Performance evaluation of a high-resolution brain PET scanner using four-layer MPPC DOI detectors

A four-layer DOI detector module comprising four detector layer boards including detectors, front-end and signal processing circuits, a data interface circuit board for transferring single-event data to a data acquisition server, and DC power supplies is shown in figure 1. The detector layer board has five detector units, each of which consists of a 32  ×  32 LYSO scintillator array with a 1.2 mm pitch and 64 MPPCs (Hamamatsu S10931-050P: the active area is 3 mm  ×  3 mm, with 3600 pixels per 50 µm, Hamamatsu Photonics K.K., Japan) in an 8  ×  8 arrangement. To fabricate a 32  ×  32 scintillator array with a 1.2 mm pitch efficiently and precisely, we adopted an internally focused laser processing technique, termed sub-surface laser engraving (SSLE). SSLE produces micro-cracks that act as optical reflectors of scintillation photons inside a monolithic scintillator plate (Moriya et al 2010). A 2D 32  ×  32 segmented array with a 1.2 mm pitch was fabricated to a monolithic LYSO scintillator plate with a 38.4 mm square cross-section. The scintillator thickness was designed to be 3 mm, 4 mm, 5 mm, and 8 mm toward the bottom layer, to ensure that the load for the count rate in every layer is equally sensitive for γ-rays, and the total scintillator thickness was 20 mm. The overall thickness for the four-layer DOI detector was 45.0 mm, including MPPCs, printed circuit boards, reflective and light shielding sheets, etc. The LYSO arrays were optically coupled to the MPPCs by a transparent silicon rubber compound (Momentive RTV615, Momentive Performance Materials, Inc., USA). The DOI detector modules were cooled using air cooling fans installed in a detector ring gantry head. Figure 2 shows a schematic drawing of a cross-section through the four-layer DOI detector.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Each detector layer board has independent front-end and signal processing circuits. In the front-end circuits, the 64-channel MPPC output signals of each detector unit are amplified and then reduced to four signals by a resistor network, and the sum signal of all 64-channel signals is generated. Each gain of the MPPCs is finely adjusted to provide the optimum bias voltage for each MPPC. In addition, a temperature sensor is located in the front of the detector layer board to compensate for temperature changes occurring with the gain of MPPCs according to the bias voltage. When a valid event is detected at one of the five units on the detector layer board, the analog signals of the unit are selected through 5:1 multiplexers in the signal processing circuits. The position signals for Anger logic calculation and the sum signal for both timing pick-off and energy discrimination are fed into the shapers and are then digitized by 10-bit analog-to-digital converters (ADCs). These digital signals are fed into a field-programmable gate array (FPGA), and signals are processed to detect γ-rays. To analyze crystal identification and energy window discrimination, look-up-tables (LUTs) were used. Additionally, timestamp data with 781.25 ps resolution were generated using a time-to-digital converter (TDC) at the moment of triggering by the timing pick-off signals. Each digital data set analyzed in the four detector layer boards was collected with a data-interface circuit board placed behind the detector module, added to DOI information and timer tag data at the indicated time interval from the scan start, and was finally transferred to a data acquisition server as 64 bits list-mode data through an optical fiber cable. Note that all single event data from each individual four-layer detector, including Compton scatter events generated between DOI crystal layers, were transferred to the data acquisition server, to allow accurate analysis of γ-ray interaction positions in post-processing.

For a typical detector unit in the first layer, the average energy resolution was 24.0%, and the coincidence timing resolution (CTR) was 850 ps FWHM with a reference detector composed of a 25 mm long truncated pyramid of BaF₂ crystal (Ohyo-Koken-kagaku, Japan) and a fast photomultiplier tube (Hamamatsu H5321, Hamamatsu Photonics K.K., Japan). Further details have been described previously by Omura et al (2012).

The detector ring consists of 32 DOI detector modules arranged in a shape with a 430 mm diameter and 168 detector rings with a 1.2 mm pitch. The 168 detector rings include the number of gaps equivalent to eight crystal segments, where there is a gap of 2.4 mm, equivalent to two crystal segments, between crystal arrays. The total number of crystals is 655 360. The transaxial FOV is 330 mm in diameter and the axial FOV is 201.6 mm, which are sufficient to cover a whole human brain. The major characteristics of the Hamamatsu brain PET scanner are listed in table 1. The gantry head, including the DOI detector modules, has a compact cylindrical shape with outer dimensions of approximately 740 mm in diameter and 770 mm in height. The detector head is able to move up and down, as well as forward and backward, and to tilt. Thus, the patient can take various positions, such as sitting, lying, and standing. Photographs of the gantry and the inner detector ring are shown in figure 3.

Zoom In Zoom Out Reset image size

The data acquisition server for single events consists of eight data transfer circuit boards based on a serially attached small computer system interface (SAS) and eight common server computers (IBM BladeCenter H, IBM, USA) operating on Linux. The single-event data 64-bit lists from 32 DOI detector modules are fed to the data transfer circuit boards, stored once in buffer memories, and then transferred to the server computers. Each data transfer circuit board handles the single-event data from four DOI detector modules, and transfers these data to the respective corresponding server computers. Single-event data include the crystal segment and DOI addresses, the energy and timing information (i.e. timestamp data representing the elapsed time from the start of a scan, with 781.25 ps resolution), and a valid flag indicating 'acceptable: within the range of the set energy window' or 'not acceptable out of the range of the set energy window'. In addition, timer tag data are inserted at every 0.8192 ms in the single-event data stream.

On-the-fly coincidence detections are performed by software after each single-event data set is sorted in time order, in units of 0.8192 ms, and corrected for the timing delay of each detector unit or crystal segment (Yu et al 2016). To be able to take a line of response (LOR) in the FOV, sets of two single events are analyzed within a coincidence time window of 4.3 ns, and a coincidence event is created if coincidence is detected. Random coincidence events are also created, using a delayed coincidence method. The coincidence data generated in this manner, as well as all single-event data, if necessary, are stored. Further details have been described previously by Isobe et al (2012).

A list-mode dynamic RAMLA (LM-DRAMA) is implemented as the default mode for image reconstruction (Tanaka et al 2003, 2010, Nakayama et al 2005). For image data corrections, normalization, attenuation, and scatter corrections are performed. In normalization, the component-based normalization (CBN) method is used (Badawi et al 1999). The system has no transmission measurement mechanics. Thus, attenuation correction is performed using emission segmented attenuation correction (E-SAC), which involves a segmented attenuation map generated from emission data (Zaidi et al 2003). The scatter correction uses the single scatter simulation (SSS) method (Watson 2000). Random correction is performed by subtracting delayed coincidence events (Rahmim et al 2004). The dead-time loss is corrected based on an empirical relation between the total single count rate and the true coincidence count rate.

There are two reconstruction modes: a high-resolution (HR) mode for maximizing the spatial resolution performance of the system and a high-speed (HS) mode for reducing the reconstruction time. The HR-mode has an image matrix size of 512  ×  512 with a 0.65 mm pixel size and 335 slices of 0.6 mm width. The HS-mode has an image matrix size of 128  ×  128 with a 2.6 mm pixel size and 83 slices of 2.4 mm width. The maximum ring differences are 167 and 40, in a unit of 1.2 and 4.8 mm for the HR and HS-modes, respectively. A filtered back-projection (FBP) algorithm for image reconstruction is also implemented to evaluate the performance in spatial resolution. We used server computers as image reconstruction engines, as well as data acquisition servers.

To evaluate the intrinsic spatial resolution of the DOI detector module, coincidence response functions (CRFs) of the detector pair located at a 40 mm distance were measured. This was done by scanning a 0.25 mm diameter ²²Na point source of approximately 0.7 MBq along the central line between them, at γ-ray incident angles of 0° and 22.5° to the crystal surface, in 0.12 and 0.185 mm scan steps, respectively. Each angle corresponds to the center and half positions of the transaxial FOV in the system ring configuration. The intrinsic spatial resolution measurement setup is shown in figure 4. The single-event data of the detector pair at each step position were sent to the data acquisition servers and stored in a hard disk. The acquiring times for each position were 20 and 10 min for 0° and 22.5°, respectively. After the measurements, coincidence detections were performed off-line by software and the coincidence counts were plotted as a function of the source position for each layer combination. The intrinsic spatial resolutions of the detector pair were derived from the FWHM values of the CRFs, by fitting a Gaussian function.

Zoom In Zoom Out Reset image size

The spatial resolution was assessed according to NEMA NU 4-2008 for small animal PET. A 0.25 mm diameter ²²Na point source (MMS09-022: EZIP, USA) of approximately 0.4 MBq was placed at distances of 10, 25, 50, 75, 100, 125, and 150 mm from the center of the transaxial FOV with both horizontal and vertical lines, where three axial positions with the center, 1/4, and 3/8 of the axial FOV from the center of the FOV were measured. The measurement time for each position was 5 min, and more than 5 M counts were obtained. The image reconstruction was performed both in 2D FBP and LM-DRAMA in the HR-mode. The voxel sizes were 0.3  ×  0.3  ×  0.3 mm³ in both 2D FBP and LM-DRAMA in the HR-mode, respectively. In LM-DRAMA, the iteration number was four, and an image-based point-spread function (PSF) (Reader et al 2002) of 1.8 mm in all directions was used. To obtain FWHM values, profiles of point source images were interpolated and analyzed according to NEMA NU 4-2008.

The scatter fraction (SF) and count rate capability were measured using a scatter phantom that had a 20 cm diameter and 70 cm long polyethylene cylinder with a hole of 6.4 mm diameter drilled at 45 mm off-center in the plane (Data Spectrum Corporation, USA). The phantom was located at the center of both the transaxial and axial FOVs in the scanner. The initial activity of the ¹⁸F line source was 423 MBq. The scan continued over 20 h, which covered a time span of 11 half-lives, which was sufficiently low for random events and dead-time losses to be negligible. The data were acquired as single events, and on-the-fly coincidence detection was carried out during data acquisition. All coincidence data were stored on hard drives. Sinograms were generated for each 10 min acquisition frame and for each slice, where the data were separated into prompt and random data. The DOI pair sinograms of 4  ×  4 combinations were gathered into a single pair, and the oblique sinograms were collapsed into a single sinogram by single-slice rebinning (Daube-Witherspoon et al 1987).

According to NEMA NU 2-2012, all frames of the measurement were analyzed into true coincidence (T) and random coincidence (R) at the count rate. The SF was also analyzed using the final acquisition data with negligible random events. Intrinsic background events due to ¹⁷⁶Lu in the LYSO scintillator were also measured by setting the scatter phantom without activities for 16 h and using this for background corrections (Watson et al 2004). To evaluate the noise equivalent count rate (NECR), NECR-1R and 2R were calculated, although random correction for the system is practically based on direct random subtraction by delayed coincidence.

The sensitivity of the system was evaluated according to NEMA NU 2-2012. An ¹⁸F line source of 3 mm inner diameter and 70 cm length (Data Spectrum Corporation, USA) was aligned with the center of the transaxial FOV, and was also aligned with a 10 cm offset from the transaxial FOV as another measurement. The line source was covered by five aluminum sleeves with increasing wall thickness (1.25 mm increments), and measurements of 10 min or 12 min were repeated for each additional sleeve. The initial activity of the ¹⁸F line source in the FOV was 4.09 MBq, where the random and the dead time-losses were negligible. The acquired data were corrected according to ¹⁸F decay and ¹⁷⁶Lu background counts estimated from 24 h measurement data without activity. The absolute sensitivity was obtained by extrapolation of the linear regression that fitted the logarithm of the count rates as a function of sleeve thickness. Sensitivity profiles for the HR and HS-modes were plotted by analyzing the data slice-by-slice.

To demonstrate the imaging performance, phantom studies were performed. For evaluating the spatial resolution capability in the reconstruction image, an ultra-micro hot-spot phantom (Data Spectrum Corporation, USA), placed at the center of the FOV, was measured in a 60 min scan. The initial activity was 5.0 MBq of ¹⁸F. The images were reconstructed by the LM-DRAMA in the HR-mode, which set the reconstruction parameters with 50 iterations, 40 subsets, and the voxel size in image matrix of 0.2  ×  0.2  ×  0.6 mm³. No attenuation and scatter corrections were employed. A 3D Hoffman brain phantom (Data Spectrum Corporation, USA), placed at the center of the FOV, was also measured with 14.1 MBq of ¹⁸F at the scan start, when the data were collected for 180 min. The images were reconstructed using the LM-DRAMA with the HR-mode, similar to those with default reconstruction parameters with four iterations. Using the values calculated under the assumption that the phantom was filled with water, attenuation and scatter corrections were employed, but no post-smoothing was applied to the images.

Clinical and pre-clinical studies were performed at Hamamatsu University School of Medicine. All animal and human experiments were approved by the Ethics Committees of Hamamatsu University School of Medicine. In the clinical study, a healthy volunteer was scanned for 62 min after the injection of 246.4 MBq of ¹¹C-MeQAA, a tracer for α7 nicotinic acetylcholine receptors. The images were reconstructed using the LM-DRAMA in both HR and HS-modes with four iterations. Attenuation with E-SAC and scatter corrections were employed, and images were processed with and without application of post-smoothing. As pre-clinical studies, ¹⁸F-FDG rat and mouse imaging studies were performed. A SD rat (8 weeks, 270.7 g) was scanned for 15 min after 45 min post-injection of 34.7 MBq of ¹⁸F-FDG, and a ddY mouse (7 weeks, 27.6 g) was also scanned for 60 min after the injection of 44.0 MBq of ¹⁸F-FDG. These animals were anesthetized by intravenous injection of chloral hydrate, and were then placed at the center of the FOV. The images were reconstructed using the LM-DRAMA in the HR-mode with four iterations, and the voxel size in the image matrix was 0.2  ×  0.2  ×  0.6 mm³. Attenuation with E-SAC and scatter corrections were employed, but no post-smoothing was applied to the images.