Feasibility of iodine-123-mIBG SPECT/CT quantication in neuroblastoma using CZT and NaI detectors

Background In [ 123 I]mIBG SPECT/CT for neuroblastoma, lesion uptake quantication could improve therapy monitoring. This study compared quantitative accuracy of a CZT (WEHR45) and NaI (LEHR) system (GE Discovery 670). Methods Volume sensitivity VS hom was estimated from a homogenous cylindrical phantom (811 ml) acquired with body contour or xed 27 cm detector radius. Relative accuracy of VS hom to retrieve true background activity concentration (AC) in an IEC body phantom was calculated. Maximum/peak contrast recovery (CR) of the sphere inserts used VS IEC estimated from the IEC phantom. In 16 children with 37 [ 123 I]mIBG SPECT/CTs (range, 1-6 per patient; CZT, 12; NaI, 25), normal organ SUVmean (liver r/l, spleen, myocardium, blood pool, spine, muscles) were calculated using VS hom and VS IEC . Iterative reconstruction (Q.Metrix) used resolution recovery with/without scatter correction (SC) by dual energy window (DEW; 159 ± 10% and 130 ± 10% keV). In 20 exams, metabolic tumor volume (MTV)*SUVmax changes of primary tumors were correlated with changes of MRI tumor volume in serial scans (Pearson). Results VS hom (cts * MBq -1 * s -1 ) using SC was slightly lower for 27 cm radius vs. body contour with CZT (48 vs. 50, p<0.001) but comparable with NaI (59 vs. 61, p=0.18). Relative error in IEC phantom AC based on VS hom for CZT vs. NaI was -1.3% vs. +4.1% with SC (p=0.22). Acquisition time reduction by 50% (CZT) showed similar VS hom and relative error. CRmax/peak overestimated true AC in largest spheres (diameter, 22-37 mm) with SC and underestimated it without SC. Average SUVmean in liver and myocardium of CZT vs. NaI patients were similar. In the remaining organs (high estimated scatter proportions of >60% [CZT] or

comparability between examinations, patients, cameras and study centers. However, attempts to derive additional prognostic or predictive information from (semi-)quantitative parameters in [ 123 I]mIBG SPECT/CT have been scarce (8).
The accuracy of quanti cation in SPECT/CT is not only in uenced by patient-speci c factors (patient geometry, attenuation, scatter within the patient) and system-speci c factors (e.g., intrinsic resolution, collimator-detector response [CDR], reconstruction algorithms) (9)(10)(11)(12)(13)(14)(15)(16). Detection properties for different radionuclides also differ between detector materials, which are commonly sodium iodide (NaI) or -more recently -cadmium zinc telluride (CZT). Brady et al. recently showed that quanti cation of [ 123 I]mIBG SPECT/CT in patients with neuroblastoma is feasible with two different NaI cameras and can provide comparable standardized uptake values (SUV) for normal organs between a low energy high resolution (LEHR) and a medium energy (ME) collimator (17).
The current study investigated a NaI and CZT camera to evaluate if quanti cation of clinical [ 123 I]mIBG SPECT/CT data using phantom-based volume sensitivity estimates is equally feasible with both cameras using standard reconstruction algorithms provided by the manufacturer. In addition to different detectors, major in uencing factors (detector positioning, acquisition time, object geometry and scatter correction) were evaluated to determine if their variation in clinical routine would substantially affect the reproducibility of quantitative data. Finally, an exemplifying analysis using quantitative [ 123 I]mIBG SPECT/CT data of NB patients examined the correlation of changes in metabolic tumor volume (MTV)*SUVmax with changes of the tumor volume in MRI during therapy as surrogates for response.
For both the homogenous and the IEC phantom, data with the CZT camera was acquired in list mode, and all data were retrospectively rebinned to reconstruct further datasets representing an acquisition time of 50% (i.e, 20 s/step).
Phantom measurements: Volume sensitivity (homogenous phantom) To estimate volume sensitivity for the homogenously lled phantom, a cylindrical volume of interest (VOI) with 11 cm diameter (volume, 1.34 liters) was created covering the whole phantom in each dataset (ROVER software, version 3.0.34, ABX advanced biochemical compounds GmbH, Radeberg, Germany).
Volume sensitivity (cts * MBq − 1 * s − 1 ) was calculated by dividing the overall counts in the VOI by the total acquisition time in s, the decay-corrected applied activity in MBq and the voxel size in mm 3 . The mean (and standard deviation; SD) of 4 separate acquisitions with identical acquisition and reconstruction settings at different time points after initial phantom lling was used (with correction for decay).
Estimated volume sensitivity was validated with the IEC phantom data. The analysis was performed using a cylindrical VOI (volume, 644 ml) placed in the background of the IEC phantom with adequate distance from the sphere inserts and the outer phantom wall. The mean volume sensitivity estimated from measurements of the homogenously lled phantom for the speci c camera and reconstruction (SC vs. non-SC) was used to determine background activity concentration (AC) in the IEC phantom in kBq/ml and its relative accuracy compared to the true AC (mean ± SD of 3 separate acquisitions each).
Phantom measurements: Contrast recovery (IEC phantom) To calculate maximum and peak contrast recovery (CR) for the IEC phantom spheres, volume sensitivity for the IEC phantom was estimated using a cuboid VOI (volume, 25.3 liters) for each dataset that covered the entire IEC phantom. Volume sensitivity was calculated analogously to the homogenous phantom (mean of 3 serial acquisitions). Using the ACCURATE tool (version v23102018, Ronald Boellaard, Amsterdam UMC, Amsterdam, The Netherlands), the maximum counts of each of the six spheres was estimated. Furthermore, the peak counts of each sphere were determined as the average counts in a spherical VOI of 1.2 cm in diameter that was automatically positioned to calculate the highest peak counts of the corresponding sphere. Sphere AC in kBq/ml was derived from sphere counts using the mean IEC phantom volume sensitivity obtained from the three serial measurements. CRmax / CRpeak were then calculated as the ratio of maximum/peak sphere AC to true AC.

Phantoms and patients: Estimation of scatter proportions
The estimated proportion of scatter relative to all reconstructed counts was calculated as (1 -SC data counts / non-SC data counts).

Patients: Characteristics
The patient analysis included 16  , spine, gluteal muscles) were derived with adequately small spherical VOI (unless the location was affected by metastases). In 6 patients with serial examinations (20 examinations; CZT, 13; NaI, 7), maximum and mean counts of the NB primary tumors were measured at each examination (38 measured values overall) after delineation of the metabolic tumor volume (MTV) with a threshold-based, background-adapted algorithm (8,19). SUVmax and SUVmean were calculated using volume sensitivity for the respective camera and reconstruction setting (non-SC, SC), voxel volume (ml) and injected activity per kg (decaycorrected). SUV in normal organs and primary tumor lesions primarily used volume sensitivity obtained from the homogenous phantom as recommended by the manufacturer (20). To account for higher scatter proportions (e.g., in larger patients or in organs with higher scatter proportions), normal organ SUV were also calculated based on volume sensitivity from the larger IEC phantom for comparison. No partial volume correction was applied. The product MTV*SUVmax was calculated for primary tumor lesions.

MRI
The primary tumor MRI volume was calculated by 0.5 * (transversal * coronal * sagittal diameter) in MRI examinations corresponding to each [ 123 I]mIBG SPECT/CT examination in the 6 patients with serial examinations.

Statistical analysis
Statistical analysis was performed using SPSS 22 (IBM Corporation, Armonk, NY, USA) and R 3.6.1 (Foundation for Statistical Computing, Vienna, Austria, 2019, http://www.R-project.org). Descriptive parameters were expressed as mean and SD (phantom data) or median, IQR and range (patient data). The t test was used for comparison of volume sensitivities or CR between paired data (50% vs. 100% acquisition time or non-SC vs. SC) or independent data (27 cm detector distance vs. body contour). SUV from patient data were compared with the Wilcoxon test (non-SC vs. SC data) or Mann-Whitney U test (CZT vs. NaI). Coe cients of variation (CV) were compared using the R package cvequality based on the asymptotic test by Feltz and Miller (21). Correlation of relative changes in MRI volume with relative changes in MTV*SUVmax on serial scans was analyzed with Pearson's correlation coe cient.
Interpretation of correlation coe cients was performed as previously proposed (22). Statistical signi cance was assumed at p < 0.05.

CRmax and CRpeak
In the three smallest spheres, CRmax and CRpeak for each sphere were similar for SC vs. non-SC data with CZT and NaI (each p > 0.05; Table 1). In the three largest spheres, CRmax and CRpeak were each signi cantly higher with SC vs. non-SC for both CZT and NaI (each p < 0.05; Table 1). In the four smallest spheres, both CRmax and CRpeak consistently underestimated true AC in non-SC and SC data (Fig. 1). In the two largest spheres, CRmax and CRpeak in SC data overestimated sphere AC while both CRmax and CRpeak in non-SC data underestimated AC.
Acquisition time reduction by 50% for the CZT affected neither mean CRmax nor CRpeak of any sphere in non-SC or SC data, respectively (50% vs. 100%, each p > 0.05; Table 1).
Estimates in scatter proportions (phantoms and normal organs) With both CZT and NaI, mean scatter proportions estimated by the DEW method in the homogenous phantom (CZT, 52.0 ± 0.7%; NaI, 30.4 ± 0.4%) were comparable to scatter proportions estimated for the right and left liver lobe and the myocardium (range of median proportions, CZT, 49.5 to 59.9%; NaI, 23.2 to 35.2%; Table 2). In contrast, mean scatter proportions in the IEC phantom (CZT, 72.3 ± 0.3%; NaI, 50.3 ± 1.0%) were similar to proportions estimated for the spleen, MBPS, spine and gluteal muscles (range of median proportions, CZT, 75.6 to 80.0%; NaI, 41.3 to 55.5%). Scatter proportions estimated by the DEW method (k = 1.0 for both cameras) are given for both phantoms (mean ± SD) and for each organ (median, IQR). For both cameras, estimated scatter proportion in the homogenously lled phantom were similar to estimated proportions in the liver and myocardium while scatter proportions in the IEC phantom were more comparable to spleen, MBPS, spine and gluteal muscles.

Normal organ SUVmean: Effect of different scatter proportions
Based on volume sensitivity obtained with the homogenous phantom (i.e., at scatter proportions that matched liver and myocardium), SUVmean for each normal organ were signi cantly higher for non-SC vs. SC data with both CZT and NaI (each p < 0.05; except for myocardium with CZT and left liver lobe with NaI; Fig. 2A).
Based on volume sensitivity from the IEC phantom (i.e., at scatter proportions that matched spleen, MBPS, spine and muscles), SUVmean were similar between non-SC and SC data in the spleen with the CZT (p = 0.64), in MBPS (p = 0.68) and spine (p = 0.07) with the NaI and in the gluteal muscles with both CZT and NaI (each p > 0.1). In contrast, normal organ SUVmean in right/left liver lobe and myocardium were signi cantly higher for SC data than non-SC data (both cameras, each p < 0.01; Fig. 2B).

Normal organ SUVmean: Comparison of CZT and NaI
Normal organ SUVmean were similar in CZT examinations vs. NaI examinations for non-SC data (each p > 0.1; Fig. 3A).
In SC data based on volume sensitivity obtained from the homogenous phantom, normal organ SUVmean were comparable between CZT vs. NaI examinations for right/left liver lobe, spleen and myocardium (each p > 0.05) but different for MBPS, spine and gluteal muscles (each p < 0.05; Fig. 3B). In SC data based on volume sensitivity from the IEC phantom, normal organ SUVmean were also comparable between CZT vs. NaI examinations for right/left liver lobe, spleen and myocardium (each p > 0.05) but different for MBPS, spine and gluteal muscles (each p < 0.05).

Coe cient of variation (CV) in normal organ SUVmean
CV of SUVmean in organs with high scatter proportion (spleen, MBPS, spine and gluteal muscles) were signi cantly higher in SC vs. non-SC data with the CZT (each p < 0.05) but comparable between non-SC vs. SC data with the NaI (each p > 0.05; Table 3). CV of SUVmean in the liver and myocardium were similar between non-SC vs. SC data with both cameras (each p > 0.05). CV with both cameras were identical between SC data obtained with either volume sensitivity from the homogenous phantom or IEC phantom, respectively (each p = 1.0).

Correlation of MRI volume changes with MTV*SUVmax changes
In 6 patients with serial examinations (n = 20; CZT and NaI mixed), Pearson correlation coe cient of changes in MRI volume with changes in MTV*SUVmax was similar for non-SC data (r = 0.62; p = 0.014) vs. SC data (r = 0.59; p = 0.021; Fig. 4).

Discussion
This study examined the feasibility of quanti cation in [ 123 I]mIBG SPECT/CT data in phantom measurements and NB patients as a basis for potential application of quantitative image parameters for predictive or prognostic purposes.
Numerous factors in uence quantitative accuracy in SPECT/CT. While attenuation is routinely addressed with CT-based attenuation correction, correction for scatter, e.g. within the patient, is not necessarily an integral part of iodine-123 SPECT/CT reconstruction in clinical routine (i.e., for visual interpretation). As highlighted by the current data, the scatter proportion estimated by the DEW method for iodine-123 in an IEC phantom geometry can account for up to half (NaI; LEHR collimator) (23) or even > 70% of the acquired counts (CZT; WEHR45 collimator). This implicates that in [ 123 I]mIBG imaging for children, which is in principle characterized by low count statistics and unfavorable noise properties, the additionally available counts in non-SC data could be bene cial for image quality and resulting con dence for visual reading in clinical care.
However, if quantitative accuracy is intended, high trueness (i.e., low systematic deviations from the true value) and precision (i.e., high reproducibility of measurements) are required. In phantom measurements, trueness of non-SC data in recovering the background activity concentration in the IEC phantom based on volume sensitivity of the smaller, homogenously lled phantom was poor (average relative error > 70%). In contrast, SC data showed high trueness (low average relative error < 5%) and high precision (low variability between serial scans) suggesting that the DEW method with a low-energy scatter window could be su cient for SC in iodine-123 SPECT/CT if only the accurate depiction of homogenous activity concentrations in a su ciently large volume is required. However, SC data considerably overestimated CR of the larger sphere inserts. This is, among other effects, due to the inability of the DEW method to account for the spatial distribution of scattered photons that originate from the sphere inward but are detected in the background and therefore systematically increase sphere-to-background ratios (24). This overestimation increases with an increase of the weighting factor k for the scatter window (k = 1.0 in the current study for both cameras) (25). In contrast, without SC, CRpeak underestimated the AC of larger sphere inserts by about 30% (NaI) or 40% (CZT) which is in accordance with Lagerburg et al. (25). The consistent results for CRmax and CRpeak show that this observation is not merely due to the chosen delineation method (CRpeak) or statistical outliers (CRmax). The general observation that CRpeak better represents sphere AC in larger spheres compared CRmax, which usually overestimates it, has been previously demonstrated for positron emission tomography (PET) (26).
In patient data, trueness usually cannot be determined unless the standard of truth is known from activity concentrations determined in vivo (e.g., in the urine (27)). However, one should aim at achieving comparability between camera systems. Similarity in average normal organ SUV between both cameras served as a surrogate assuming that normal organ SUV should be similar on average -even in different patient samples. Under this premise, SUV in non-SC data were similar between both cameras for all normal organs. Intuitively, the similarity of estimated scatter proportions between the homogenous phantom and organs such as liver and myocardium as well as between the IEC phantom and organs such as MBPS and spine would suggest that the DEW method could overall account for varying scatter geometries in different organs. Consequently, normal organ SUV in SC data should also be similar between both cameras -irrespective of the examined organ and its estimated scatter proportion. In contrast, in speci c organ geometries with high estimated scatter proportions (spleen, MBPS, spine, gluteal muscles), SUV in SC data were different between CZT and NaI patients. It must be hypothesized that the use of a unique k value for both systems accentuates these effects of SC. Furthermore, SC could not reduce the imprecision in patient data when the variation of normal organ SUVmean (CV) between different examinations was used as a surrogate.
In summary, surrogates for both trueness and precision in the current patient data imply that SC based on the DEW method with a low-energy scatter window is insu cient for quantitative accuracy in clinical iodine-123 SPECT/CT with both CZT (WEHR45) and NaI cameras (LEHR collimator). In both detectors, the DEW method is limited by the inability to account for spatial distribution of scattered photons (see above). Moreover, it relies on calibration of the k factor which is then applied indiscriminately to the acquired dataset, and DEW cannot account for downscatter from the 529 keV iodine-123 peak (28,29); both adaptations would require a third energy window above the photopeak window. In CZT detectors, overestimation of scatter using the DEW method will result from the detector-speci c low-energy tail which is caused by contamination from photons that are unscattered but detected with lower energy (28,30).
Considering these spatially variant and invariant sources of error in SC for iodine-123 with the simpli ed DEW method, Monte Carlo based SC in combination with CDR modelling may be superior in achieving accurate quantitative data in the complex and variable geometry of the patient body (13,31); however, these algorithms are resource consuming and not commonly integrated in clinical SPECT/CT systems.
Brady et al. recently investigated SUV in [ 123 I]mIBG SPECT/CT acquired with two NaI cameras (LEHR and ME collimator) in 43 patients with NB. Using Monte Carlo based SC and volume sensitivity from a homogenously lled cylindrical phantom, normal organ SUV (salivary glands, heart, liver, adrenal glands, urinary bladder) were -on average -similar between both cameras. However, considerable variation in SUV of all organs remained between examinations. IQR of liver SUV was 1 to 2 and therefore comparable to CZT examinations in the current study but lower than in NaI examinations in the present analysis (IQR, 1 to 3). It may be noted that even with optimal image acquisition and processing, physiological variability of normal organ SUV will occur, which will itself vary between organs (e.g., due to varying sympathetic innervation of the left ventricular myocardium (32)). If normal organ SUV variation remains high, it could ultimately limit the potential of normal organ SUV in [ 123 I]mIBG imaging to serve as physiological intraindividual reference as has been commonly proposed for the liver or MBPS SUV in [ 18 F] uorodeoxyglucose (FDG) PET (33,34).
Independent of the SC method used for quantifying patient data, an appropriate calibration procedure (e.g., scan protocol, phantom geometry) must be chosen to estimate volume sensitivity for SPECT data. If the employed SC method is accurate, it would be su cient to obtain volume sensitivity from a homogenously lled cylindrical phantom (20). Appropriateness of the volume sensitivity could be examined under varying scatter properties using a body phantom. However, the current results suggest that further steps will be required to ensure that the SC method is appropriate for normal organ and lesion quanti cation in patients. This may ultimately require in vivo measurements of activity concentrations, e.g. in the urine (27).
Further in uencing factors were examined in the current study (detector radius, acquisition time reduction with the CZT). Differences in volume sensitivity for the homogenous phantom between acquisition with body contour or xed 27 cm detector radius were < 5% in SC data. This is facilitated by resolution recovery as part of image reconstruction which aims at compensation for CDR including its variation at different detector radii along the angular range (35,36). Although relative differences were small, volume sensitivity with the CZT at 27 cm radius was signi cantly lower than with body contouring. This may be due to higher dependency of effective spatial resolution from source-to-collimator distance compared to NaI detectors (37)(38)(39). Consequently, variation in detector distance among patients could add systemspeci c variance in quantitative accuracy. However, the currently chosen differences in detector radii were comparably large, especially considering the context of a pediatric population, with the aim of identifying the consequential deviations under extreme conditions -while (considerably) smaller deviations can usually be expected in clinical routine. Furthermore, volume sensitivity for both phantoms as well as CRmax and CRpeak of the sphere inserts did not differ relevantly between 100% or 50% acquisition time (CZT). Availability of data and material The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.  Normal organ SUVmean (non-SC vs. SC) Boxplots of SUVmean either based on volume sensitivity obtained from the homogenous phantom (A) or IEC phantom (B) are separated by organs and for non-SC vs. SC data. Outliers are displayed as circles or asterisks. In organs with high scatter proportion (spleen, MBPS, spine and muscles), SUVmean in non-SC data and SC data are similar if based on volume sensitivity from the IEC phantom (matching high scatter proportion). In contrast, non-SC and SC data are more similar in liver and myocardium (low scatter proportion) if volume sensitivity is obtained with the homogenous phantom (matching low scatter).

Figure 3
Normal organ SUVmean (CZT vs. NaI) Boxplots of SUVmean from non-SC (A) or SC data (B) for CZT vs.
NaI, both based on volume sensitivity obtained from the homogenous phantom. Outliers are displayed as circles or asterisks. SUVmean in non-SC data show lower differences between CZT and NaI than SC data. Results for SC data based on volume sensitivity obtained from the IEC phantom are not displayed as they are almost identical to data in (B).