mGluR5 and GABAA receptor‐specific parametric PET atlas construction—PET/MR data processing pipeline, validation, and application

Abstract The glutamate and γ‐aminobutyric acid neuroreceptor subtypes mGluR5 and GABAA are hypothesized to be involved in the development of a variety of psychiatric diseases. However, detailed information relating to their in vivo distribution is generally unavailable. Maps of such distributions could potentially aid clinical studies by providing a reference for the normal distribution of neuroreceptors and may also be useful as covariates in advanced functional magnetic resonance imaging (MR) studies. In this study, we propose a comprehensive processing pipeline for the construction of standard space, in vivo distributions of non‐displaceable binding potential (BP ND), and total distribution volume (V T) based on simultaneously acquired bolus‐infusion positron emission tomography (PET) and MR data. The pipeline was applied to [11C]ABP688‐PET/MR (13 healthy male non‐smokers, 26.6 ± 7.0 years) and [11C]Flumazenil‐PET/MR (10 healthy males, 25.8 ± 3.0 years) data. Activity concentration templates, as well as V T and BP ND atlases of mGluR5 and GABAA, were generated from these data. The maps were validated by assessing the percent error δ from warped space to native space in a selection of brain regions. We verified that the average δABP = 3.0 ± 1.0% and δFMZ = 3.8 ± 1.4% were lower than the expected variabilities σ of the tracers (σABP = 4.0%–16.0%, σFMZ = 3.9%–9.5%). An evaluation of PET‐to‐PET registrations based on the new maps showed higher registration accuracy compared to registrations based on the commonly used [15O]H2O‐template distributed with SPM12. Thus, we conclude that the resulting maps can be used for further research and the proposed pipeline is a viable tool for the construction of standardized PET data distributions.

a reference for the normal distribution of neuroreceptors and may also be useful as covariates in advanced functional magnetic resonance imaging (MR) studies. In this study, we propose a comprehensive processing pipeline for the construction of standard space, in vivo distributions of non-displaceable binding potential (BP ND ), and total distribution volume (V T ) based on simultaneously acquired bolus-infusion positron emission tomography (PET) and MR data. The pipeline was applied to [ 11 C]ABP688-PET/MR (13 healthy male non-smokers, 26.6 ± 7.0 years) and [ 11 C]Flumazenil-PET/MR (10 healthy males, 25.8 ± 3.0 years) data. Activity concentration templates, as well as V T and BP ND atlases of mGluR 5 and GABA A , were generated from these data. The maps were validated by assessing the percent error δ from warped space to native space in a selection of brain regions. We verified that the average δ ABP = 3.0 ± 1.0% and δ FMZ = 3.8 ± 1.4% were lower than the expected variabilities σ of the tracers (σ ABP = 4.0%-16.0%, σ FMZ = 3.9%-9.5%). An evaluation of PET-to-PET registrations based on the new maps showed higher registration accuracy compared to registrations based on the commonly used [ 15 O]H 2 O-template distributed with SPM12. Thus, we conclude that the resulting maps can be used for further research and the proposed pipeline is a viable tool for the construction of standardized PET data distributions. The advantages of simultaneously acquired positron emission tomography (PET) and magnetic resonance imaging (MR) data have been extensively reported in numerous previous studies (Herzog et al., 2011;Herzog, 2012;Catana, Drzezga, Heiss, & Rosen, 2012;Torigian et al., 2013). Simultaneous PET/MR brain imaging enables the combination of in vivo data relating to neuroreceptor systems (obtained from PET) with anatomical and structural information acquired under exactly the same conditions using multiparametric MR (Sander, Hansen, & Wey, 2020). This application provides a perfect basis for the construction of specific neuroreceptor distribution maps.
Brain atlases showing standard distributions of the most important neuroreceptors in the healthy brain may aid the study of the molecular mechanisms underlying psychiatric conditions. Furthermore, such maps may also be useful as covariates in functional MR (fMRI) studies.
Indeed, an association between the fMRI signal, the relative receptor occupancy, and the level of neurotransmitter (concretely dopamine) has been demonstrated previously (Mandeville et al., 2013). Thus, consideration of receptor availability provides additional information for more advanced analyses of fMRI studies.
In this study, we aimed to construct atlases showing parametric total volume of distribution (V T ) and non-displaceable binding potential (BP ND ), in conjunction with normalized activity concentration templates, based on bolus-infusion PET and simultaneously acquired MR data. For this purpose, we designed a coherent PET/MR neuroimaging processing pipeline in NiPype (Gorgolewski et al., 2011). Based on data from healthy subjects, the pipeline was then used to establish in vivo maps of V T and BP ND for use as a reference in advanced studies of psychiatric and neurologic diseases. The pipeline was further used in the creation of [ 11 C] ABP and [ 11 C]FMZ activity concentration templates to provide tracerspecific target templates for PET-to-PET registration.
[ 11 C]ABP688 is a recently developed radioligand that binds to the allosteric site of the metabotropic glutamate receptor subtype 5 (mGluR 5 ; Ametamey et al., 2007). Previous investigations have shown sufficient test-retest reliability for this radiotracer (Smart et al., 2018) and have demonstrated its ability to detect physiological changes in endogenous glutamate levels (DeLorenzo et al., 2011). Glutamate is the main excitatory neurotransmitter in the brain (Meldrum, 2000), and disturbances in the glutamatergic system are hypothesized to be involved in the development of numerous psychiatric and neurological diseases, including Parkinson's disease, depression, anxiety, and schizophrenia (Niswender & Conn, 2010).
[ 11 C]Flumazenil is a well-established and widely used radiotracer that binds to the benzodiazepine binding site of the γ-aminobutyric acid class A (GABA A ) receptors (Odano et al., 2009). GABA is the main inhibitory neurotransmitter in the brain (Petroff, 2002), and disturbances in GABAergic neurotransmission is associated with several psychiatric diseases such as major depressive disorder, schizophrenia, and bipolar disorder (Chiapponi, Piras, Piras, Caltagirone, & Spalletta, 2016), as well as anxiety disorders, epilepsy, and insomnia (Möhler, 2006).
As the two selected ligands (in the following abbreviated as [ 11 C] FMZ and [ 11 C]ABP) are suitable for the investigation of the fundamental inhibitory and excitatory neurotransmitters involved in several psychiatric and neurologic diseases, the generation of atlases for the corresponding receptors is of high scientific importance. In this work, we demonstrate the validity of our pipeline for the creation of BP ND and V T maps and evaluate whether the resulting atlases reflect accurate parametric values by performing region-of-interest (ROI) analyses in a native space and a template space.
We also investigate the effect of omitting the parameter estimation step from the proposed pipeline on the generation of normalized activity concentration templates for use in direct PET-to-PET registration, as this would be particularly beneficial for PET imaging applications that focus on specific neurotransmitter systems. Direct PET-to-PET normalization is usually necessary for conducting group-wise evaluation when PET data are acquired without an MR image.
Although hybrid PET/MR systems are gaining relevance in research and have significant diagnostic advantages over the more commonly used PET-CT (computed tomography; Von Schulthess & Schlemmer, 2009;Zaidi, Mawlawi, & Orton, 2007), they remain rare in clinical applications (Ehman et al., 2017 Rosa et al., 2014). However, both of the aforementioned templates primarily show the distribution of gray matter, which barely coincides with the distribution of neuroreceptors. Therefore, the use of specific templates is more beneficial, as already demonstrated for two carbonyl-11 C-labeled tracers-[ 11 C]WAY-100635 and [ 11 C]Raclopride (Meyer, Gunn, Myers, & Grasby, 1999). Furthermore, a dynamic 4D [ 11 C]Raclopride template has been constructed based on PET data acquired using a high-resolution research tomography (Bieth, Lombaert, Reader, & Siddiqi, 2013). The usefulness of these types of registration templates has thus been demonstrated for the group-wise evaluations of PET datasets (Della Rosa et al., 2014;Meyer et al., 1999)

| Subject selection and data acquisition
Twenty-three subsets of data from two previous studies conducted in our institute using a hybrid BrainPET/MR scanner (Herzog et al., 2011) were finally considered for the construction of the novel atlases, and a subset from a larger study investigating the role of the mGluR 5 in schizophrenia was used for the generation of the [ 11 C]ABP maps. At the time of data analysis, 15 healthy male non-smokers could be considered for further processing. Another subset of data from this study was recently analyzed and published (Régio Brambilla et al., 2020). The [ 11 C]FMZ data originate from a study of 20 healthy male participants, which aimed to optimize the bolus-infusion scheme for the [ 11 C]FMZ acquisitions   Elmenhorst et al., 2016). In both studies, the PET tracer was injected as a bolus plus constant infusion (B/I). Information about the mean injected activity, mean age of the participants, bolus fraction (Kbol), and the acquisition time for both studies is given in Table 1. The simultaneous PET/MR data acquisition protocol involved the concurrent acquisition of PET data in list-mode with structural and functional MR data acquisition. Structural MR data were acquired immediately after the bolus injection, before the tracer reached the equilibrium state. Once the tracer was expected to have reached the steady-state, a "resting statetaskresting state" paradigm was applied as described in Neuner et al. (2018). In the case of [ 11 C]FMZ, the considered list-mode PET data were framed at 20 Â 5 min, and 20 Â 2 min framing was used to reconstruct the first 40 min of the [ 11 C]ABP data. The framed data were reconstructed into a 256 Â 256 Â 153 image matrix with a 1.25 Â 1.25 Â 1.25 mm isometric voxel size, using 3D ordinary Poisson-ordered subset expectation maximization (3D OP-OSEM) with 32 iterations and two subsets (Zhang et al., 2014). To eliminate any effects, the tasks in the measurement protocols may have had on the data, only frames up to the first task phase were considered for the purposes of this work. Thus, it was not necessary to reconstruct frames for the entire duration of the measurements. This is further explained in Section 2.2.2. During reconstruction, attenuation correction was applied based on an initially acquired MR image. Details relating to this template-based attenuation correction method are discussed elsewhere (Kops & Herzog, 2008). Additionally, corrections for dead time, decay, random coincidences (variance reduction of randoms), and scatter were applied.
The T 1 -weighted, structural MR images were acquired with a magnetically prepared rapidly acquired gradient echo (MPRAGE) sequence in 176 sagittal slices of 1 mm thickness with the following MR parameters: repetition time (TR) = 2,000 ms, echo time (TE) = 3.03 ms, flip angle α = 9 , GRAPPA factor = 2.

| Data preparation
The ECAT7 PET files and DICOM MR files were converted into the common NifTi file format using dcm2niix (Li, Morgan, Ashburner, Smith, & Rorden, 2016). This file format is readable with all of the relevant software that tools that were used in the pipeline. Two preliminary steps were applied: first, a brain extraction step was performed to increase the accuracy of the nonlinear registration step. Second, the PET files underwent a time-activity curve (TAC) analysis to ensure that the equilibrium condition for accurate estimation of BP ND and V T was met.

| Brain extraction
In order to achieve maximal registration accuracy with the advanced normalization tools (ANTs) method (Avants et al., 2011), the use of skull-stripped T 1 -weighted images is recommended to estimate the optimal registration into template space (Pustina & Cook, 2017). Several skull-stripping methods are established in neuroimaging. However, none of the tools tested in the course of this work (FSL BET, ANTs Brain Extraction, SPM NewSegment) were able to consistently output accurate brain extractions when using one fixed set of parameters. Thus, instead of individually optimizing the parameters to reach the needed performance of the scripts, the binary brain masks produced by the ANTs brain extraction shell script were each edited manually in the FSL image viewer FSLeyes (McCarthy, 2020) to ensure a minimal amount of missing or non-brain tissue voxels. An example of manual correction for one [ 11 C]FMZ subject is shown in Figure 1.

| TAC evaluation
For the estimation of BP ND and V T in the later stages of the pipeline, the equilibrium of the tracer in plasma and tissue must be already established. To ensure that a constant equilibrium at the time of interest was reached, TACs were plotted and analyzed prior to the pipeline T A B L E 1 Additional information regarding the groups of volunteers as well as the respective acquisition time (A T ), administered dose (A D ), and bolus fraction K bol in the two considered studies processing. In both datasets, the start of the equilibrium was expected to occur approximately 30 min after injection (Neuner et al., 2018;Régio Brambilla et al., 2020). The TACs in the whole gray matter (GM) and the respective reference regions were normalized to the mean activity in GM during the expected equilibrium phase. Previous studies have shown that the GM of the cerebellum is a suitable reference region for [ 11 C]ABP (Akkus et al., 2013;Régio Brambilla et al., 2020;Smart et al., 2018). The definition of the cerebellar GM in the Neuromorphometrics brain atlas included in SPM12 was used for BP ND calculation in the ABP subjects. The data used to create this atlas originated from the OASIS project (www.oasis-brains.org), and the labels were provided by Neuromorphometrics, Inc., under academic subscription (www.neuromorphometrics.com).
In the case of [ 11 C]FMZ, the pons is a brain region with a negligible amount of GABA A receptors (Odano et al., 2009). Here, a manually drawn mask was created in FSLeyes, guided by the use of the MNI space template referred to in Section 2.3.2. The fuzzy edges of this mask were reduced by eroding it and then dilating it using the same spherical kernel for both operations.
Only scans that contained frames with a deviation below 10% of normalized activity concentration during the time of the first restingstate phase were accepted for atlas construction. The TACs of the subjects finally considered, including the respective selected frames, are given in the Supporting Information Material.
A further purpose of the TAC evaluation was to find the frames in which the equilibrium condition was already established and coincided with the first resting-state phase of the protocol mentioned in Section 2.1. Frames were accepted if they covered more than 50% of resting-state acquisition. This way, any influence of the following tasks on the PET data could be ruled out. During resting-state acquisition, the subjects were instructed to close their eyes and to stay calm without thinking about anything specific.
Following initial visual inspection, TAC evaluation, and motion correction, two subjects of the [ 11 C]ABP study, and 10 subjects of the [ 11 C]FMZ study were excluded from further consideration leaving a total of 10 subjects from the [ 11 C]FMZ study and 13 healthy male non-smokers from the [ 11 C]ABP dataset.

| NiPype image processing pipeline
The data for the construction of the PET atlases were pre-processed using one coherent pipeline of processing steps that were assembled in NiPype ( Pypes workflows as a basis for parameter estimation. In the approach proposed here, emphasis was placed on an optimized selection of preprocessing steps to minimize error sources. This selection will be described in detail in the following paragraphs. The image processing pipeline consists of three major parts: (a) MR and PET data pre-processing; (b) estimation of the transformation into a standardized space and; (c) PET parameter calculation. A detailed graph illustration of the implemented pipeline is given in Figure 2.
To be able to apply the pipeline, the user is required to install the software requirements mentioned above and must simply provide the following.
1. The raw structural MR scans in addition to the respective binary brain masks.
2. The reconstructed and TAC evaluated PET data.

A list of tuples, defining the frames of interest for each subject
(starting frame no., no. of frames). Currently, the pipeline expects all neuroimaging data in NifTi format.

(Optional) For
Additionally, the FWHM of a Gaussian kernel for PET smoothing can be set, and there is a choice of whether or not to apply partial volume correction (PVC). With these inputs given, the processing steps outlined in the following sections were run fully automatically. For the purposes of this work, the FWHM is set to 2.5 mm and PVC is included.

| MR and PET data pre-processing
Raw PET scans were initially motion corrected using the SPM12 routine "Realign." The pipeline is set to apply motion correction with respect to the first frame. The frames of interest were extracted from the motion-corrected frames according to the input tuple, smoothed (SPM; Gaussian kernel sized 2.5 mm FWHM), and averaged (FSL).
After the bias-correction of the raw structural MR scans using the ANTs N4 bias correction method, with the option to normalize the intensity range set to true (Tustison et al., 2010), the MR image and the input brain mask were co-registered to the previously generated average PET frame by applying the SPM12 co-registration method.
The last step of this pipeline workflow is PVC. Therefore, the coregistered MR images were segmented using SPM12 NewSegment, and the resulting tissue probability maps were given to an implementation of the RBV + Labbé approach within the PETPVC package.

| Transformation estimation
From the previous steps, the bias-corrected and co-registered MR images and brain masks were taken as input to the estimation of the transformations into the standard space. The masks were used to extract the brain segments from the whole-head MR scan. The extracted brain images were then given as input to the ANTs registration method, which was set to calculate optimal transformations in a rigid, then an affine, and lastly a nonlinear symmetric normalization (SyN)

| PET parameter calculation
The advantage of the applied B/I infusion protocol is that it enables the estimation of BP ND and V T using a simple ratio method. The B/I acquisition protocol has been previously validated (Carson, 2000) and optimized for [ 11 C]ABP (Burger et al., 2010) and [ 11 C]FMZ . Metabolite correction of venous blood plasma was performed following the method described in Mauler et al. (2020). At true equilibrium, it is possible to calculate V T and BP ND according to Equations (1) (Carson et al., 1993) and (2) using the activity concentration in tissue C T , plasma C P , and the non-displaceable tissue concentra- The estimations of BP ND and V T were performed in subject space before the transformation into MNI space was applied. For BP ND cal- to avoid the inclusion of these bordering voxels into the analysis, as well as for restricting the masks to voxels with a high probability of being GM. The output was then used to calculate BP ND after (2). The same was performed for V T following (1).

| Pipeline outputs and template construction
To aid quality control, the pipeline was set to output several intermediate results: motion correction parameters, bias corrected and coregistered MR, resulting tissue segments, mean PET frames of interest). In addition, the forward and inverse transforms estimated during the pipeline execution were also output, as well as the normalized activity concentration and parametric versions of the PET images in native and template space.
The templates were thus constructed by averaging the resulting files in template space. The BP ND and V T versions of the templates were masked to remove irrelevant voxels. This was achieved by using a dilated version of the binary brain mask that is distributed in addition to the MNI template used for this work. The dilation kernel was set to a spherical kernel with a radius of 2.5 mm to ensure that no relevant information in the templates was cut off by the mask. The activity concentration templates were left unmasked.

| Constructed atlases
Example slices of the constructed BP ND templates for GABA A and mGluR 5 are presented in Figure 3. Extensive data from other modalities for these two tracers were not readily available for a comprehensive comparison of the resulting data, so no validation was possible in this regard.

| Assessment of the warp effect
The results for a selection of large, medium, and small ROIs for both sets of subjects are given in Figure 4. The assessment of the warp effect in general revealed only a very slight difference between the native and the warped space values for both subject groups. No clear systematic difference between the native space and warped space values was observed in either group. For the [ 11 C]FMZ group, the maximal δ across all subjects in the selected ROIs was found in the GM at 7.8%. The highest average δ was observed in the right thalamus at (3.8 ± 1.4)%. The maximal δ in the [ 11 C]ABP group was found in the right thalamus at 7.1%, and the left frontal pole exhibited the highest average δ of (3.0 ± 1.0)%. Despite the low differences, a two-sided Wilcoxon signed-rank test revealed that 9 out of 14 tested reasons were significantly different at a significance level of p < .05.

| PET-to-PET registration tests
Example slices of the constructed normalized activity concentration templates of [ 11 C]ABP and [ 11 C]FMZ are illustrated in Figure 5. In order to demonstrate the more realistic tracer distribution of the The results of the registration tests for both groups of subjects are given in Figure 6. Three different cases were evaluated: (a) using Testing whether the difference in MNI spaces has a significant impact on the registration quality revealed that all cases which were found to be statistical significant in the previous case remained statistically significant in this test. Additionally, the difference in gSSim of the SPM ON registration was found to be statistically significant in this case although it had been not statistically significant before.
Smoothing was shown to have almost no effect on the significance of any difference.

| DISCUSSION
In this work, we aimed to construct parametric maps containing infor-  Gispert et al., 2003;Meyer et al., 1999).

| Considerations regarding pipeline construction
Our pipeline was assembled using the Python-based neuroimaging package NiPype (Gorgolewski et al., 2011), which, due to its modular nature, offers the possibility of exchanging individual steps to adapt to specific needs. Furthermore, the proposed pipeline can be applied F I G U R E 6 Evaluation of the registration quality using different templates and two registration methods: antsRegistrationSyNQuick (ANTs SQ ) and SPM Old Normalize (SPM ON ).   to ascertain whether it has any impact on the registration quality (dotted bars). Significance of the differences between the relevant groups is indicated by asterisks (*p<.05, **p<.01, ***p<.001, ****p<.0001) or not significant (no label), according to a paired two sample t test multiple comparisons correction was not applied with minimal user intervention by simply providing the reconstructed bolus-infusion PET data, raw structural MR scans, a target template file in NifTi format, and the material for V T and/or BP ND estimation as inputs. However, the metabolite corrected venous blood plasma activity at the time of interest must be given as an input to the pipeline in the form of a text file for estimation of V T , and a binary NifTi file masking the reference region is necessary as an additional input to calculate BP ND . Thus, the pipeline is not immediately applicable to all types of PET data in its current form.
A crucial step during the pipeline is the applied spatial normalization routine. There are two general approaches to spatial normalization of functional images which have been discussed previously (Ashburner & Friston, 1999): (a) direct calculation of the transformation parameters from the subject-space PET image to a standardized PET template; (b) the use of a structural scan from another imaging modality, that is in co-registration with the PET image, to estimate the transformation and then to apply this transformation to the PET image in a second step.
Higher registration quality using the latter approach was first proposed by Ashburner and Friston (1999). Subsequently, this approach was investigated, and the findings of Gispert et al. (2003) and Martino et al. (2013) indicate a significantly improved registration quality when using an MR image to estimate the spatial transformation. This significant difference is due to the fact MR offers a superior spatial and structural resolution compared to PET (Ashburner & Friston, 1999), whereas PET is advantageous for the visualization of metabolic processes. Thus, most previous studies have employed varieties of an MR-based approach to register PET images to a template space (Della Rosa et al., 2014;Gispert et al., 2003;Vállez Garcia et al., 2015). Due to the above advantages, we elected to apply the MR-based method to estimate the transformation.
However, the dynamic [ 11 C]Raclopride template, described in Section 1 of this article, was constructed using a purely PET-based approach (Bieth et al., 2013).
In the following, some methods applied in the proposed approach are briefly discussed. During this work, two bias correction methods were tested: the bias correction method included in the current segmentation routine of SPM12 and the N4 bias correction (Tustison et al., 2010) method that is included in ANTs. As realistic phantom data are difficult to obtain, the methods were assessed by visual inspection. Here, the N4 bias correction seemed to remove the bias field best when using the manually edited brain mask, restricting the algorithm to include only brain voxels in the estimation. The image intensity ranges were standardized to each other using an option in the ANTs N4 bias correction routine. To be able to perform BP ND estimation, as well as being able to warp the parametric images, the corrected MR images were then warped into MNI 2009c nonlinear space. The ANTs registration routine was shown to outperform a variety of other methods in an evaluation of 14 nonlinear registration methods (Klein et al., 2009). In addition, ANTs offers several parameters that can be adjusted to optimize the output. Thus, ANTS registration was applied as the method of choice in this work.
An important quality control step is the examination of the resulting motion correction parameters. Observation of these parameters is necessary due to the attenuation correction method applied (see Section 2.1). Large movements cause errors in the reconstruction, and scans that contained motion larger than the image resolution were excluded from consideration. This was achieved by applying motion correction with respect to the first frame. Only scans with motion parameters less than 3 mm in x, y, z-direction or 3.5 of roll, pitch, or yaw were considered for further processing.

| Applicability of the tracer-specific maps and templates
To be able to use the constructed in vivo maps of neuroreceptor distribution, it was necessary to assess the impact of the warp on the para-  (Smart et al., 2018) when BP ND was estimated using the simplified reference tissue model (Lammertsma & Hume, 1996).
Therefore, the highest average deviation of 3.0 ± 1.0% is smaller than the expected general variability of [ 11 C]ABP. Although the percent errors were lower than the expected variabilities of the tracers, a Wilcoxon signed-rank test revealed that the differences were significant in 8 of the 14 tested cases. This suggests that the transformation, or the applied way of extracting the values, introduced a systematic bias towards lower values in warped space, which should be carefully considered before use.
We also applied the pipeline to generate MNI space activity concentration images that were normalized by the mean activity in In both cases, only the level of statistical significance changed in some cases, but all tests remained statistically significant.

| Limitations
When interpreting our study, some limitations should be kept in mind.
First, although fully automatic processing is an implemented option in the pipeline, it is questionable whether current skull stripping methods can provide the necessary accuracy in a robust way. Therefore, we chose to manually edit brain masks in a prior step and give them as input into the pipeline in addition to the raw PET and MR images. A robust skull-stripping routine that offered results visually imperceptible from manual extractions would improve the usability of the proposed pipeline considerably.
Second, the quality of the intermediate steps that is, motion correction, co-registration, PVC, and nonlinear registration, ultimately limits the quality of the resulting images but is often impossible to quantify. Thus, these steps were visually inspected for inconsistencies, which is an error-prone inspection method.  (Nørgaard et al., 2021) in which gene expression and histology were considered. It may thus be possible to further validate our results using the aforementioned tools in an ensuing study.
Finally, in addition to the very low sample size (four and five subjects respectively), the approach applied to evaluate the registration quality is intrinsically limited. The tests were based on the assumption that MR-based registrations of PET images are more accurate than intermodal registrations of PET images, as previously hypothesized in Ashburner and Friston (1999) and reviewed in Gispert et al. (2003) and Martino et al. (2013). However, MR-based registrations cannot be expected to be without errors, meaning that the "ground truth" images to which the results were compared do not actually reflect a ground truth. This test could be improved by using simulated data to quantify the registration accuracy in a more robust way.

| CONCLUSION
In this work, we demonstrate the reliable applicability of a PET/MR NiPype image processing pipeline specifically designed to give accurate, parametric in vivo PET template construction. The pipeline provides a neuroimage pre-processing method that is easy to implement for arbitrarily large bolus-infusion PET/MR datasets using state-ofthe-art image processing tools at minimal user intervention. The pipeline outputs were further used to construct MNI space distributions of V T , BP ND , and normalized activity concentration of [ 11 C]FMZ and [ 11 C]ABP in healthy male humans. The validations performed lead to the conclusion that the pipeline is suitable for application in a variety of advanced analyses of multimodal datasets. However, the constructed maps should only be used after careful consideration of the warp effect. Based on the extensive testing of the applicability of the normalized activity concentration templates for PET to PET registration, we conclude that using the constructed templates instead of an arbitrary perfusion template should benefit registration performance.

ACKNOWLEDGMENTS
Parts of this work were presented at the 28th Annual Meeting of the Open access funding enabled and organized by Projekt DEAL.

CONFLICT OF INTERESTS
The authors declare that there are no conflict of interests. Irene Neuner: Study design and setup, approval ethics and BfS, funding, and revision of the paper.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study as well as the resulting maps and templates are openly available in Jülich DATA at https:// doi.org/10.26165/JUELICH-DATA/HDVEEF.

ETHICS STATEMENT
Both studies of which data were analyzed in this work, as well as the corresponding analyses, were approved by the Ethics Committee of the Medical Faculty at the RWTH Aachen University and the German Federal Office for Radiation Protection (Bundesamt für Strahlenschutz). Written informed consent was obtained from all participants before the measurement.