Introduction

Alzheimer disease (AD) is the most prevalent cause of dementia, accounting [1] for 60–80% of cognitively impaired patients. Pathological changes in the brain with AD are characterized by extracellular amyloid β (Aβ) aggregates, intracellular tau neurofibrillary tangles, and neuron death [2]. Abnormal proteins such as Aβ and tau accumulate in the brain before symptoms appear. The National Institute of Aging-Alzheimer’s Association (NIA-AA) proposed a research framework to define AD as a biological construct rather than a clinical consequence in its 2018 update. This framework was based on the AT(N) model, which describes patients based on AD-specific biomarkers targeting Aβ (A), tau (T), and neurodegeneration (N) [3]. A more accurate characterization of AD using these biomarkers should provide an understanding of the sequence of events leading to the cognitive impairment that is associated with AD, as well as the multifactorial etiology of dementia.

Tau PET imaging is a more recently developed tool that can minimally invasively assess tau pathophysiology. The distribution of tau neurofibrillary tangles is classified as Braak stages defined at autopsy [4, 5]. Tau PET images have revealed that the cortical uptake of tau PET tracers not only corresponds to a Braak stage, but is also associated with markers of neural injury or cortical gray matter atrophy [6,7,8,9,10]. Tau PET images are useful to predict AD progression for staging because the density, extension, and regional distribution of tau deposits can be determined [11]. Several tau PET tracers such as [18F]flortaucipir (also known as [18F]AV1451), [11C]PBB3, [18F]THK5351, [18F]GTP1, [18F]MK6240, [18F]PI2620, [18F]PM-PBB3, [18F]RO948, and [18F]S-16) have already been assessed in proof‑of‑concept studies [12,13,14,15,16,17,18,19]. The first generation of tau PET tracers [18F]flortaucipir, [18F]THK5351, and [11C]PBB3 were thought to be tau specific, but some binding was off-target [20,21,22,23]. Second-generation tau PET tracers had better affinity and more selectivity that resulted in reduced off-target binding [20]. However, the injected dose, uptake duration, and scan duration varied among these tracers [12,13,14,15,16,17,18, 24, 25]. Images of patients with AD also differed from between first- and second-generation tau PET tracers due to variations in chemical structures and image acquisition parameters [20].

Longitudinal and cross-sectional standardization of tau PET images was therefore required. Thus, the Japanese Society of Nuclear Medicine (JSNM) proposed phantom test procedures and criteria to standardize brain [18F]FDG and amyloid PET imaging [26]. Several studies then determined optimal reconstruction conditions for brain [18F]FDG and amyloid PET imaging using an iterative reconstruction method based on the JSNM phantom test criteria [27, 28]. The FDA approved [18F]flortaucipir as the first tau PET ligand in 2020 [29]. The Molecular Imaging-based Precision Medicine Task Group published an international consensus regarding [18F]flortaucipir PET imaging protocols and results for clinical purposes [30]. Although the Task Group recommended three-dimensional ordered-subset expectation maximization (3D-OSEM) as the reconstruction algorithm for [18F]flortaucipir PET images, reconstruction parameters such as iterations and subsets were not mentioned. The Alzheimer’s Disease Neuroimaging Initiative (ADNI)-3 provided reconstruction parameters for [18F]flortaucipir PET imaging for PET scanners from all vendors. However, they proposed identical reconstruction conditions for [18F]FDG, amyloid, and [18F]flortaucipir imaging [31]. Tau PET imaging to evaluate the progression of tau tangles should be able to detect early tau deposition in the medial temporal lobe and define an accurate threshold of tau positivity [5, 20, 32,33,34]. Therefore, the reconstruction conditions for tau PET imaging using [18F]FDG, or amyloid should not be identical. They should be concomitantly optimized with scan protocols for tau PET tracers. However, the methodology of phantom test to determine and optimize reconstruction conditions for tau PET imaging has not been addressed compared with brain [18F]FDG or amyloid PET imaging.

This Japanese multicenter study aimed to develop an objective methodology for phantom test protocols that could determine dedicated reconstruction parameters to standardize tau PET imaging. We developed an analytical volume of interest [1] template based on the pathophysiological characteristics of tau deposition. We then validated it using phantom test procedures and a VOI analytical method, then optimized the reconstruction conditions for tau PET imaging.

Materials and methods

Determination of phantom conditions

Here, we initially estimated the activity of [18F]flortaucipir, [18F]THK5351, and [18F]MK6240 tau PET tracers in the brain with reference to published scan parameters [35,36,37]. The conventional injected doses (MBq) and uptake durations (min) were 370 and 70, 185 and 40, and 185 and 90 for [18F]flortaucipir, [18F]THK5351, and [18F]MK6240, respectively [35,36,37], and the acquisition duration was 20 min for all three tracers. The estimated whole brain activity at the start of PET acquisition was 4.0, 1.0, and 1.5 MBq for [18F]flortaucipir, [18F]THK5351, and [18F]MK6240, respectively. The standard brain volume was equivalent to 1,200 mL. The activity concentration in a normal brain was 3.33, 0.83, and 1.25 kBq/mL for [18F]flortaucipir, [18F]THK5351, and [18F]MK6240, respectively.

We then calculated activity in the phantom based on estimated brain activity. The distribution of tau PET tracers in the brain corresponded to that of tau deposition defined in terms of Braak stages [4, 5]. The accumulation of tau in the entorhinal cortex or hippocampus due to early AD pathology appears as a hot region in tau PET imaging. We hypothesized that a Hoffman 3D brain phantom (Data Spectrum Corporation, Hillsborough, NC, USA) could simulate the local tau deposition. Therefore, we considered that the target (local accumulation) and reference regions could be simulated with the Hoffman phantom and a cylindrical phantom (Itoi Plastics Co. Ltd., Kobe, Japan), respectively, containing different amounts of activity. The average activity concentration of the three tau PET tracers in the whole brain was ~ 2.0 kBq/mL. The volume of the cylindrical phantom and the amount of activity in it were 6 L and 12.0 MBq, respectively, at the start PET image acquisition. The concentration of activity in the Hoffman phantom was taken as the standardized uptake value ratio (SUVR) in the medial temporal lobe. The SUVR of the medial temporal lobe or hippocampus calculated from the activity concentrations of [18F]flortaucipir, [18F]THK5351, and [18F]MK6240 in the cerebellar cortex as a reference region was ~ 2.0 [38,39,40,41,42,43,44,45]. Thus, the concentration of activity in the Hoffman phantom was twice that in the cylindrical phantom. The volume of the Hoffman phantom and the activity in it were 1.2 L and 4.8 MBq, respectively, at the start of PET acquisition.

Characteristics of equipment at four sites

Radioactivity in the phantoms was quantified using two brands of dose calibrators (Nippon RayTech Co., Ltd., Tokyo, Japan and Capintec Inc., Ramsey, NJ, USA), and four PET/computed tomography (CT) scanners (one from GE Healthcare, Milwaukee, WI, USA and three from Siemens Healthineers, Erlangen, Germany). Table 1 shows the PET/CT systems and dose calibrators. Reconstruction conditions except for the iterations and the Gaussian filter (Table 1) proceeded under the clinical conditions for brain PET examinations at each site. The performance of the PET/CT scanners has been described elsewhere [46,47,48,49].

Table 1 Characteristics of equipment at four sites
Full size table

Phantom experiment

Computed tomography images were acquired from all scanners to correct attenuation, scatter, and other issues except for the point-spread function. Thereafter, images were acquired from the Hoffman and cylindrical (inner diameter, 16 cm; inner length, 30 cm) phantoms initially containing 4.8 and 12.0 MBq of [18F]FDG, respectively, for 20 min based on the protocols for [18F]flortaucipir, [18F]THK5351, and [18F]MK6240.

Creation of a VOI template for Hoffman phantom images

We created a VOI template (Fig. 1) to optimize tau PET images based on Braak stages using PMOD v. 3.8 (PMOD Technologies LLC, Zurich, Switzerland). Each VOI was placed according to the distribution of tau pathology in patients with AD. The locations and amounts of voxels in the inferotemporal cortex, lateral temporal lobe, precuneus, white matter (WM), and cerebellar cortex were 357, 375, 365, 576, and 769 voxels, respectively. The inferotemporal cortex, lateral temporal lobe and precuneus comprised the gray matter (GM) VOI and the cerebellar cortex was the reference VOI. The VOI template was created based on the digital Hoffman phantom and then coregistered to the CT image acquired by the Discovery MI.

Fig. 1
figure 1

Volume-of-interest template for phantom tau PET images. Each volume of interest indicates inferotemporal cortex (red), lateral temporal lobe (green), precuneus (blue), white matter (yellow), and cerebellar cortex (light blue)

Full size image

Image registration

The CT images acquired by Biograph 16, Biograph mCT Flow, and Biograph Vision were first coregistered to their respective CT images acquired by Discovery MI. Then, the PET images from each scanner were coregistered to their corresponding CT images. The aligned PET coordinates were matched with the coordinates of the VOI template. Finally, the aligned VOI template was applied to each coregistered PET image for VOI analysis.

Image analysis

Determination of iteration number

The data acquired from Hoffman phantom images were reconstructed with one to ten iterations and no post-filter. Mean activity concentrations in the GM and WM were measured using the VOI template. The ratio of gray-to-white matter contrast (contrast [%]) was calculated as:

$${\text{Contrast}} \left( \% \right) = \frac{{\left( {GM_{p} /WM_{p} - 1} \right)}}{{\left( {GM_{d} /WM_{d} - 1} \right)}} \times 100,$$

where GMp and WMp in the brain, and GMd and WMd in the digital Hoffman phantom PET images are GM and WM activities, respectively, in VOIs. The GMd and WMd values provided a true gray-to-white ratio of 4 and were applied to the image coregistered to the digital phantom.

The recovery coefficient (RC) in GM and WM was defined as the image-derived mean activity concentration determined as contrast divided by the activity concentration of the stock solution in the Hoffman phantom. The RC at the GM and WM was calculated as:

$$RC_{GM} = \frac{{GM_{p} }}{{{\text{activity}} {\text{concetration}} {\text{in }}GM}},$$
$$RC_{WM} = \frac{{WM_{p} }}{{{\text{activity}} {\text{concetration}} {\text{in}} WM}}.$$

The activity concentration in the GM was derived from the net phantom activity (measured using a dose calibrator at each site) divided by the fillable volume (1.14 L) of the Hoffman phantom. The activity concentration in WM was 25% of that in GM.

Contrast and RCs in GM and WM were measured using PMOD v. 3.8. The convergence of contrast or RC in one to ten iterations was defined as optimal iteration for tau PET imaging.

Determination of Gaussian filter magnitude

Data acquired from the cylindrical phantom was reconstructed using the optimal number of iterations determined as contrast (%), RC, and Gaussian filter magnitudes of 0‒10 mm at full width at half maximum (FWHM). A circular ROI (13 cm diameter) was placed on the center of the cylindrical phantom image (Fig. 2) to evaluate noise as a coefficient of variation (CV) calculated as:

$$CV \left( \% \right) = \frac{SD}{{Mean}} \times 100,$$

where SD is the standard deviation of ROI activity, and Mean is the mean activity within a circular ROI. The optimal magnitudes of Gaussian filters were determined from CV < 15%.

Fig. 2
figure 2

Region of interest to calculate image noise

Full size image

Results

Numbers of iterations

Figure 3 shows contrast in the four scanners as a function of the number of iterations. Contrast increased along with iteration numbers and converged at ≥ 4 iterations. Contrast at convergence points was 85% and 70–75% in two scanners each.

Fig. 3
figure 3

Contrast as a function of iterations in four scanners

Full size image

Figure 4 shows the RC in four scanners as a function of iteration numbers. The RCs at GM and WM, respectively, increased and decreased with increasing numbers of iterations. The RCs in GM and WM converged iteration ≥ 4. The RC in GM underestimated (RC = 1.0), whereas that in WM overestimated (RC = 0.25) the true activity. The RCs at GM and WM in each one scanner were converged to a different value. The RC error from true activity was −15% for GM and 15% for WM. Thus, optimal iterations were 4 for all scanners.

Fig. 4
figure 4

Recovery coefficients (RC) as functions of iterations in four scanners. a Gray matter (GM). b White matter (WM). Dashed lines, true activity (RC) = 1.0 in GM and 0.25 in WM

Full size image

Determination of Gaussian filter magnitude

Figure 5 shows image noise in four scanners as a function of Gaussian filter magnitudes. The CV decreased as the filter magnitude increased. Two images each acquired using two scanners each had less (CV = 10–15%) and more (CV = 25%) image noise. Images from two scanners had CVs < 10% when the Gaussian filter magnitude was 2 or 4 mm at FWHM. Thus, optimal FWHM of Gaussian filter were 4 mm for Biograph 16 and Biograph mCT Flow and 3 mm for Biograph Vision and Discovery MI.

Fig. 5
figure 5

Magnitude of Gaussian filter in four scanners

Full size image

Phantom images

Figure 6 shows the physical indices of brain and cylindrical images acquired using four scanners under tau-specific optimized reconstruction conditions. Contrast and CV were 67.2% and 8.6% for the Biograph 16, 73.8% and 7.3% for the Biograph mCT Flow, 75.3% and 4.9% for the Biograph Vision, and 70.2% and 6.9% for the Discovery MI, respectively. The images have good contrast and low image noise.

Fig. 6
figure 6

Brain and cylindrical phantom images acquired using four scanners and reconstructed under tau-specific conditions. Biograph 16 (a; It, 4; GF, 4 mm), Biograph mCT Flow (b; It, 4; GF, 4 mm), Biograph Vision (c; It, 4; GF, 3 mm), and Discovery MI (d; It, 4; GF, 3 mm). It iteration, GF Gaussian filter

Full size image

Discussion

The methodology required to determine reconstruction conditions for tau PET imaging has not been established and standardized. We simulated the brain activity of three tau PET tracers using brain and cylindrical phantoms. We developed a VOI template to analyze tau PET images based on the pathophysiological distribution of tau in the brain. We determined the number of iterations and Gaussian filter parameters as two of the reconstruction conditions for different PET/CT scanners at four Japanese institutions.

We estimated the brain activity of [18F]flortaucipir, [18F]THK5351, and [18F]MK6240 from published data at the start of image acquisition to define phantom conditions. The process referred to JSNM phantom test procedures to define brain [18F]FDG activity using four amyloid PET tracers [26]. We used brain uptake and kinetic information from dosimetry studies of several tau PET tracers to estimate the activity of tau PET tracers in the brain [35,36,37]. We determined the activity of the three tau PET in the cylindrical and Hoffman phantoms were 2.0 and 4.0 kBq/mL, respectively. However, experimental conditions from phantoms should be established for each individual tau PET tracer. If images are acquired using more than one tau PET tracer, several experiments are needed to optimize the reconstruction condition for each of them. Several complex experiments can lead to measurement error while concomitantly exposing operators to needless amounts of radiation. We considered that the phantom experiment to determine the activity condition of three tau PET tracers was reasonable and could be generalized to standardize tau PET imaging protocols. The convergence of PET images using an iterative reconstruction algorithm depends on the target activity, the acquired PET counts, and the target size or shape. We previously found that the convergence rate of contrast was independent of target activity because the conditions of the iterative reconstruction algorithm were the same regardless of phantom activity [27, 28]. The high levels of whole-brain activity in tau PET images in first-generation tau PET tracers due to non-specific brain uptake were higher than the second-generation tau PET tracers. The activity of 2.0 kBq/mL herein was in the mid-range of brain activity in both generations of tau PET tracers. The [18F]flortaucipir and [18F]MK6240, which had the highest and lower activities, respectively, in this study, are currently being used in clinical examinations or practice. However, as second-generation tracers may eventually supersede the first-generation tracers, adjustments in phantom activity may be necessary. Therefore, for the time being, we recommend including tracers with higher brain activity, such as [18F]flortaucipir, in the first-generation tau PET tracer simulations. Our phantom experiment objectively determined appropriate reconstruction conditions for tau PET images because mid-range activity can be applied to more recent tau PET tracers. The activity in the Hoffman phantom containing 4.0 kBq/mL was equivalent to an SUVR of ~ 2.0 at the medial temporal lobe and hippocampus in a patient with AD [38,39,40,41,42,43,44,45]. The role of tau PET imaging is to define the density and extent of tau deposition in patients with AD. Therefore, tau PET imaging should contribute to the early diagnosis of AD based on the detection tau deposition in the entorhinal cortex or hippocampus corresponding to an early Braak stage [32,33,34]. Tau pathology visualized as [18F]flortaucipir accumulation was visually undetectable in Braak stage I–III [50, 51]. The cutoff SUVR required to distinguish Braak stages I–III from IV is 1.29 [51]. Second-generation tau PET tracers could track longitudinal tau accumulation in asymptomatic and symptomatic AD [52, 53]. The cutoff of SUVR in second generation tau PET tracers to distinguish AD from non-AD is 1.35 [54]. We propose a phantom activity rate of 2.0. However, we plan to re-evaluate the phantom activity to validate the early detection of tau accumulation.

Our new tau-specific VOI template for analyzing phantom tau PET images based on the Braak stages covered the inferotemporal cortex, lateral temporal lobe, precuneus, white matter (WM), and cerebellar cortex that are frequently characterized by tau deposition in patients with AD. The six Braak stages were categorized based on the distribution and developmental sequences of lesions. The tau PET-specific VOI corresponded to Braak stages I/II, III/IV, and V/VI that were anatomically defined as trans-entorhinal, limbic, and isocortical. Other clinical studies have analyzed tau PET images using an AD-signature temporal meta-ROI or an anatomical definition of VOI also based on Braak stages [42, 45, 55,56,57,58]. Dore et al. developed a universal cortical tau mask comprising the tau PET ligands [18F]flortaucipir, [18F]GTP1, [18F]MK6240, [18F]PI2620, [18F]PM-PBB3, and [18F]RO948 to evaluate tau accumulation during the AD continuum in areas common to tau PET ligands [59, 60].

Optimal reconstruction conditions for tau PET images from four PET/CT scanners were determined as contrast, RC, and image noise calculated from brain and cylindrical phantom images. Images of the brain and cylindrical phantom were acquired with sufficient image quality for visual assessment by physicians. Contrast and RC converged at ≥ 4 iterations regardless of scanner generation or vendor. The convergence rate was equivalent to that in a previous study because it was independent of phantom activity, locations of quantified activity in the phantom, and the shape of the VOI template [27]. Brain PET imaging in dementia has been compared longitudinally or cross-sectionally among patients or with a reference database [9]. The recent Research for life (EARL) study (EANM Research Ltd., Vienna, Austria) investigated the harmonization of image quality and quantitative capability in clinical brain PET images [61]. We used the RC proposed in the EARL study as a quantitation index to harmonize PET images and validate our method. The RCs in GM and WM were 0.88–1.01 (true value, 1.00) and 0.25–0.27 (true value, 0.25), respectively. Our quantitative capability surpassed that of an earlier study [61]. In particular, the high-end SiPM-PET/CT Biograph Vision scanner, allowed precise quantitation due to the spatial and temporal resolution being better than that of PMT–PET [48]. The RC in the present study decreased the effects of spill-in or -out because our VOIs were separated between GM and WM, unlike those in an earlier study [61]. The image noise (CV) achieved < 10% with Gaussian filters set at 2–4-mm (at FWHM) in four scanners. The acceptance criteria of image quality in the JSNM phantom test procedure for [18F]FDG and amyloid PET imaging has been defined as CV < 15% [26]. Tau PET imaging should be able to detect local tau deposition and warm tau tracer accumulation in small structures or tissues such as the medial temporal lobe [5, 20, 32,33,34]. Brain activity and PET counts were lower than those of [18F]FDG and [18F]florbetapir [26] in tau images, especially when measured using second generation tracers. Noise on tau PET images caused by statistically fewer PET counts hindered the early detection of local or warm tau depositions. We proposed that image noise should be a CV ≤ 15%.

The image quality and impression of the image using optimized reconstruction conditions were different among these scanners. The difference that led to affecting the outcome of interpretation in tau PET imaging was unknown because clinical images were not assessed in this study. However, clinical studies using tau PET imaging have used a number of scanners [32,33,34]. The affecting result of different resolution may be low or nothing. We argue that reconstruction conditions need to be optimized to educe scanner performances for tau PET imaging using phantom study.

The present study has some limitations. We did not investigate reconstruction conditions other than iterations and Gaussian filters, and we applied clinical conditions for standard brain PET imaging at each site. The Molecular Imaging-based Precision Medicine Task Group and ADNI-3, respectively, recommended pixel sizes of 2–4 and < 2.0 mm to acquire [18F]flortaucipir PET images [30, 31]. The pixel size in four scanners was equivalent to that in previous studies and it was also appropriate in the present study. When pixels ≤ 1.0 mm were used to detect early tau deposition, the Gaussian filter was adjusted to suppress image noise. Another limitation of the present study is that the Hoffman phantom simulated the distribution of flow or metabolic tracers in the brain [62]. However, the distribution of tau deposition differs from the flow of metabolic tracers. A tau-specific phantom should be developed to determine optimal qualitative and quantitative reconstruction conditions for tau PET imaging. We selected the Hoffman phantom because it is widely available.

Conclusions

We estimated the activity of 4.0 and 2.0 kBq/mL in brain and cylindrical phantoms based on the brain activity of some tau PET tracers that were used in previous tau PET imaging studies. We developed a tau-specific VOI based on tau deposition corresponding to Braak stages. We optimized the reconstruction parameters of iteration numbers and the Gaussian filter magnitude using tau PET images acquired from a phantom by four PET/CT scanners. The image quality and quantitative capability were sufficient under our conditions.