Skip to main content

Quantitative SPECT (QSPECT) at high count rates with contemporary SPECT/CT systems



Accurate QSPECT is crucial in dosimetry-based, personalized radiopharmaceutical therapy with 177Lu and other radionuclides. We compared the quantitative performance of three NaI(Tl)-crystal SPECT/CT systems equipped with low-energy high-resolution collimators from two vendors (Siemens Symbia T6; GE Discovery 670 and NM/CT 870 DR).


Using up to 14 GBq of 99mTc in planar mode, we determined the calibration factor and dead-time constant under the assumption that these systems have a paralyzable behaviour. We monitored their response when one or both detectors were activated. QSPECT capability was validated by SPECT/CT imaging of a customized NEMA phantom containing up to 17 GBq of 99mTc. Acquisitions were reconstructed with a third-party ordered subset expectation maximization algorithm.


The Siemens system had a higher calibration factor (100.0 cps/MBq) and a lower dead-time constant (0.49 μs) than those from GE (75.4–87.5 cps/MBq; 1.74 μs). Activities of up to 3.3 vs. 2.3–2.7 GBq, respectively, were quantifiable by QSPECT before the observed count rate plateaued or decreased. When used in single-detector mode, the QSPECT capability of the former system increased to 5.1 GBq, whereas that of the latter two systems remained independent of the detectors activation mode.


Despite similar hardware, SPECT/CT systems’ response can significantly differ at high count rate, which impacts their QSPECT capability in a post-therapeutic setting.


Quantitative SPECT (QSPECT) is increasingly used in the clinics with both diagnostic and therapeutic radionuclides. It is of particular utility in dosimetry-driven protocols of personalized radiopharmaceutical therapy, in which the administered activity can be highly variable [1,2,3,4]. We have previously shown the importance of dead-time correction in a 177Lu post-therapeutic setting [5]. However, using 177Lu and 99mTc, we previously observed divergent behaviours between the two detectors of a SPECT/CT system at high count rate, which lead to image artifacts and limitations in QSPECT accuracy [6, 7]. This motivated us to investigate this issue more thoroughly, and also with other systems. Indeed, patients administered a high activity and/or retaining a high fraction thereof can challenge the QSPECT capability of a SPECT/CT system. In a post-therapeutic setting, this can translate into increased dosimetric errors, as well as suboptimal planning and delivery of a personalized radiopharmaceutical therapy.

In this work, we evaluated key quantitative characteristics of three dual-head NaI(Tl)-crystal SPECT/CT systems from two major vendors at high count rate using 99mTc (for practicality), including the per-detector response.

Materials and methods

SPECT/CT systems

Three contemporary dual-head Anger systems with 9.5 mm-thick NaI(Tl) crystals and a multislice CT subsystem were evaluated: System A was a Symbia T6 SPECT/CT (Siemens Healthineers, Germany). Systems B and C were, respectively, Discovery NM/CT 670 and NM/CT 870 DR (GE Healthcare, USA). Systems A and B were equipped with low-energy high-resolution collimators, while System C, with low-energy high-resolution and sensitivity ones. All energy depositions in the detector cause dead time [6, 8, 9]. Therefore, along with the 99mTc 140-keV photopeak (20% width) and the lower and upper scatter windows (10%), three additional energy windows were added to cover a wide spectrum (18 to 504 keV). All acquisitions were repeated with only the detector 1, only the detector 2, and both detectors activated. The existence of a “fast” mode on Systems B and C was brought later to our attention and was not activated for any experiment. However, it was confirmed by the vendor that no such mode exists on System A.

Planar acquisitions

For planar acquisitions, twenty 0.5-mL Eppendorf tubes were filled each with approximately 600 (System A) and 700 MBq (System B and C) of 99mTc. The tubes were placed 3 cm apart on a 30 × 23 cm, 5 mm-thick cardboard positioned equidistantly between the detectors [8]. Dynamic planar acquisitions (40 frames; 15 s/frame; 256 × 256 matrix) were performed, while adding the tubes one at a time on even-numbered frames, so that the activity was constant during odd-numbered ones. The experiment was repeated after 20 h of decay. The total activity thus ranged from 60 MBq up to 12 GBq (System A) and from 70 MBq to 14 GBq (Systems B and C).

Calibration and pileup

The systems under study were assumed to be paralyzable [7, 10]. For System A, we previously ruled out a cascaded paralyzable–non-paralyzable system, as the non-paralyzable constant was smaller than the paralyzable constant [6, 8, 11]. Sorenson described the paralyzable model of Anger cameras as [10]:

$$R_{o} = R_{t} \cdot e^{{ - R_{t} \cdot \tau }}$$

where τ, Ro and Rt correspond to the dead-time constant, the observed count rate and the true count rate, respectively. As detailed previously [6, 8], we define the camera calibration factor (CF) as the dead-time-free primary photons count rate (scatter-subtracted photopeak count rate) per activity (A). Equation (1) can be modified to express the wide-spectrum count rate (RWo) in relation with the observed primary photons count rate (RPo) as follows [6, 8], [10]:

$$R_{Wo} = CF \cdot A \cdot \frac{{R_{Wo} }}{{R_{Po} }} \cdot e^{{ - CF \cdot A \cdot \frac{{R_{Wo} }}{{R_{Po} }} \cdot \tau }}.$$

CF and τ were determined with the full range of quantifiable data obtained during the planar acquisitions by nonlinear regression of Eq. (2) using Python 3.6 (Lmfit package, least-square minimization) [6, 12, 13]. To illustrate the observed vs. expected count rate relationship, we set the expected wide-spectrum count rate as equal to \(CF \cdot A \cdot \frac{{R_{Wo} }}{{R_{Po} }}\) and assumed the same τ [8], knowing that in fact the count losses due to pileup are affecting the primary counts to a greater extent than the wide-spectrum counts. The upper limit of the quantifiable range was defined as the activity or RWo at which the observed count rate decreases (as expected from a paralyzable system), or plateaus (in case a system has absolute maximum count rate limit). Look-up tables were created to retrieve the dead-time correction factor, i.e., the ratio of the true to the observed primary count rates, for a given observed wide-spectrum count rate, RWo, during SPECT acquisition [6, 8].

To study the pileup effect, the repartition of counts between the energy windows was plotted against activity.

SPECT/CT acquisitions

A water-filled NEMA 2012/ IEC 2008 phantom (Biodex Medical Systems, USA) was customized by adding two saline bags of 250 and 500 mL inside the phantom to simulate a kidney and a large liver lesion, respectively, and two spheres to simulate additional lesions (Fig. 1a; Table 1). Acquisitions were performed over up to 26 h: 13 acquisitions of 96 projections (System A) or 9 acquisitions of 90 projections (for each Systems B and C), step-and-shoot mode, non-circular orbit, 128 × 128 matrix, 5 or 10 s per frame depending on the activity. For each dual-detector acquisition, detectors 1 and 2 started at 0° and 180°, respectively. In single-detector mode, the activated detector started at 0° and acquired all projections over 360°. Low-dose CT acquisition was performed subsequently (110 or 120 kVp, 40 or 70 mAs). Data was reconstructed using a vendor-neutral software (SPECTRA Quant, MIM Software, USA) with ordered subset expectation maximization (5 iterations, 6 subsets), CT-based attenuation, resolution recovery, and triple energy window scatter corrections.

Fig. 1
figure 1

The NEMA phantom (a) containing a 500-mL (#1) and a 250-mL (#2) bags, large sphere (#3), small sphere (#4) and the cold-water-filled cylinder (#5). Maximum intensity projection (b) and selected transaxial fusion slices from System A with both detectors (c; 4.93 GBq total activity) and only detector 1 activated (d; 5.12 GBq total activity), respectively. The scale is expressed in terms of counts per pixel

Table 1 NEMA phantom initial activity distribution

SPECT quantification and image quality

Using the CT, volumes of interest (VOIs) were manually drawn around the different compartments of the NEMA phantom. The contour of the phantom itself was automatically defined based on the CT (−400 HU threshold). A 200-mL VOI was also drawn in the background of the main compartment, far from the spill out from high-activity objects. The contours of the small saline bag and the whole phantom were both expanded by 1 cm, to include spilled-out counts [6, 14]. The estimated background activity included in the VOI expansion of the small saline bag was subtracted from its total activity. The mean VOI counts per second was quantified, dead-time corrected (by multiplying counts with the dead-time correction factor corresponding to the average acquisition RWo), divided by the actual volume, and compared to the known activity concentration.


Planar acquisitions

As activity increased, the observed count rate of System A depended on detector activation (Fig. 2a). When both activated, detectors 1 and 2 diverged above 375 kcps and eventually plateaued at specific levels (480 and 290 kcps, respectively). However, those same detectors saturated at different levels (700 and 475 kcps, respectively) when they were individually activated. While the count rate drop at higher activity that is typical of a paralyzable system was not as obvious for System A during this particular experiment, we did observe it in the past when acquiring with 99mTc using a more transparent medium-energy collimator (Additional file 1: Fig. S1). Conversely, the response of Systems B and C was independent of the detector configuration and exhibited a maximum observed count rate of 205 kcps, as predicted by Sorenson’s model, before depressing at higher activities (Fig. 2b). For the following results, A1 and A2 refer to System A with only one (detector 1) and both detectors activated, respectively.

Fig. 2
figure 2

System A (Symbia T6; a and c) and B (Discovery 670; b and d) observed count rate vs. activity (a and b) and expected count rate (c and d; only quantifiable non-grayed data points fitted to the paralyzable model) during planar acquisitions

Paralyzable model calibration

System A (CF = 99.97 ± 0.07 cps/MBq; τ = 0.4949 ± 0.0004 μs; Fig. 2c) was more sensitive and less prone to dead time than Systems B (CF = 75.37 ± 0.16 cps/MBq; τ = 1.738 ± 0.004 μs; Fig. 2d) and C (CF = 87.46 ± 0.30 cps/MBq; τ = 1.740 ± 0.008 μs; data not shown because it is similar to Fig. 2d). The different CFs between Systems B and C are only attributable to the collimators’ design. The maximum quantifiable unattenuated 99mTc activities as visually determined from Fig. 2 were equal to 6.58, 2.79, 5.02, and 4.34 GBq, respectively, for Systems A1, A2, B, and C. We confirmed that the systems do not behave as cascaded paralyzable–non-paralyzable systems, as fitting data to such model result in non-paralyzable constant that is smaller than the paralyzable one, which is not allowed (data not shown) [10].

Pileup effect

The repartition of counts between the different energy windows during planar acquisitions is presented in Fig. 3 for Systems A1 and B. Because of the pulse pileup, as activity increased, the percentage of counts in the photopeak decreased, while that in both upper scatter windows increased. The phenomenon was more pronounced with System A. Because scatter events are subtracted from photopeak events to obtain the primary counts, the pileup effect amplifies the dead-time loss of primary counts at high count rate. Nevertheless, the Sorenson’s model fits the data well.

Fig. 3
figure 3

System A1 (Symbia T6 with only detector 1 activated; a and c) and System B (Discovery 670, both detectors activated; b and d) observed count rates (a and b) and percentage of counts (c and d) for the different energy windows are plotted against activity. GLS1 = general scatter 1 (18–53 keV). GLS2 = general scatter 2 (53–112 keV). LS = lower scatter (112–126 keV). PP = photopeak (126–154 keV). US = upper scatter (154–168 keV). GUS = general upper scatter (168–504 keV). WS = wide spectrum (18–504 keV)

SPECT quantification

The observed count rate per projection during the SPECT acquisitions was consistent with the planar data, showing detector configuration-dependent saturation levels only for System A (Fig. 4). The paralyzable behaviour of systems is again evidenced by lower count rates in angular positions of maximum exposure at high activity (in particular Fig. 4a and d). After dead-time correction, accurate quantification (i.e., less than 5% deviation from true activity) of the heterogeneous attenuating phantom was achieved up to 5.12, 3.34, 2.66, and 2.30 GBq, respectively, for Systems A1, A2, B, and C (Fig. 5). However, it must be noted that for System A2, the phantom placement was favorable to the system design (hottest area primarily imaged by detector 1), and 1.64 GBq would not have exceeded the lower saturation level of detector 2 in any projection (Fig. 4c, dashed line).

Fig. 4
figure 4

Observed wide-spectrum count rate vs. projection number for System A (Symbia T6) with only the detector 1 (a), only the detector 2 (b), or both detectors activated (c), and System B with both detectors activated (d; Discovery 670; visually similar to NM/CT 870 DR) during SPECT acquisitions of the NEMA phantom. The dashed lines correspond to the maximum quantifiable activity

Fig. 5
figure 5

SPECT quantification accuracy for the whole NEMA phantom and the small saline bag (CT-based VOIs expanded by 1 cm for both) vs. activity (a and b) and observed averaged wide-spectrum count rate (c and d), for System A1 and A2 (Symbia T6 with only detector 1 and both detectors activated, respectively) and System B (Discovery 670 with both detectors activated), without (a and c) and with (b and d) dead-time correction using the paralyzable model

SPECT image quality

With System A2, we observed image artifacts when the activity was greater than 3.34 GBq, attributable to the differential detector behaviours (Fig. 1c). These artifacts were not observed with System A1 (Fig. 1d), B and C.


We previously demonstrated the necessity to correct for dead time in post-therapeutic 177Lu imaging with NaI(Tl)-crystal cameras [5]. Here, we comprehensively investigated the divergent detector behaviour of System A at high count rate, which can compromise the accuracy of quantification and dosimetry, as well as the image quality (Fig. 1c) [6]. This issue was suspected in few of our patients treated with high activities (> 20 GBq) of 177Lu-octreotate and exhibiting high tumour retention in our personalized peptide receptor radionuclide therapy regimen [2]. In these occasional, extreme clinical situations, we found that the quantitative capacity of System A at high count rate can be extended by performing acquisitions with detector 1 only. When such cases are anticipated, the camera count rate can be verified on the monitor before starting the acquisition in the appropriate dual- or single-detector mode. To avoid an excessive exam duration in single-detector mode, the time per projection can probably be reduced without risking to significantly degrade the image quality owing to the very high count rate.

For practical reasons, we used 99mTc on the basis that, for a given system, the wide-spectrum count rate as the determinant of dead time is expected to be relatively radionuclide-independent. Indeed, for System A, previously characterized as a paralyzable system, we observed a similar wide-spectrum count rate-based τ of 0.49 μs for 99mTc, vs. 0.55 μs for 177Lu [6], suggesting a limited influence of the distribution of counts within the wide spectrum on the loss of primary counts at high count rate.

To extend the scope of our work, we evaluated two additional SPECT/CT systems (differing only by their collimators) from another leading manufacturer. At 1.74 µs, the τ of Systems B and C is substantially larger than that of System A and consistent with that determined by others at 1.66 µs using 99mTc [15]. It is likely that τ of Systems B and C would be similar if measured using 177Lu. While collimator design can partly account for differences in sensitivity between cameras having otherwise the same crystal thickness, it should not, in principle, affect τ.

We later learned that a “fast” acquisition mode is available on Systems B and C (not on System A), which we did not investigate [16]. According to the vendor, this mode enables higher count rates at the expense of a lower spatial resolution and is seldom used for clinical SPECT. While the impact of this fast mode on quantitation and image quality warrants further investigation, we nevertheless compared all systems in their respective “normal” acquisition mode.

As observed by others [17], determining the dead-time constant of a system based only on the photopeak count rate (and adjacent scatter windows) would not take into account the increasing pileup effect (Fig. 3). This would eventually lead to quantification errors. We also previously demonstrated that a dead-time constant based on the photopeak count rate is dependent on the object/subject geometry (i.e., volume and shape of attenuating/scattering medium), while one based on the wide-spectrum count rate is not [6, 8]. Hence, we recommend using the latter to determine the dead time affecting the primary events.

Comparison of dead time from different manufacturers has already been done by others [7]. Our study, however, focused on the response of modern SPECT/CT systems and its impact on quantification. Our results point to some advantages of System A which features higher sensitivity, shorter dead time, and extended QSPECT capacity at very high count rate. However, its maximum observed count rate capacity is limited by an absolute total count rate limit of ~ 700 kcps in both single- or dual-detector mode, rather than the maximum count rate predicted by the Sorenson model of 743 kcps for each detector independently. This is caused by a downstream electronics bottleneck, while the prioritization of the detector 1 in dual-detector mode is the result of the design of this system.

However, new CZT-based SPECT/CT systems with improved energy resolution may become the state-of-the-art technology for personalized theranostic applications in the future [18, 19]. If the downstream electronics would not impose a bottleneck count rate limit (e.g., the plateau seen with System A), a CZT camera equipped with multiple small detector subunits (i.e., each counting only a small fraction of the total events) could in theory exhibit very little dead time at the system level.


We found significant differences in key features of modern NaI(Tl)-crystal SPECT/CT systems affecting their quantitative performance at high count rate. These are relevant to consider for accurate QSPECT, especially in the context of personalized radiopharmaceutical therapy.

Availability of data and materials

Please contact the corresponding author for the data used in this manuscript.





Calibration factor


Coefficient of variation


Cadmium zinc telluride


Quantitative single-photon emission computed tomography

RPo :

Observed primary photon count rate

RWo :

Observed wide-spectrum photon count rate


Dead-time constant


Volume of interest


  1. Del Prete M, Arsenault F, Saighi N, Zhao W, Buteau F-A, Celler A, et al. Accuracy and reproducibility of simplified QSPECT dosimetry for personalized 177Lu-octreotate PRRT. EJNMMI Phys. 2018;5(1):25.

    Article  PubMed  PubMed Central  Google Scholar 

  2. Del Prete M, Buteau F-A, Arsenault F, Saighi N, Bouchard L-O, Beaulieu A, et al. Personalized 177Lu-octreotate peptide receptor radionuclide therapy of neuroendocrine tumours: initial results from the P-PRRT trial. Eur J Nucl Med Mol Imaging. 2019;46(3):728–42.

    Article  CAS  PubMed  Google Scholar 

  3. Sundlöv A, Sjögreen-Gleisner K, Svensson J, Ljungberg M, Olsson T, Bernhardt P, et al. Individualised 177Lu-DOTATATE treatment of neuroendocrine tumours based on kidney dosimetry. Eur J Nucl Med Mol Imaging. 2017;44(9):1480–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Garske-Román U, Sandström M, Fröss Baron K, Lundin L, Hellman P, Welin S, et al. Prospective observational study of 177Lu-DOTA-octreotate therapy in 200 patients with advanced metastasized neuroendocrine tumours (NETs): feasibility and impact of a dosimetry-guided study protocol on outcome and toxicity. Eur J Nucl Med Mol Imaging. 2018;45(6):970–88.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Desy A, Bouvet GF, Frezza A, Després P, Beauregard J-M. Impact of dead time on quantitative 177Lu-SPECT (QSPECT) and kidney dosimetry during PRRT. EJNMMI Phys. 2020;7(1):32.

    Article  PubMed  PubMed Central  Google Scholar 

  6. Frezza A, Desport C, Uribe C, Zhao W, Celler A, Després P, et al. Comprehensive SPECT/CT system characterization and calibration for 177Lu quantitative SPECT (QSPECT) with dead-time correction. EJNMMI Phys. 2020;7(1):10.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Silosky M, Johnson V, Beasley C, Kappadath SC. Characterization of the count rate performance of modern gamma cameras. Med Phys. 2013;40(3): 032502.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Beauregard J-M, Hofman MS, Pereira JM, Eu P, Hicks RJ. Quantitative 177Lu SPECT (QSPECT) imaging using a commercially available SPECT/CT system. Cancer Imaging. 2011;11(1):56–66.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Arnold JE, Johnston AS, Pinsky SM. The influence of true counting rate and the photopeak fraction of detected events on Anger camera deadtime. J Nucl Med. 1974;15(6):412–6.

    CAS  PubMed  Google Scholar 

  10. Sorenson JA. Deadtime characteristics of Anger cameras. J Nucl Med. 1975;16(4):284–8.

    CAS  PubMed  Google Scholar 

  11. Uribe CF, Esquinas PL, Gonzalez M, Zhao W, Tanguay J, Celler A. Deadtime effects in quantification of 177Lu activity for radionuclide therapy. EJNMMI Phys. 2018;5.

  12. Zhao W, Esquinas PL, Hou X, Uribe CF, Gonzalez M, Beauregard J-M, et al. Determination of gamma camera calibration factors for quantitation of therapeutic radioisotopes. EJNMMI Phys. 2018;5(1):8.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Uribe CF, Esquinas PL, Tanguay J, Gonzalez M, Gaudin E, Beauregard J-M, et al. Accuracy of 177Lu activity quantification in SPECT imaging: a phantom study. EJNMMI Phys. 2017;4.

  14. Tran-Gia J, Lassmann M. Characterization of noise and resolution for quantitative 177Lu SPECT/CT with xSPECT Quant. J Nucl Med. 2019;60(1):50–9.

    Article  CAS  PubMed  Google Scholar 

  15. Halty A, Badel J-N, Kochebina O, Sarrut D. Image-based SPECT calibration based on the evaluation of the Fraction of Activity in the Field of View. EJNMMI Phys. 2018;5(1):11.

    Article  PubMed  PubMed Central  Google Scholar 

  16. Gregory RA, Murray I, Gear J, Leek F, Chittenden S, Fenwick A, et al. Standardised quantitative radioiodine SPECT/CT Imaging for multicentre dosimetry trials in molecular radiotherapy. Phys Med Biol. 2019;64(24): 245013.

    Article  CAS  PubMed  Google Scholar 

  17. Heemskerk JWT, Defrise M. Gamma detector dead time correction using Lambert W function. EJNMMI Phys. 2020;7(1):27.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Ljungberg M, Pretorius PH. SPECT/CT: an update on technological developments and clinical applications. Br J Radiol. 2017;91(1081):20160402.

    Article  PubMed  PubMed Central  Google Scholar 

  19. Peterson TE, Furenlid LR. SPECT detectors: the Anger Camera and beyond. Phys Med Biol. 2011;56(17):R145–82.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references


We are thankful to Ann Pascale Gilbert, NMT, Assistant Chief Nuclear Medicine Technologist. We also thank Alexander Hans Vija, at Siemens Healthineers, and Sébastien Clermont, at GE Healthcare, for their insight regarding the systems under study.


A.D. is the recipient of a Bourse interfacultaire from Cancer Research Centre, Université Laval. J.M.B. is supported by a Clinical Research Scholarship from the Fonds de recherche du Québec – Santé. This work was funded by the Canadian Institutes of Health Research (CIHR) operating grant MOP-142233 to J.M.B. No other potential conflict of interest relevant to this article was reported.

Author information

Authors and Affiliations



A.D. participated in the design of the study, collected and analyzed the data, and drafted the manuscript. G.B. helped with data analysis. N.L. and E.C. helped with data collection. G.B., N.L., E.C., E.T., and P.D. revised the manuscript. J.M.B. designed the study, supervised the project, contributed to data analysis, and edited the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Jean-Mathieu Beauregard.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors read the manuscript and consented for its publication.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1

: Fig. S1. Observed count rate vs. activity with System A (Symbia T6, only detector 1 activated) equipped with medium-energy low-penetration collimators.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Desy, A., Bouvet, G.F., Croteau, É. et al. Quantitative SPECT (QSPECT) at high count rates with contemporary SPECT/CT systems. EJNMMI Phys 8, 73 (2021).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: