This readout scheme allows one to characterize the scintillation light distributions in monolithic crystals [16]. The PET scanner has been assembled without any forced cooling approach, but simply requiring the room to be at a stable temperature in between 20 and 25°C (variation ± 0.5°C). The DAQ system was installed in a cabinet under the patient’s bed, where the patient is in supine position.
For each detector block, thin multicoaxial cables (SAMTEC) have been used to exchange the 24 (12 + 12 row and columns) analog signals, the trigger signal (sum of all 24), and the temperature sensor, as well as the amplifiers and SiPM bias. The trigger logic has been programmed to digitize signals only when two trigger signals are within a coincidence window of 5 ns. Charge integrators with a window of 250 ns, and 12-bit precision, are used.
Every detector allows coincidences with its 13 opposite detectors. This defines a transaxial and axial FOV of about 300 and 46 mm, respectively. The axial FOV can be increased to about 80 mm by axially displacing the ring and allowing certain image overlapping (multiacquisition process).
In order to facilitate patients to properly position in the bed and into the scanner, the system has two movable parts that open and close with an accuracy of about 0.5 mm. In Fig. 2, from left to right, one can see the detector ring when it is open, installed in the acquisition bed, with a patient in supine position, and an example in the lateral cubit, respectively.
Calibration data
For every detector block, we run calibration processes of the impact position for both the planar and DOI coordinates, as well as for the energy. This calibration process is either based on 1D polynomials [19] or Voronoi diagrams [20] (see Fig. 3). Similar image performance is obtained with either method. See reference [20] for a detailed comparison of both calibration procedures. Planar coordinates are calculated by raising the 12 digitized signals for each projection to the power of two, before center of gravity (CoG) calculation. The DOI coordinate is estimated for each event as the average for rows and columns (r,c) of the ratio of the sum of all 12 signals (photon energy, E) to its maximum value (퐸/Imax)r,c, [16]. Independently of the calibration procedure, a continuous DOI correction for each line of response (LOR) is carried out. After calibration, list-mode data files are generated, prior to reconstruction.
Simulation platform
We have carried out Monte Carlo simulations of the proposed prostate PET system using GATE v7.2 [21]. The code source is Geant4. For the simulations, the already described LYSO scintillator block was modeled and placed in a ring configuration as schematically shown in Fig. 1. It has to be mentioned that, for simplicity, we have defined back-to-back (511 keV) sources instead of positron emitter sources and therefore some variations were expected between the simulation results and the experimental performance of the detectors.
The simulations results have been used to evaluate the scanner configuration in terms of sensitivity and NECR performance. We established a time resolution of 3 ns, and the time coincidence window to 5 ns, mimicking the real scanner. To make the simulation data more realistic, a Gaussian energy blurring of 15% (based on previous experiments performed with these blocks [17]) was included, as well as energy windows of both 30% and 50% (± photopeak position). A paralyzable deadtime of 1 μs was also considered, as estimated from the real DAQ system.
CASToR reconstruction platform
To reconstruct the list-mode data, we have made use of the CASToR 1.1 platform version [22] and the Siddon projector [23]. CASToR is an open source code that enables to reconstruct not only PET data, but also SPECT and CT as well. CASToR is a generic application that does not estimate correction factors such as normalization, attenuation, scattered, or random counts. Therefore, it is necessary to externally introduce the required correction information.
In particular, we have employed the OSEM algorithm (ordered subsets expectation maximization) with 2 subsets for every reconstruction. The number of iterations has been optimized for the different measurements as it will be described later. 3D images were generated in a binary raw file with a matrix size of 416 × 416 × 50 voxels with 1 × 1 × 1 mm3 voxel size. Since monolithic crystals allow one to define the pixel size, somehow based on the measured detector resolution [24], we have selected virtual detector pixels of 1 × 1 mm2, which is also a compromise between statistics and computational cost.
Normalization correction
Normalization coefficients were calculated across the entire FOV using a custom designed phantom which ensures that all system LORs cross it. The designed phantom consists on a fillable PMMA annulus with inner and outer diameters of 290 mm and 300 mm (50 mm axial), respectively. This particular geometry helps minimizing the number of scattered events [25]. The ring was filled with a solution of FDG and an activity of 7 mCi and positioned in the center of the FOV. Sequential acquisitions over 9 h were carried out. Figure 4 left shows the ProsPET scanner together with the normalization phantom and, on the right, the measured normalization map.
Attenuation correction
Attenuation maps were either obtained through CT images acquired in a separated system (Philips Gemini-TF 64) or by segmentation in the case of uniform phantoms [26]. In the case of CT-based maps, they were co-registered with the PET acquisitions using a specific software, called ITK-SNAP [27]. We recalculated all voxels to values considering the linear attenuation coefficients of each material. For every data acquisition, we have obtained a correction map that is the combination of the aforementioned normalization and the attenuation one. Figure 5 shows the attenuation and combined attenuation-normalization maps for the image quality phantom (to be defined later). On the left image, all the white values are established to 0.096 cm−1, the water value, and the rest are set to 0.
System spatial resolution and sensitivity
All tests were performed without including smoothing filters or random and scatter corrections. The spatial resolution is defined as the width of the reconstructed image point spread function (PSF), calculated using the full width at half-maximum (FWHM). Data was acquired using a 0.25 mm in diameter spherical 22Na source. This radioactive source had an activity of 22 μCi, and thus, the random coincidence rate is almost negligible, as well as the percentage of dead time losses. Each acquisition lasted 10 min. We designed and constructed holders to place the source in different positions along both the transaxial and axial axes; see Fig. 6. The source was moved along the radial direction in steps of 20 mm, both at the center and at 3/8 of the axial FOV.
The optimal number of iterations for the spatial resolution evaluation was studied using the data acquired when the source was located at the center of the field of view (CFOV). The data was reconstructed both with and without including DOI information.
Using the same radioactive source, we also evaluated the system sensitivity. A new holder that allows one to move the source across the axial direction in steps of 5 mm was designed for this purpose; see Fig. 6. Each measurement lasted 10 min.
Quality phantom study
We have evaluated the contrast-to-noise ratio (CNR) and the contrast of reconstructed images using a custom-made phantom. The CNR was calculated as follows:
$$ \mathrm{CNR}=\frac{\mathrm{Mean}\kern0.5em \mathrm{Hot}\kern0.5em \mathrm{Spot}\kern0.5em \mathrm{VOI}-\mathrm{Background}}{\mathrm{Background}\kern0.5em \mathrm{Standard}\kern0.5em \mathrm{Deviation}} $$
(1)
The contrast, in percentage, was calculated using the mean value of each VOI and the background level:
$$ \mathrm{Contrast}\left(\%\right)=100\times \frac{\mathrm{Mean}\kern0.5em \mathrm{Hot}\kern0.5em \mathrm{Spot}\kern0.5em \mathrm{VOI}-\mathrm{Background}}{\mathrm{Mean}\kern0.5em \mathrm{Hot}\kern0.5em \mathrm{Spot}\kern0.5em \mathrm{VOI}} $$
(2)
The phantom used in this study, named quality phantom (QP), is made out of PMMA and has an outer diameter of 135 mm and 103 mm height. The QP includes 6 insert tubes with different diameters (4.5, 6, 9, 12, 15, 20 mm) and 60 mm height each, at an off-center radius of 35 mm [28]. In addition to filling the inserts with FDG, the background of the phantom was also filled but with a different FDG concentration. Two insert-to-background concentrations were measured, namely 38 and 18. The acquisition of the phantom images lasted 8 min each.
Both the background level and its standard deviation have been calculated. We have defined 12 different volumes of interest (VOIs), with the size of the small insert, distributed along uniform areas of the phantom and obtained the mean value of each one. Thereafter, we generated 6 VOIs that fit each insert dimensions, but with a centered height of 25 mm. The CNR and contrast values as a function of the number of iterations were tested after normalization and attenuation corrections.
For comparison purposes, data acquisitions of the QP using the whole-body PET Gemini-TF 64 (Philips) [29] were carried out about 10 min after the measurements performed with the ProsPET system.
Count rate performance
In PET imaging, effects such as scattered and random coincidences might generate an image blurring, consequently producing a wrong determination of the radioactive distribution. The intrinsic radiation of LYSO scintillators might also generate undesired random events [30]. Therefore, in order to optimize the quality of the image, it is critical to estimate the percentage of all these losses as a function of the imaged activity. To study the different contributions, we have used a phantom made out of high-density polyethylene with 170 mm length and 60 mm in diameter (see Fig. 7c). It was placed at the scanner center FOV. The phantom has a drilled hole of 3.2 mm in diameter and at 13 mm off-radial direction [24]. Through this hole, a silicone tube (170 mm long) with 1 mm (inner) and 3 mm (outer) diameters was inserted and filled with FDG. The initial activity was 5.45 mCi. We carried out sequential acquisitions every 10 min, during a total time of 18 h.
For each acquisition, a sinogram was generated, where the true, scatter, and random counts are estimated following the NEMA NU 4-2008 protocol [31]. For validation purposes, similar data was obtained using GATE v7.2 simulations.
Case example: patient study
Finally, we have tested the ProsPET system with a patient that was diagnosed with PCa. The patient was injected with 7.5 mCi of a 18F-choline solution. About 90 min post-injection, the patient was scanned in the Philips Gemini-TF (PET/CT) where each bed position lasted 90 s. The CT image was used to determine the distance between the top of the head and the prostate and, therefore, properly positioning the patient within the ProsPET system. The ProsPET acquisition was carried out 30 min after the whole-body PET/CT one. This person signed an informed consent form.
For each of the two axial FOVs we acquired 7 min data. We applied a smoothing post-filter to the final image of 5 mm FWHM transaxial and axial with 3.5 sigmas in the convolution kernel.