Evaluation of cost-effective system designs for long axial field-of-view PET scanners

Abstract

Objective.

Current commercial PET scanners have excellent performance and diagnostic image quality primarily due to improvements in scanner sensitivity and time-of-flight (TOF) resolution. Recent years have seen the development of total-body (TB) PET scanners with longer axial field-of-view (AFOV) that increase sensitivity for single organ imaging, and also image more of the patient in a single bed position thereby enabling multi-organ dynamic imaging. While studies have shown significant capabilities of these systems, cost will be a major factor in their widespread adoption in the clinic. Here we evaluate alternative designs that achieve many advantages of long AFOV PET while utilizing cost-effective detector hardware.

Approach.

We utilize Monte Carlo simulations and clinically relevant lesion detectability metric to study the impact of scintillator type (LSO or BGO), scintillator thickness (10-20 mm), and TOF resolution on resultant image quality in a 72 cm long scanner. Detector TOF resolution was varied based on current scanner performance, as well as expected future performance from detector designs that currently hold most promise for scaling into a scanner.

Main results.

Results indicate that BGO is competitive with LSO (both 20 mm thick) if we assume that it uses TOF (e.g., Cerenkov timing with 450 ps fwhm and Lorentzian distribution) and the LSO scanner has TOF resolution similar to the latest PMT-based scanners (~ 500-650 ps). Alternatively, a system using 10 mm thick LSO with 150 ps TOF resolution can also provide similar performance. Both these alternative systems can provide cost savings (25-33%) relative to a scanner using 20 mm LSO with ~50% of effective sensitivity, but still 500-700% higher than a conventional AFOV scanner.

Significance.

Our results have relevance to the development of long AFOV PET, where reduced cost of these alternative designs can provide wider accessibility for use in situations requiring imaging of multiple organs simultaneously.

1. Introduction

The last several years have seen large changes in the design of PET scanners with a major focus being on extending the scanner axial FOV. Standard commercial whole-body (WB) PET systems using Lu-based scintillators (L(Y)SO) are available in axial range of 20-30 cm and have system TOF resolution of the order of 214-385 ps (Hsu et al., 2017, Pan et al., 2019, van Sluis et al., 2019, Chen et al., 2020, Li et al., 2020), leading to very high performance in terms of intrinsic and effective sensitivity (boosted due to TOF gain), as well as image quality. In parallel, there has been a push towards Total-Body (TB) PET with long AFOV systems (> 60 cm) after the advent of the commercial United Imaging uExplorer (Spencer et al., 2020) and Siemens Biograph Vision Quadra (Alberts et al., 2021), as well as the research PennPET Explorer (PPEx) system (Karp et al., 2020). In addition to providing very high sensitivity, these scanners provide the capability to image multiple organs simultaneously, thereby opening new research and clinical areas for dynamic imaging (Badawi et al., 2021, Pantel et al., 2022). While studies have shown significant performance gains with these systems, cost remains prohibitive for most clinical and research sites. Hence, there is potential to develop new long AFOV systems that provide all the advantages of TB-PET imaging at a reasonable cost (Vandenberghe et al., 2022).

Short of any new technical developments in detector design, there remain three areas of detector/system development that can lead to new cost-effective designs of long AFOV scanners: introducing axial or transverse gaps within the scanner, reducing the thickness of L(Y)SO crystal used, and replacing L(Y)SO with BGO.

The use of incomplete detector coverage to create scanners with longer AFOV without increasing the cost has been considered by several groups (Salomon et al., 2011, Feng et al., 2018, Efthimiou et al., 2019, Li et al., 2019). Yamaya, et al., proposed an OpenPET geometry, with separated two rings of detectors and found a loss of axial resolution and image artifacts as the axial gap exceeded the detector width, thereby establishing a practical limit to the fraction of missing lines of response (LORs) (Yamaya et al., 2008, Yamaya et al., 2009). Other studies introduced multiple small gaps (Karakatsanis et al., 2019, Zein et al., 2020) or checkerboard detector designs (Akl et al., 2019, Efthimiou et al., 2019) with up to 50% detector reduction and found comparable quality images to those from fully populated systems, especially with a specialized TOF reconstruction algorithm (Zheng et al., 2019). Potential benefits of TB-PET were also seen in studies performed on the PPEx (Pantel et al., 2020), even though a delay in updating the firmware in the hardware led to initial evaluations being performed with inter-ring gaps where each gap corresponded to 40% of the active ring width. Such pilot studies provide experimental evidence of operating a TOF scanner with inter-ring gaps, without artifacts or variation of quantitative accuracy (Daube-Witherspoon et al., 2020).

In the past, use of thinner (<10 mm) LSO crystals has been investigated to achieve improved imaging performance in a 72 cm long scanner when compared to an 18 cm long scanner using 20 mm thick LSO (Surti et al., 2013). Another study showed that improved sensitivity and noise equivalent count (NEC) rate performance is achieved with a fixed crystal volume when using thinner crystal and a longer scanner AFOV (Yoshida et al., 2013).

Finally, a third option is to re-consider BGO where 20 mm thick BGO provides 30% higher coincidence efficiency relative to 20 mm thick LSO at less cost. Previously, simulations demonstrated good NEC performance in a 1 m long BGO scanner (Zhang and Wong, 2016). Since then, it has also been recognized that BGO can provide TOF capability making using of the Cerenkov photons emitted within the crystal (Brunner et al., 2014, Brunner and Schaart, 2017, Kwon et al., 2016). While it is not clear what a practical implementation of such a system would cost relative to one based on LSO, the goal for this work is a systematic evaluation of the imaging performance of these different design concepts, utilizing our knowledge of the current detector capabilities that are scalable to a full system. For image evaluation the metric we use is a numerically calculated measure of the area under the localized receiver operating characteristic (LROC) curve for detecting and localizing small lesions.

2. Methods

We performed EGS4-based system simulations for a cylindrical PET scanner geometry using pixelated crystals. Scanner ring diameter was fixed at 85 cm while the AFOV was 72 cm. As shown in figure 1(a), scanners with AFOV larger than ~70 cm do not lead to any significant gains in peak (or maximal sensitivity) due to attenuation in the phantom and reduced gains in geometric sensitivity (Surti et al., 2020). Hence for our work here, we will analyse lesions placed only in the central slice with peak sensitivity, since the peak sensitivity of the 72 cm AFOV scanner is representative of all TB-PET scanners, where the peak sensitivity increases only slightly beyond 72-cm (see Fig 1.)

Figure 1:

(a) Calculated plot of geometric sensitivity as a function of axial position in a 35 cm diameter cylindrical phantom filled with water. Results are shown for varying seamier axial length, (b) Reconstructed image of the central transverse slice for the simulated lesion phantom. The distribution of 16, 1-cm diameter lesions at radial distances of 7 and 13 cm can be visualized here.

For the detector we simulated LSO crystals 10 or 20 mm thick and BGO crystals 20 mm thick. Crystal cross-sections were fixed at 4 x 4 mm² with pitch of 4.07 x 4.07 mm². For LSO detectors TOF resolution (fwhm) was varied between 150 ps (a practical target for the near future using scalable electronics especially with 10 mm thick crystals), 240 ps (representative of modern digital PET systems) (Hsu et al., 2017, Pan et al., 2019, van Sluis et al., 2019, Chen et al., 2020, Li et al., 2020, Karp et al., 2020), 500 - 650 ps (representing previous generation non-digital PET systems) (Surti et al., 2007, Jakoby et al., 2011, Bettinardi et al., 2011, Kolthammer et al., 2014) while assuming Gaussian distributions. The BGO scanner was modeled as non-TOF as well as TOF using a Gaussian as well as Lorentzian TOF kernel of 450 ps fwhm (G 450ps and L 450ps, respectively). The Gaussian TOF kernel was used only for comparison purpose since measurements to date with BGO demonstrate Lorentzian timing kernels. Excellent Cerenkov timing performance has been reported (with Lorentzian kernels) (Kwon et al., 2019, Kratochwil et al., 2020), but a Lorentzian TOF kernel with 450 ps fwhm represents measurements performed with realistic BGO crystal sizes (3x3x20 mm³) and currently scalable electronics – a Lorentzian with 560 ps fwhm was measured by Kwon using NUV-HD SiPMs (Kwon et al., 2016), while a Lorentzian with 400 ps fwhm was measured by Brunner using the PDPC array (Brunner and Schaart, 2017). In figure 2 we show Gaussian distribution plots with fwhm values of 240 ps and 450 ps, as well as a Lorentzian distribution with 450 ps fwhm. System energy resolution was set at 11% for LSO and 16% for BGO leading to collected events energy window of 440-650 keV and 400-650 keV respectively.

Figure 2:

Distribution plots for a Gaussian with 240 ps (blue) and 450 ps (red) fwhm values, and a Lorentzian with 450 ps fwhm (black) value. All three functions are normalized to an integral value of one.

We simulated cylindrical lesion and uniform phantoms (35 cm diam. x 80 cm long) and only true (T) and scatter (Sc) coincidences were simulated. Lesion phantom had 16, 1 cm diameter hot lesions (spheres) placed in the central transverse slice with a water background (figure 1(b)). The lesion to background activity uptake ratio was set at 3:1. Eight lesions were distributed at a radial position of 7 cm from the centre, while the other eight were distributed at a radial position of 13 cm from the centre to capture the impact of parallax error. The uniform water-filled phantom had an activity concentration that is the same as the lesion phantom background (0.1 μCi/cc). Five independent data replicates were simulated for each phantom and scanner design. Scan times of 8.5, 17, 34, and 68 seconds were simulated. The list-mode data output from the simulations is reconstructed using spherical blob-basis functions (Matej and Lewitt, 1996) and a list-mode OSEM algorithm with Gaussian TOF kernel, 25 subsets, and normalization, attenuation, and scatter corrections built into the system model (Popescu, 2004). The blob basis functions are on a body-centred cubic grid with an optimized blob spacing of 6 mm (blob radius of 7.5 mm) and the reconstructed blob image is converted into 2 x 2 x 2 mm³ image voxels by summing the magnitude and intensity of each blob at a given image voxel centre. No point spread function (PSF) modelling was performed in the image reconstruction for this study. Attenuation images were produced analytically. Scatter was estimated using a TOF-extended 3D single scatter simulation correction (Werner et al., 2006). The TOF kernel used for scatter estimation was modelled either as a Lorentzian or Gaussian distribution. Normalization data were generated by performing uniform phantom simulations (40 cm diameter x 80 cm long) with a very high number of coincident events (an average of ~ 5 counts per line of response). All data were used for image reconstruction without setting an upper limit for the maximum ring difference (MRD).

Lesion detectability was measured using the generalized scan statistics methodology (Popescu and Lewitt, 2006) to first estimate the signal (lesion contrast) and noise (background nodule contrast) probability density functions (pdfs), followed by calculation of the LROC curve (Swensson, 1996) from first principles. The area under the LROC curve (ALROC) was subsequently used as our metric for scanner performance evaluation. The error in the ALROC value was determined as the standard deviation of the results over 100 bootstrap data sets. The ALROC value also changes as a function of the number of iterations of the reconstruction algorithm and scanner design, and for our work we use the maximal ALROC value for evaluation.

3. Results

Table 1 summarizes the basic statistics from the lesion phantom simulations for sensitivity and scatter fraction (SF=T/(T+Sc), where T are the true or unscattered coincidences and Sc are the scattered coincidence), as well as the relative noise equivalent counts (NEC*) (NEC*=TxT/(T+Sc)) where the impact of random coincidences is not included). Worse energy resolution of BGO translates into a higher SF, and the resultant gain in T sensitivity relative to 20 mm thick LSO is 15%, while the NEC’ is ~ 10% lower. For comparison purposes, a scanner with axial gaps modelled on the earlier configuration of the PPEx scanner (each gap corresponding to 40% of the active ring width) will have a 50% reduction in intrinsic sensitivity leading to a relative NEC* of 0.18.

Table 1:

Basic imaging statistics from phantom simulations performed with the 3 detector designs.

	Crystal thickness (mm)	SF (%)	Rel. (T+Sc) sensitivity	Rel. T sensitivity	Rel. NEC*
LSO	20	40	1.00	0.60	0.36
LSO	10	40	0.39	0.23	0.14
BGO	20	53	1.46	0.69	0.32

3.1. Performance of scanner using 20 mm thick LSO crystals

In figure 3 we show central transverse slices from representative reconstructed images of the lesion phantom simulated in a scanner using 20 mm thick LSO crystals. These images and corresponding ALROC values indicate that 17 secs. image from a scanner with 240 ps TOF resolution is similar to the 34 secs. image from a scanner with 500 ps TOF resolution – consistent with the effective sensitivity gain due to improved TOF performance.

Figure 3:

Central transverse slices from reconstructed images for a scanner using 20 mm thick LSO crystals. Results are shown for two scan times (17 and 34 secs.) in scanners with 500 ps (G 500ps) and 240 ps (G 240ps) TOF resolution. Below each image is the ALROC value with its estimated error or standard deviation.

3.2. Performance of scanner using 10 mm thick LSO crystals

Figure 4 shows central transverse slices from selected reconstructed images from a scanner using 10 mm thick LSO crystals. These results show the expected relative gain in performance with the improved 150 ps TOF resolution.

Figure 4:

Central transverse slices from reconstructed images for a scanner using 10 mm thick LSO crystals. Results are shown for three scan times (17, 34, and 68 secs.) in scanners with 240 ps (G 240ps) and 150 ps (G 150ps) TOF resolution. Below each image is the ALROC value with its estimated error or standard deviation.

3.3. Performance of scanner using 20 mm thick BGO crystals

Figure 5 shows central transverse slices from reconstructed images from a scanner using 20 mm thick BGO crystal with varying TOF capabilities. These results show that a scanner with 450 ps Lorentzian TOF distribution has a factor two effective gain in sensitivity compared to non-TOF scanners (68 secs non-TOF is similar to 34 secs L 450ps), while the 450 ps Gaussian TOF distribution has another factor of two gain (17 secs G 450ps is similar to 34 secs L 450ps).

Figure 5:

Central transverse slices from reconstructed images for a scanner using 20 mm thick BGO crystals. Results are shown for three scan times (17, 34, and 68 secs.) in Non-TOF and TOF scanners with 450 ps Lorentzian (L 450ps) and 450 ps Gaussian (G 450ps) TOF resolution. Below each image is the ALROC value with its estimated error or standard deviation.

3.4. Comparison of imaging performance of scanners using 10 mm and 20 mm thick LSO

Figure 6 shows a plot of ALROC vs scan time for scanners using 10 mm or 20 mm thick LSO crystals. As expected, improved TOF resolution leads to improved ALROC results. In addition, a scanner using 10 mm thick LSO crystals needs improved TOF capability (150 ps resolution) to overcome reduced intrinsic sensitivity of shorter crystals.

Figure 6:

ALROC vs scan time in scanners using 20 mm and 10 mm thick LSO (LSO and 10mmLSO, respectively). Results are shown for non-TOF and TOF scanners with varying TOF resolution (Gaussian distribution).

3.5. Comparison of imaging performance of 20 mm thick LSO vs 20 mm thick BGO scanners

Figure 7 shows a plot of ALROC vs scan time for scanners using 20 mm thick LSO and BGO crystals. Non-TOF BGO scanners perform significantly worse and have 4 times worse effective sensitivity relative to LSO scanners with 240 ps TOF resolution (benchmark commercial LSO scanners). With 450 ps Lorentzian TOF resolution, the BGO scanners will be capable of performing as well as LSO scanners with 500-650 ps TOF resolution.

Figure 7:

ALROC vs scan time in scanners using 20 mm thick LSO and BGO. Results are shown for non-TOF and TOF scanners with varying TOF resolution (Gaussian distribution labelled with G, Lorentzian distribution labelled with L).

3.6. Comparison of imaging performance of scanners using 10 mm thick LSO and 20 mm thick BGO crystals

Figure 8 plots ALROC vs scan time for scanners using 10 mm thick LSO and 20 mm thick BGO crystals. Once again non-TOF BGO scanner performs significantly worse overall, while a BGO scanner with 450 ps Lorentzian TOF resolution is similar to a scanner using 10 mm thick LSO with 150 ps Gaussian TOF resolution.

Figure 8:

ALROC vs scan time in scanners using 10 mm thick LSO and 20 mm thick BGO. Results are shown for non-TOF and TOF scanners with varying TOF resolution (Gaussian distribution labelled with G, Lorentzian distribution labelled with L).

4. Discussion

Our results show that the relative performance (in terms of ALROC) of a scanner using 10 mm thick LSO crystals is similar to a scanner using 20 mm thick LSO but with a factor of 2-3 increase in scan time. This is consistent with the relative increase in NEC* (x2.6) with 20 mm thick LSO as shown in Table 1. A scanner utilizing 10 mm thick LSO crystals and achieving a TOF resolution of 150 ps performs similar to a scanner with 20 mm thick LSO and TOF resolution of 500-650 ps, which represents performance of the last generation of PMT-based TOF scanners (non-digital) developed between 2006-2015.

Comparing 20 mm thick LSO with 20 mm thick BGO our results clearly show reduced performance of BGO without any TOF capability. However, with potential for achieving TOF capability in BGO detectors when using the Cerenkov light mechanism, some of these performance differences can be reduced or eliminated. However, in current practical detector designs the TOF capability (450 ps fwhm) instead of having a Gaussian distribution has longer tails leading to Lorentzian distribution (Brunner and Schaart, 2017). Our results clearly show that a Lorentzian TOF distribution with 450 ps fwhm is worse than a 450 ps fwhm Gaussian distribution. When compared to a detector using 20 mm thick LSO crystals, we see that the BGO detector with this TOF capability is similar (90%) in performance to the last generation of non-digital TOF scanners using LSO. In fact, this performance is very similar to that achieved in detectors using 10 mm thick LSO with 150 ps TOF resolution.

In Figure 9 we summarize relative sensitivity and costs for fixed AFOV scanner designs: a standard PET scanner design that uses 20 mm thick LSO crystals, a standard PET scanner with axial gaps corresponding to 40% of the active ring width, a PET scanner using 10 mm thick LSO crystals, and a PET scanner using 20 mm thick BGO crystals. Developing scanners with gaps (or more generally, incomplete detector coverage) is one way to reduce cost of long AFOV PET systems (Vandenberghe et al., 2023). For the calculations, the scanner with axial gaps is modelled on the PPEx scanner where initial DAQ limitations restricted full scanner ring readout, but did not have an impact on image quality or quantitative accuracy beyond that of reduced intrinsic sensitivity (50%). In this figure we assume that the cost of a BGO is 1/3 of LSO for the same crystal volume, which only considers the reduced cost of the scintillator, but not the cost of manufacturing crystal arrays which should be similar. We also assume that the cost of photosensor (SiPM) and electronics is the same, and their combined cost is the same as the cost of 20 mm LSO (or LYSO) in a fixed scanner AFOV. In order to achieve TOF capability with BGO using Cerenkov light we may need more specialized electronics at a higher cost, that may further reduce any advantages of using BGO in a long AFOV system. From our past experience, the additional cost for mechanical assembly will be ~15% of the total scintillator, SiPM, and electronics cost, plus any additional cost related to acquisition of a CT and patient bed. From this figure it is clear that non-TOF BGO scanner is not competitive based on its low effective sensitivity per unit cost. The three alternative TOF scanner designs (20 mm thick LSO with axial gaps, 10 mm thick LSO, and BGO with Lorentzian TOF) will also have lower effective sensitivity for fixed total cost, but they can be considered practical alternatives when designing a fixed length, long AFOV (≥ 70 cm) PET scanner where system cost is likely the driving factor. However, in designing shorter AFOV (25-30 cm) systems similar to the traditional whole-body PET scanners that use with multiple bed positions to scan a whole patient, there will still be a need for high effective sensitivity. Hence, these alternative designs will lead to much higher cost in achieving similar effective sensitivity to that of a traditional design (20 mm LSO with 240 ps TOF resolution).

Figure 9:

Plot of relative cost and effective sensitivity of different scanner designs with a fixed AFOV. Relative cost only includes an estimate of the scintillator and SiPM, but ignores electronics, and other manufacturing or mechanical costs.

We note that our study does not include the effect of random coincidences and detector deadtime. Due to the long decay time of BGO scintillation signal, the deadtime in BGO detectors will be higher. However, in modern detector designs utilizing small SiPMs this effect is expected to be minor in normal clinical range. Relative random coincidences will, however, be higher in the BGO detector since the long tails in the Lorentzian TOF distribution will still require a much wider coincidence timing window relative to an LSO based scanner (about a factor of 3-4 higher based on the use of a 5-6 ns coincidence timing window for LSO and up to 20 ns for BGO). Since this study is not intended to investigate high count rate scenarios, we believe this assumption is warranted, in order to simplify the study. However, our work clearly shows that the different LSO systems evaluated here performed similar or better than the BGO scanners. Hence, the overall imaging performance and the effective sensitivity as estimated here for BGO represents the best-case scenario.

In our work we use 150 ps TOF resolution as a target for the 10 mm thick LSO crystals. The commercially available Siemens Vision scanner currently achieves 214 ps TOF resolution using 20 mm thick LSO (van Sluis et al., 2019). With thinner crystals and ongoing improvements in SiPMs, it is conceivable that 150 ps TOF resolution with 10 mm thick LSO is practical to achieve in the near future. On the other hand, while recent work has shown an ability to achieve improved TOF performance with BGO (better than the 450 ps Lorentzian modelled here) (Kwon et al., 2019, Kratochwil et al., 2020), achieving this performance in a large PET system with thousands of channels is currently not practical.

5. Conclusion

Clinical task-based metrics were used to evaluate imaging performance of two alternative, cost-effective detector designs, one using 10 mm thick LSO and the other using 20 mm thick BGO. Our results indicate the 10 mm thick LSO crystal with 150 ps TOF resolution will perform similar to the last generation of non-digital PET using 20 mm thick Lu-based crystals an achieving 500-650 ps TOF resolution. BGO without TOF capability has significantly worse performance. However, if TOF capability based on Cerenkov timing as demonstrated in a practically scalable detector design is achieved (Lorentzian TOF distribution with 450 ps fwhm), then such a detector will be competitive in performance to the 10 mm thick LSO detector. The cost of these two systems is also expected to be similar, assuming similar costs for electronics and manufacturing. The impact of higher random coincidences in the BGO scanner relative to an LSO scanner (due to a wider coincidence timing window) was however not included in this work, and is expected to further degrade its performance at activities likely to be encountered in typical imaging scenarios.