All measurements reported in this manuscript have been taken with SAFIR-I. The PET insert was permanently installed inside the constant magnetic field of the MRI system during all tests. No Magnetic Resonance (MR) coils were used concurrently and no simultaneous MR acquisitions took place to preclude excessive experimental complexity. Full MR-compatibility of SAFIR-I has previously been shown [1]. A well counter dose calibrator (MedisystemFootnote 1 Medi 405) was used to determine the activities of the employed sources.

The data collection, processing, and analysis followed the NEMA-NU4 protocol (for convenience simply denoted “the protocol” henceforth) [9], with minimal adaptations detailed in subsections – below. For the coincidence search, a CTW of \(500\,\hbox \) and an energy window of 391 to \(601\,\hbox \) [1, 4] were applied in all cases. All reconstructions were performed using methods implemented in Software for Tomographic Image Reconstruction (STIR) [10], with a voxel size of \(0.55\,\hbox \times 0.55\,\hbox \times 1.1\,\hbox \).

The performance reports for each test were again based on the protocol’s instruction. Finally, the achieved performance results were compared to the performance parameters reported for a set of reference scanners featuring similarly sized crystals, including the SAFIR prototype system, as tabulated in [3].

A \(^\)Na point source embedded in a \(1-\hbox ^\) acrylic cube (Eckert & Ziegler Isotope Products, High-Resolution Marker “NEMA,” MMS09-022, source diameter \(0.25\,\hbox \)) was used. Its activity was \(316.3\,\hbox \), a level at which dead time losses are negligible for SAFIR-I. As demanded by the protocol, the axial source positions were \(0\,\hbox \) (axial center) and \(13.5\,\hbox \) (one quarter of the axial FOV). Supplementary to the mandatory offset evaluations, measurements at \(0\,\hbox \) radial offset were taken. At each source location, \(6 \times 10^\,\hbox \) were collected to satisfy the requirement of acquiring at least \(10^5\) coincidences. Instead of the animal bed on a single-direction rail, a point source holder facilitating point source positioning in two directions was used, as shown in Fig. 3.

Point source holder for SAFIR-I. The radius of the white alignment plates corresponds to the inner radius of the detector; the 3D-printed tip (black) accepts \(1-\hbox ^\) cubic sources (\(^\)Na) for detector calibration and NEMA-NU4 measurements. Design by R. Becker for the SAFIR collaboration

The data analysis was performed using STIR’s FBP3DRP reconstruction algorithm without any filtering or smoothing of the data (as stipulated). As a back-projection filter the STIR-default Colsher filter (\(\alpha = 1\), cut-off 0.5 cycles in both axial and planar directions) was used. A known downside of this algorithm is its requirement for cylindrical projection data; for a detector with block geometry like SAFIR-I, direct interpolation of the data to fit the cylindrical geometry leads to image degradation with severe streak artifacts [3, 5, 11, 12]. In order to partially ameliorate the degradation, the raw data from SAFIR-I were first sorted into projection data using its true, generic block geometry and then rebinned into a cylindrical projection [5, 11, cf.], prior to reconstruction.

As SAFIR-I can fit mice and rats, scatter phantoms for those two use-cases were obtained (QRM Micro-PET Scatter Phantom mouse size / rat size, see Fig. 4).

The NECR phantoms used. Top: Mouse-like. Bottom: Rat-like. Standard Luer lock connectors and plugs on both ends enabled to fill and seal the active volume

For measurements of the background due to intrinsic radioactivity of SAFIR-I’s LYSO crystals, \(42\,\hbox \) and \(48\,\hbox \) of continuous data could be acquired, for the mouse-like and rat-like phantom, respectively.

For the acquisition of count rate data, both phantoms were filled with a solution of \(^\)F in water. In the case of the mouse-like phantom the liquid volume was \(0.225\,\hbox \); in the case of the rat-like phantom it was \(0.450\,\hbox \).

In order to get a good resolution on any count rate peak potentially presenting at high activities, 16 measurements were taken at shorter \(600\,\hbox \) intervals, before the interval time was extended to \(3200\,\hbox \) for the remaining 18 measurements. The acquisition time for the first measurement was \(1\,\hbox \) and for subsequent acquisitions it was exponentially increased to the nearest full second to compensate for the decay. For these experiments, the start and end activities for the mouse-like phantom were \(506.1\,\hbox \) and \(580.2\,\hbox \); for the rat-like phantom they were \(506.1\,\hbox \) and \(564.3\,\hbox \), respectively.

The scatter fraction was determined for the last data point in each case. The protocol’s instruction was followed in all other points.

The same \(^\)Na point source as in subsection was used, at an activity of \(313.6\,\hbox \). Per measurement point, \(8 \times 10^\) coincidences were acquired in order to obtain a smoother axial sensitivity curve [3, cf.].

Furthermore, the point source holder (see Fig. 3) only allowed for a step width of an estimated \((1.0 \pm 0.1)\,\hbox \), as opposed to the step size of one slice thickness (i.e., \(1.1\,\hbox \)) which the protocol asked for. Hence, the axial FOV was covered in 55 instead of 49 steps. To account for the extra steps when going in either direction from the axial center in the sum calculation of the total system sensitivity, all sensitivities but the central value were weighted with a factor of 48/54. Furthermore, considering the uncertainty associated with the mechanical setting of the source position, the positions were determined from the data.

In all other aspects, the steps outlined by the protocol were followed.

An IQ phantom was obtained (QRM Micro-PET IQ Phantom, see Fig. 5) and its active volume filled with a solution of \(^\)F in water. Due to a temporary issue with the well counter, the start activity was \(3.35\,\hbox \), i.e., marginally below the protocol-requested \(3.52\,\hbox \). The acquisition time was \(20\,\hbox \). The image data were random, attenuation, scatter and normalization corrected [5, cf.]; the Singles-Prompts (SP) method [13] was used for the estimation of randoms and the Single Scatter Simulation (SSS) method [14, 15] was used to estimate scatter (see [5] for details). The attenuation maps were generated manually using STIR’s generate_image utility in conjunction with reconstructed measurement frames for reference, and including the known attenuation properties of the phantom materials. This approach is similar to the segmentation method [16]. A quantitative calibration of voxel values was achieved by imaging a cylindrical phantom of comparable dimensions to the IQ phantom but without any internal structures, hence larger active volume. Accordingly, the start activity was increased to \(6.66\,\hbox \) for the calibration phantom measurement in order to achieve a similar average activity concentration for an identical acquisition time of \(20\,\hbox \).

The IQ phantom used. Markers for rapid positioning and alignment were added on orange tape. Yellow tape (left) was used to fix the phantom in place on the animal bed

The images were reconstructed using a Maximum-Likelihood Expectation- Maximization (MLEM) algorithm with 30 iterations and a Gaussian inter-update filter with FWHM of \(1.1\,\hbox \times 1.1\,\hbox \times 2.2\,\hbox \), i.e., twice the edge lengths of the reconstructed voxels, was applied [3, cf.]. The voxel value calibration factor CF was determined by inspecting a centered cylindrical Volume Of Interest (VOI) inside the calibration phantom with a radius of \(75\,\%\) the phantom’s inner radius to prevent edge effects. There, the mean reconstructed number of counts was compared with the expected number of decays leading to the emission of \(511\,\hbox \) photons based on the known activity, i.e.,

$$\begin \mathrm} = \frac}}. \end$$

(1)

An \(^\)F-to-\(\beta ^\) branching ratio of \(96.86\,\%\) [17] was used for the calculation. This factor was subsequently applied to the IQ phantom data, leading to a calibrated number of reconstructed counts (denoted “calibrated counts” hereafter). To test the calibration, cylindrical VOI was drawn inside the central uniform region of the IQ phantom (again centered on the phantom’s axis and with a radius of \(75\,\%\) the phantom’s inner radius). The accuracy A (in percent) of the calibration could then be estimated by comparing the mean of the calibrated counts in the VOI with the expected number of decays in the same volume, according to [3, 5, cf.]

$$\begin A = \frac~-~\text }}~\times ~100\,\%. \end$$

(2)

The procedures described in the protocol were followed in all remaining points, including the definitions of Regions Of Interest (ROIs)/VOIs, and the determinations of Spill-Over Ratio (SOR) and Recovery Coefficient (RC) values.