Accuracy of holmium-166 SPECT/CT quantification over a large range of activities

Abstract

Background

Quantitative imaging is a crucial step for dosimetry in radionuclide therapies. Traditionally, SPECT/CT imaging is quantified based on scanner-specific conversion factors or self-calibration, but recently absolute quantification methods have been introduced in commercial SPECT reconstruction software (Broad Quantification, Siemens Healthineers). In this phantom study we investigate the accuracy of three quantification methods for holmium-166 SPECT/CT imaging, and provide recommendations for clinical dosimetry.

Methods

One cylindrical phantom, filled with a homogeneous holmium-166-chloride activity concentration solution, was imaged at one time point to determine a scanner-specific conversion factor, and to characterize the spatial dependency of the activity concentration recovery. One Jaszczak phantom with six fillable spheres, 10:1 sphere-to-background ratio, was imaged over a large range of holmium-166 activities (61-3130 MBq). The images were reconstructed with either an ordered subset expectation maximization (OSEM, Flash3D-reconstruction; scanner-specific quantification or self-calibration quantification) or an ordered subset conjugate gradient (OSCG, xSPECT-reconstruction; Broad Quantification) algorithm. These three quantification methods were compared for the data of the Jaszczak phantom and evaluated based on whole phantom recovered activity, activity concentration recovery coefficients (ACRC), and recovery curves.

Results

The activity recovery in the Jaszczak phantom was 28–115% for the scanner-specific, and 57–97% for the Broad Quantification quantification methods, respectively. The self-calibration-based activity recovery is inherently always 100%. The ACRC for the largest sphere (Ø60 mm, ~ 113 mL) ranged over (depending on the activity level) 0.22–0.89, 0.76–0.86, 0.39–0.72 for scanner-specific, self-calibration and Broad Quantification, respectively.

Conclusion

Of the three investigated quantification methods, the self-calibration technique produces quantitative SPECT images with the highest accuracy in the investigated holmium-166 activity range.

Background

Over the past decades, the medical isotope holmium-166 (¹⁶⁶Ho) has become a widely explored therapeutic radionuclide [1]. Its primary clinical application is the ¹⁶⁶Ho poly-L-lactic acid microsphere used in medical procedures such as transarterial radioembolization (TARE) or direct intratumoral injection of the microspheres [2]. During TARE, radioactive microspheres are injected in (a subbranch of) the hepatic artery, and are distributed through the perfused liver volume via the blood flow. The microspheres ultimately lodge in the arterioles, where they locally irradiate the tissue [3].

¹⁶⁶Ho emits high-energy beta particles (T_1/2: 26.8 h; max energy 1.77 MeV/50.5%, 1.85 MeV/48.2%),a primary gamma peak at 80.57 keV/6.55%, which is typically used for planar scintigraphy and single photon emission computed tomography with computed tomography (SPECT/CT) imaging, and a spectrum of high energy gammas with lower abundance (1.38 MeV/0.93%, 1.58 MeV/0.19%, 1.66 MeV/0.12%) [4, 5]. Scattered high energy photons, bremsstrahlung photons from the beta radiation, as well as characteristic x-rays from the lead collimator [6], all contribute towards contamination of the main ¹⁶⁶Ho photopeak window. The scattered high energy, and the bremsstrahlung photons can be estimated for using window-based scatter correction in the reconstruction of the SPECT images, but since the spectrum of the characteristic lead x-rays overlaps with both the main ¹⁶⁶Ho photopeak window as well as possible adjacent scatter windows, these interactions can only be accounted for using a dedicated Monte Carlo-based scatter correction method. In ¹⁶⁶Ho-TARE, SPECT imaging is used for both pre-treatment dosimetry and therapy planning [7], as well as for post-treatment dosimetry and dose-response evaluation [8]. As the ¹⁶⁶Ho microspheres are paramagnetic, magnetic resonance imaging-based (MRI) dosimetry can also be performed, as an alternative to traditional nuclear medicine based dose evaluation [9, 10]. For treatment optimization based on patient dosimetry, a thorough understanding of dose-response relationships after radionuclide therapy, and verification of the absorbed doses, is vital [11]. This requires standardized and optimized quantitative imaging and dosimetry protocols, for clinical studies to incorporate [8].

Although SPECT/CT-based dosimetry for ¹⁶⁶Ho-TARE treatment planning and evaluation is considered the current gold standard [6], its value is limited by the low spatial resolution, motion artifacts, and the lack of standardized quantification. Traditionally, SPECT/CT images have been quantified based on a pre-determined calibration factor measured in a phantom [12,13,14,15,16]. One major drawback of this method is that it is phantom dependent and/or scanner-specific which can make it cumbersome to quantitively compare absorbed doses in multi-scanner, multi-center studies. An alternative method is to quantify the SPECT/CT images based on a patient self-calibration conversion factor [4, 15]. This method is uniquely applicable for localized radionuclide therapies such as TARE, as all the activity can be imaged within one field-of-view of the SPECT (possible extrahepatic deposition can be estimated for based on SPECT or planar scintigraphy) and because the microspheres have no biological half-life (permanent entrapment of the radioactive particles). A single bed position, single time point SPECT/CT scan is therefore sufficient for the determination of the self-calibration conversion factor. A third option for quantification, recently introduced by some vendors, is absolute SPECT quantification in commercial software [17, 18], which could potentially aid in increasing the reproducibility and the reliability of the reported absorbed doses.

In this phantom study, we quantitively compare three different ¹⁶⁶Ho-SPECT/CT-based quantification methods; scanner-specific conversion, self-calibration, and a commercial software for absolute quantification (Broad Quantification, Siemens Healthineers, Erlangen, Germany). Furthermore, we evaluate the accuracy of these quantification methods for the purpose of post-treatment dosimetry over a large range of ¹⁶⁶Ho activities, and we provide recommendations for post-treatment absorbed dose evaluation in clinical practice.

Table 1 Activity in the phantoms present at injection and at the time point of scanning. The dead-time [%] was read from the workstation during the data acquisition. Measurements at time points marked with * were repeated three times

Methods

Phantom preparation

Two phantoms were utilized in this study. Firstly, one cylindrical phantom was filled with a homogeneous distribution of ¹⁶⁶Ho-chloride dissolved in demineralized water (HolmiumSolution, Quirem Medical B.V., Deventer, The Netherlands). This phantom had an inner diameter of 20 cm, an inner height of 20 cm, a volume of 6283 mL, and an activity concentration of 0.049 MBq/mL at the injection time point. The second phantom was a flanged, rod-less, Jaszczak phantom with six fillable spheres, see Fig. 1 for sphere numbering and configuration. The background compartment of the Jaszczak phantom had a volume of approximately 6.7 L and the fillable spheres had inner diameters of 9.9, 15.4, 19.8, 24.8, 31.3, 60 mm (and approximate volumes 0.5, 2.0, 4.0, 8.0, 16.0, 113 mL). The Jaszczak phantom was also filled with homogeneous solutions of ¹⁶⁶Ho-chloride. The activity concentration was 4.10 MBq/mL in the spheres and 0.41 MBq/mL in the background compartment at the injection time point, resulting in a sphere-to-background activity concentration ratio ($\:\text{S}:\text{B}{\text{G}}_{\text{r}\text{a}\text{t}\text{i}\text{o}}$) of 10:1. The total activity in the Jaszczak phantom ranged from 3.26 GBq to 61 MBq throughout the experiment, see Table 1. The ¹⁶⁶Ho activities were measured in a dose calibrator (manufactured in 2019, equipped with ionization chamber VIK-202, Comecer, Joure, The Netherlands). In both phantoms the ¹⁶⁶Ho-chloride solution was saturated with 50 mM ethylenediaminetetraacetic acid to bind the holmium ions, forming water-soluble complexes.

Fig. 1

CT images of the Jaszczak phantom with the fillable spheres. Axial view on the left and sagittal view on the right. The sphere numbering presented in this image is used throughout this paper

SPECT/CT system

Both phantoms were imaged using a Symbia Intevo Bold SPECT/CT scanner equipped with the proprietary absolute quantification software Broad Quantification and the additional TrueCalc software (Siemens Healthineers, Erlangen, Germany). The purpose of TrueCalc is to improve the detector response at high count rates by reducing the impact of detector dead-time and give consistent quantitative accuracy across a wide range of activities [17, 19, 20]. The SPECT/CT scanner was manufactured and installed in 2019, uses software version VB22A, features a dual-headed detector configuration with a field-of-view of 53.3$\:\times\:$38.7 cm² and a NaI crystal with a thickness of 9.5 mm. Prior to the measurements the SPECT/CT scanner had been peaked with a ~ 5 MBq ¹⁶⁶Ho point source, and the Broad Quantification calibration procedure for ¹⁶⁶Ho had been carried out in accordance with the operator manual, see supplementary information.

Data acquisition

Imaging of the homogeneous phantom was acquired at one time point (total activity of 250 MBq at scanning). Imaging of the Jaszczak phantom was acquired at 14 different time points with three repetitions at the 10th, 12th, 13th, and 14th time point. The dead-time [%] estimate was read from the workstation during the acquisition of the data (‘Tomo Acquisition’-activity, ‘Analyzer’-tab), see Table 1. All SPECT data were acquired with the following settings: a photopeak energy window at 81 keV (15% width), two adjacent scatter windows (8% width), non-circular orbit over 360°, step and shoot, 2 × 60 projections, 20 s projection time, and a 128 × 128 matrix size. Parallel hole, Medium Energy Low Penetration (MELP) collimators were used, and the checkbox for ‘Enable Quantitative Acquisition’ had been ticked. When acquiring quantitative data, with the additional dead-time correction software (TrueCalc), the SPECT scanner generates two different projection datasets per SPECT acquisition; one dataset that is corrected for the estimated count losses due to dead-time effects, and a second dataset that is uncorrected. The measured counting rate performance per time point, for these two projection datasets, can be found in the Supplemental information. The dead-time estimates, presented in Table 1, are unaffected by the TrueCalc software. Additionally, low-dose CT data for attenuation correction and structural information were acquired for each dataset. The positioning of the Jaszczak phantom was standardized across all measurements, see Fig. 1.

Image reconstruction

The acquired SPECT data for both phantoms were reconstructed using two different Siemens proprietary algorithms; Flash3D and xSPECT. The widely used Flash3D method is a 3D ordered subset expectation maximization (OSEM) method (utilized the uncorrected projection data). For Flash3D, the data were reconstructed with 10 iterations 8 subsets, and a voxel size of $\:4.80\times\:4.80\times\:4.80\:\text{m}{\text{m}}^{\text{3}}$. The xSPECT algorithm utilizes an ordered subset conjugate gradient (OSCG) method and Broad Quantification was enabled, meaning that the images were reconstructed in the unit of Bq/mL. Additionally, the xSPECT algorithm also applied the TrueCalc count loss correction in the image reconstruction (utilized the projection data corrected for count losses). For xSPECT, the data were reconstructed with 36 iterations 1 subset, a voxel size of $\:4.88\times\:4.88\times\:4.88\:\text{m}{\text{m}}^{\text{3}}$, and decay corrected back to the injection time point. Triple energy window scatter correction (automatic scatter window weight of 0.94) and CT based attenuation correction was applied for both reconstruction methods. If nothing else is indicated, no post-reconstruction filtering was applied to the Flash3D data and a 15 mm full width at half maximum (FWHM) post-reconstruction filtering was applied to the xSPECT data. See supplementary information for the choice of image reconstruction parameters.

Data analysis

All image delineation and all data analysis were performed using MATLAB (R2021a, The MathWorks Inc., Natick, Massachusetts).

Image delineation

All volumes-of-interests (VOIs) for spheres, background and whole phantom volumes, were semi-automatically created based on the known phantom geometry and CT based positioning on a high resolution CT (voxel size $\:0.98\times\:0.98\times\:1.5\:\text{m}{\text{m}}^{\text{3}}$). The background VOI in the Jaszczak phantom was defined by subtracting the sphere VOIs from the whole phantom VOI. To minimize loss of data precision in the analysis, the reconstructed SPECT data were transformed to the high resolution CT voxel space by means of an affine transformation and nearest neighbor interpolation. All analyses were carried out using the high resolution grid.

SPECT quantification

Three methods for quantifying the SPECT data for the Jaszczak phantom were compared in this study;

• 1. Scanner-specific conversion factor measured in the Flash3D-reconstruction of the cylindrical homogeneous phantom (CF_homogeneous) [12,13,14,15,16],

• 2. Self-calibration based on the Flash3D-reconstruction and delineation of the Jaszczak phantom for each dataset (CF_self) [4, 15], and.

• 3. Broad Quantification absolute quantification software (based on xSPECT-reconstruction).

For the two Flash3D-based quantification methods, the conversion factor (CF) was calculated using the following formula;$\:\text{C}\text{F}=\frac{\left(\frac{{\mu\:}}{\text{t}\cdot\:\text{n}\cdot\:{\text{V}}_{\text{v}\text{o}\text{x}\text{e}\text{l}}}\right)}{\left(\frac{\text{A}}{{\text{V}}_{\text{V}\text{O}\text{I}}}\right)}\hspace{1em}\hspace{1em}\left[\frac{\text{cps}\text{/}\text{mL}}{\text{MBq}\text{/}\text{mL}}\right],\hspace{1em}$

where $\:{\mu\:}$ is the mean counts in the VOI to be quantified, $\:\text{t}$ is the time per projection [s], $\:\text{n}$ is the number of projections, $\:{V}_{voxel}$ is the voxel volume [mL], $\:A$ is the true total activity in the VOI [MBq], and $\:{V}_{VOI}$ is the volume of the VOI [mL]. The CF_homogeneous was calculated based on the activity present at scanning in a two liter cylindrical VOI centered in the homogeneous phantom. The CF_self was calculated based on the total activity present in the Jaszczak phantom at the start of each SPECT acquisition, and a VOI encompassing the whole phantom. Each Flash3D-reconstruction was processed to have the unit of counts per second per milliliter [cps/mL], and was thereafter quantified by dividing each voxel in the image with either the CF_homogeneous or CF_self. Lastly, the quantified images [MBq/mL] were decay corrected to the injection time point.

The Broad Quantification-based quantification method directly quantifies the xSPECT-reconstruction in units of Bq/mL and decay corrects it to the injection time point, see supplementary information for information about the calibration procedure.

Homogeneity evaluation in cylindrical phantom

Based on the quantitative SPECT images of the cylindrical homogeneous phantom, the homogeneity of the reconstructions were evaluated through assessment of activity concentration profiles (ACPs) parallel to the phantom’s symmetry axis. Three voxel row profiles and two cylindrical VOIs were evaluated. The ACPs for the cylindrical VOIs were obtained by averaging the activity concentration values, for all voxels within a specified radius, per slice in the SPECT data. Furthermore, cumulative activity-volume histograms (cAVH) have been calculated for one VOI encompassing the whole phantom volume, and two cylindrical VOIs centered in and having the same length, but with a smaller radius (either 30 mm or 70 mm), as the cylindrical phantom. The cAVHs were qualitatively compared to the theoretical ideal cAVH for a homogeneous distribution, which is a step function. Images without post-reconstruction filtering and with a 15 mm FWHM Gaussian filter were compared in this analysis.

Activity recovery in Jaszczak phantom

The percentage of recovered activity was computed for each of the three SPECT quantification methods and each time point of the Jaszczak phantom measurements. All the voxels in a VOI encompassing the whole phantom, i.e. the same VOI as was used for the determination of CF_self, was considered for this calculation.

Activity concentration recovery coefficient

The activity concentration recovery coefficients (ACRC) were calculated for the sphere VOIs, and the background VOI for each SPECT quantification method and for each time point of the Jaszczak phantom measurements. The ACRC was defined as$\:ACRC=\frac{A{C}_{quantifiedSPECT}}{{AC}_{true}},$

where$\:{\:AC}_{quantifiedSPECT}$ is the mean activity concentration of the sphere or background VOI in the quantified SPECT image [MBq/mL], and $\:{AC}_{true}$ is the true activity concentration in the same VOI [MBq/mL].

Recovery curve

Recovery curves that can be used for correcting images for losses related to partial volume effects (PVEs), i.e. ACRC as a function of sphere diameter (d), were created. The ACRCs for the three repeated measurements at time point 10, 12, 13 and 14 were averaged for each quantification method and the dependency of the activity level was visualized. A 3-parameter (a, b, c) logistic function on the form

$$\:ACRC\left(d\right)=a\left(1-\frac{1}{1+{\left(\frac{d}{b}\right)}^{c}}\right)$$

was fitted to the data [13]. The a-coefficient represents the asymptote that the fitted curve is converging towards when increasing the spherical diameter of a hot object, indicating the theoretically highest ACRC possible for each quantification method and activity level.

Results

Conversion factors for SPECT quantification

Conversion factors used for quantifying the Flash3D-reconstructions were calculated based on the homogeneous phantom and for each measurement of the Jaszczak phantom, see Fig. 2. The value of CF_homogeneous (19.26 cps/MBq) is similar to the CF_self determined at 206–217 MBq (19.26–19.29 cps/MBq).

Fig. 2

Conversion factors (CF) for two quantification methods utilizing the Flash3D reconstructions: CF_homogeneous and CF_self. The CF_homogeneous has been measured and calculated for one time point (at 250 MBq) and the CF_self was calculated for each measurement of the Jaszczak phantom (total of 22 measurements)

Homogeneity cylindrical phantom

The ACPs for the cylindrical homogeneous phantom can be found in Fig. 3. The locations of the voxel profiles and volumes are visualized in Fig. 3a, b and the profiles in Fig. 3c-f. The ACPs for the displayed voxel rows of the unfiltered Flash3D-reconstruction vary between 53% and 181% of the AC_true within the phantom (a margin of 30 mm from the edge was considered), and for the xSPECT-reconstruction the corresponding value vary between 0% and 297% (Fig. 3c, d). After filtering this variation has been reduced to 62–153% and 52–200% of AC_true, for Flash3D and xSPECT respectively (Fig. 3e, f). The ACPs averaged over a radius of 30 mm is varying between 77% and 132% of AC_true for the unfiltered Flash3D, and 75% and 151% for the unfiltered xSPECT-reconstructions. This is reduced to 85–123% and 87–130% for the filtered Flash3D and xSPECT data, respectively. The ACPs averaged over a radius of 70 mm can be considered to conform to AC_true as the deviation is within ± 10% (93–107% unfiltered Flash3D, 91–109% unfiltered xSPECT, 96–103% filtered Flash3D, 95–104% filtered xSPECT).

The cAVH-curves for the cylindrical homogeneous phantom can be found in Fig. 4. The cross-sections of the VOIs considered for this analysis are visualized in Fig. 4a, b and the corresponding curves can be found in Fig. 4c-f. None of the cAVH-curves follow the step function (indicated in grey), but there is better agreement for the filtered data (Fig. 4e, f) compared to the unfiltered data (Fig. 4c, d). In Fig. 4d, the heterogeneous activity concentration distribution resulting from the unfiltered xSPECT-reconstruction (visualized in Fig. 3b, d) is shown. According to this graph more than 5% of the total phantom volume holds less than 0.001 MBq/mL, and more than 4% of the volume contains more than two times the actual activity concentration (0.098 MBq/mL).

Fig. 3

Visualization of the homogeneity in the homogeneous cylindrical phantom. A cross-section of the homogeneous cylindrical phantom is shown in A and B for unfiltered Flash3D and xSPECT-reconstructions, respectively. The locations and the VOIs considered for the visualized activity concentration profiles (ACPs) are also depicted in a and b, the profiles are orthogonal to the cross-sectional plane. The ACPs for the datasets shown in a and b, are plotted in c and d. In e and f, the same ACPs are shown, but for filtered versions of the datasets in a and b. The distinctive appearance of the voxel row ACPs (solid lines; purple, green, cyan) is highly spatially dependent. The ACP that has been average over a radius of 70 mm is considered to match the true homogeneous activity concentration

Fig. 4

(a,b) A cross-section of the homogeneous cylindrical phantom for filtered Flash3D and xSPECT-reconstructions (same slice as in Fig. 3), respectively, including the cross-sections of the VOIs considered in the cumulative activity-volume histograms (cAVHs). (c-f) cAVHs of the homogeneous phantom for the unfiltered data (c,d) and filtered data (e,f).The accuracy of the cAVH-curves, improves for the filtered the data compared to the unfiltered data. Even in the most accurate cAVH, cylindrical VOI with a radius of 70 mm (green dotted line) in e, the activity concentration is underestimated for 58% of the volume and overestimated for 42% of the volume

Activity recovery

The percentage of the recovered activity for each quantification method and measurement of the Jaszczak phantom is visualized in Fig. 5. In general, CF_homogeneous underestimates the amount of activity for measurements performed at high activities and overestimates the amount of activity for measurements at low activities; 28% of the activity is recovered at time point 1 (3130 MBq), and 115% of the activity is recovered at time point 14 (61 MBq). As a direct result of the definition of the CF_self, the recovered activity for this method is always 100%, independent of the activity level. Similarly to CF_homogeneous, the accuracy of the Broad Quantification method is also hampered at the measurements performed at high activities, but in contrast to CF_homogeneous this method does not overshoot the recovered activity at low activities; 57% of the activity is recovered at time point 1, and 97% of the activity is recovered at time point 14.

For CF_homogeneous, only the data acquired with 206–393 MBq present in the phantom results in a maximum deviation of 10% from the true total amount of recovered activity. For the Broad Quantification a maximum deviation of 10% is obtained for data acquired at 61–464 MBq.

Fig. 5

Percentage of recovered activity in the Jaszczak phantom, based on the quantified SPECT/CT images for each measurement time point and quantification method. For CF_homogeneous, the recovery is 28–115% depending on the activity level, and for Broad Quantification it is 57–97%. The recovery for CF_self is always 100%

Activity concentration recovery coefficient

The ACRC for each quantification method as a function of total activity at each measurement time point, for each VOI in the Jaszczak phantom, can be found in Fig. 6. A linear model has been fitted to the ACRC datapoints to indicate the trend over the full range of measured activities. In general, CF_self has the highest ACRCs for the spheres, with the exception of the measurements at the lowest measured activity (time point 14; 61–64 MBq) where the CF_homogeneous overestimates the recovered activity in the whole image, which also results in an overshoot in the ACRC for all VOIs. The Broad Quantification method performs better than CF_homogeneous at high activities (above approximately 2000 MBq), but worse at the lower activities (below approximately 1000 MBq). One contributing factor for this is the lower contrast recovery that is achieved when utilizing the xSPECT-reconstruction method compared to the Flash3D-reconstruction method, see supplementary information. CF_self slightly overestimates the ACRC in the background VOI (by 2–4%) for all measurement time points because of the imperfect contrast recovery, see supplementary information, in combination with the definition of CF_self which forces the true activity to be distributed in the image.

The impact of lower detection efficiency at high activities, as a result of higher dead-time (see Table 1), is also visible in Fig. 6. The ACRC for Sphere 6 and CF_self is however relatively stable across all measurement time points (range: 0.76–0.86), but drops slightly at the lower end of the measured activities, probably because of the lower count rate statistics.

Fig. 6

Activity concentration recovery coefficients, as a function of total activity in the Jaszczak phantom at the measurement time point, for the three quantification methods. An ACRC = 1 is considered ideal (dotted line). A linear model (dashed lines) has been fitted to the data to visualize the trend

Recovery curves

The ACRCs for time point 10, 12, 13 and 14, as a function of sphere diameter, can be seen in Fig. 7. The data has been averaged over the three consecutive repetitions for each of these time points, and are presented by their mean and standard deviation. Recovery curves (3-parameter logistic function) have been fitted to the data and are also shown in the graph. The coefficient of determination (R²) was > 0.99 for all fitted curves. The asymptote value from this fit is presented in Table 2.

Fig. 7

Activity concentration recovery coefficients (ACRCs) as a function of sphere diameter, for each of the three quantification methods (a-c). The ACRCs from the three repeated measurements, of time point 10, 12, 13, and 14 (687–727 MBq, 371–393 MBq, 206–217 MBq, and 61–64 MBq), have been averaged, and the mean and standard deviation are plotted in the graphs. Recovery curves (3-parameter logistic functions) have been fitted to the data, with R² > 0.99 for all curves

Table 2 The a-coefficient from the fitted recovery curves (3-parameter logistic functions) representing the asymptote that the curve would converge to with increasing hot sphere diameter

Discussion

In the presented study we have evaluated the accuracy of three SPECT quantification methods for ¹⁶⁶Ho across a wide activity range. The Flash3D-reconstruction algorithm combined with the self-calibration quantification method (CF_self) resulted in the most accurate activity recovery (Fig. 5) and the highest sphere recovery over the whole activity range (Fig. 6).

No linear relationship between the measured counts per second in the photopeak energy window, and the activity present in the phantom at scanning has been found for our data, see counting rate performance in the Supplementary information. This is due to nonlinearities in the measured count rate with respect to the scanned activity level, dead-time effects and pulse pile-up [4, 21]. This indicates that a conversion factor based on a single ¹⁶⁶Ho activity level (CF_homogeneous) would not be suitable to use for quantification of data at different ¹⁶⁶Ho activity levels (Fig. 2). Furthermore, we have demonstrated that applying CF_homogeneous results in significant underestimation of the recovered activity at high activity levels and overestimation of the recovered activity at low activity levels (Fig. 5). In contrast to a recent guideline [12], we can therefore not recommend this methodology for quantifying ¹⁶⁶Ho-SPECT/CT images for clinical purposes. Additionally, the absolute quantification method (Broad Quantification) also gravely underestimates the recovered activity at high ¹⁶⁶Ho activity levels, but is more accurate than CF_homogeneous at low activity levels (Fig. 5).

In the homogeneous phantom (Fig. 3), we see that the unfiltered xSPECT-reconstruction contains high levels of noise and is therefore not suitable for clinical evaluation of homogeneous activity distributions. After filtering, this noise is reduced but the variation in the voxel values is still large. The noise levels in the unfiltered Flash3D-reconstruction is comparable to the filtered xSPECT-reconstruction, however both reconstruction methods (independent of the filtering parameters) produce images in which the voxel values are highly location dependent and there is hence a large uncertainty in the measured voxel value.

Examining the cAVHs of the cylindrical homogeneous phantom (Fig. 4), we see that the imaged heterogeneity (i.e. the uncertainty in the measured voxel value) is spread out through the whole phantom as the curves for the different VOIs, for each separate reconstruction method and filtering parameter, are very similar. Considering that the true activity concentration in the cylindrical phantom is uniform and that the ¹⁶⁶Ho-SPECT/CT images are inherently heterogeneous, neither of the investigated reconstructions methods can be deemed to produce images which are suitable for voxel-level assessment of homogeneous distributions [22]. An additional drawback with utilizing activity- or dose-volume histograms to assess activity-/dose distributions for clinical assessment is that no spatial information is retained. Considering that microsphere distributions after TARE are inherently heterogeneous [23] and that this information is lost in a dose volume histogram, this metric should be used with caution in radionuclide therapy when evaluating the tumor coverage [24].

From the Jaszczak phantom measurements, we verified the quantification accuracy over a large activity range, determined ACRCs, and constructed the associated recovery curves. Figure 6 indicates poor recovery (ACRC < 0.36) of the two smallest spheres, regardless of quantification method or activity level, making small lesion dosimetry challenging with ¹⁶⁶Ho-SPECT/CT imaging. The visibility of the smallest spheres appears to be worse for ¹⁶⁶Ho-SPECT than for other radionuclides [25,26,27]. The measured ACRCs are however not only dependent on quantification method, radioisotope, and activity level, but also on sphere position within the phantom, $\:\text{S}:\text{B}{\text{G}}_{\text{r}\text{a}\text{t}\text{i}\text{o}}$, PVE compensations, and delineation method. Two of the most frequently used delineation methods are anatomical- (CT-/MRI-based), as used in this study, and threshold-based [28]. Considering all these factors, it is therefore not always easy or even appropriate to directly compare the ACRCs from different studies.

When performing clinical dosimetry, especially when imaging small lesions, the recovery curves in Fig. 7 could be used to correct for PVEs [13]. This type of partial volume correction (PVC) method is region-based, meaning that voxel-level evaluation of the absorbed dose distribution must not be carried out after this correction has been applied [29]. In Fig. 7a (CF_homogeneous), the recovery curve shifts upwards as the activity in the phantom decreases (later time point). This shift is however not a result of better image quality or higher quantitative accuracy at a later time point, but a result of the bias in the recovered activity that is introduced when CF_homogeneous is used to quantify the data. In contrast, Fig. 7b (CF_self) shows that the recovery curves are comparable for time points 10, 12, and 13, but that the curve is shifted downwards for time point 14 as a result of lower count statistics in the acquired data (i.e. higher image noise) and hence lower ACRCs. Unless this lower count statistics can be compensated for by increasing the total scan time, this indicates that clinical dosimetry should not be carried out based on ¹⁶⁶Ho-SPECT imaging with too low activity. The data in Fig. 7c (Broad Quantification) has an overall different shape compared to Fig. 7a, b, with lower ACRCs and these recovery curves are further from convergence. The asymptote of the fitted recovery curves, see Table 2, is nevertheless predicting a higher ACRC for the Broad Quantification method (range: 0.80–1.10) than for the other two quantification methods, but for sphere diameters (hot object size) which are larger than what is currently used in this Jaszczak phantom. For example, when extrapolating the recovery curve for the Broad Quantification method and time point 13 (not shown in the figure), an ACRC of 0.90 (10% loss) is only reached for a spherical diameter of 157 mm which is larger than the average normal liver size [30].

Even though quantifying the ¹⁶⁶Ho-SPECT/CT images using CF_self is the most robust for the total activity recovery with respect to the imaged activity level (Fig. 5), recovery curve PVC is still dependent on the activity level (independent of which quantification method is used), and applying this correction method for clinical data is therefore a rather complex endeavor. Another drawback with recovery curve PVC is that when it is used in combination with CF_self-based quantification, then the total amount of activity in the patient would be overestimated.

The activity level significantly affects the ¹⁶⁶Ho-SPECT/CT image quality. Too high activity leads to limited quantification accuracy due to dead-time effects, while reduced count statistics can be an issue at low activities. Clinically administered ¹⁶⁶Ho activities can exceed 8 GBq [3, 31, 32], and could potentially be even higher when personalized image-guided dose escalation is developed [33]. When personalizing the treatment, the amount of administered activity would not always be known before the start of the procedure and it could be difficult to determine the optimal time point for the SPECT/CT scan upfront. There is therefore a clinical need for limiting the negative effects of a high dead-time on the image quality. TrueCalc has previously been shown to improve the accuracy of the measured activity concentration for technetium-99m (^99mTc) and lutetium-177 (¹⁷⁷Lu) scanned at high count rates [20]. In this study we were however not able to compare the performance of the Broad Quantification with and without TrueCalc (when TrueCalc has been installed on a clinical SPECT/CT scanner it is always active when acquiring xSPECT data in ‘Quantitative Acquisition’’-mode), and can therefore not investigate the effect of TrueCalc for ¹⁶⁶Ho-SPECT/CT imaging as a dead-time correction method. Other dead-time correction methods, based on the acquisition of the wide-spectrum count rate, have been suggested for ¹⁷⁷Lu [34, 35]. However, because some of the high-energy photons of ¹⁶⁶Ho are lying outside of the measurable energy interval (up to about 700 keV, depending on the SPECT/CT scanner), these methods did not seem appropriate for ¹⁶⁶Ho. Our data suggests that performing ¹⁶⁶Ho-SPECT/CT imaging with approximately 200–400 MBq would ensure < 10% error in the total amount of recovered activity, independent of the quantification method used.

To reliably use the CF_self for quantification of clinical ¹⁶⁶Ho-SPECT/CT images, it is important to monitor the activity administered to the patient. The administered activity should be based on a pre-treatment measurement of the activity in the v-vial in a calibrated dose calibrator, and corrected based on the a post-treatment measurement of the residual activity in the v-vial and the delivery set. If such a measurement is not possible, it is reasonable to assume a residual activity of 3.5–6.9% after a TARE procedure with ¹⁶⁶Ho microspheres [36]. Additionally, the value of the activity used in the determination of CF_self should be corrected for the fraction of activity that reaches the targeted region or organ (in cases of shunting). This can be done either based on the ¹⁶⁶Ho-SPECT/CT image or on planar scintigraphy [37]. Another point of concern in clinical SPECT-based dosimetry for TARE is the degenerative effect of respiratory motion on the images, as this can introduce motion induced spill-in and spill-out of counts, as well as misalignment between the nuclear medicine image and the anatomical image [15]. Respiratory induced liver motion alone has been shown to on average decrease the activity recovery and tumor/non-tumor ratios with 24% points in a simulation study of a ^99mTc digital phantom [38]. One option for mitigating this effect is to apply dilation correction of tumor VOIs when assessing the tumor doses [4, 7]. Preferably however, the projection data should be corrected for motion artifacts before being reconstructed [39], but this technique is not available on all SPECT/CT scanners. An even better option for eliminating artifacts originating from respiratory motion would be to acquire the imaging data in breath hold. The acquisition time for ¹⁶⁶Ho-SPECT is currently too long (approximately 20 min acquisition in our clinical protocol) to be able to efficiently include respiratory gating in the clinical workflow. As an alternative to conventional nuclear imaging, MR imaging acquired in breath hold could be used to measure the distribution of ¹⁶⁶Ho microspheres after TARE [9, 10]. MRI has the advantage of being much faster than SPECT (2–3 min for acquisition of a whole liver), as well as having higher resolution (in-plane resolution 2 × 2 mm², slice thickness 4 mm), and being quantitative. Additionally, MRI can also be used for image guidance during the TARE procedure [31].

Conclusion

In conclusion, for clinical dosimetry we recommend to reconstruct the SPECT images using the Flash3D algorithm and quantify the images based on a self-calibration CF. For this it is essential to accurately measure the activity that has been administered to the patient, using a calibrated dose calibrator. The injected activity should also be corrected for possible shunting from the targeted region. This fraction can be estimated either on the ¹⁶⁶Ho-SPECT/CT image or on planar scintigraphy, and should be used to adjust the activity present in the calculation of the CF_self.