2.3. 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 (Gorgolewski et al., 2011)—a Python package comprising a variety of neuroimaging software wrapped into Python interfaces so that they can be connected in a single programming language. The use of NiPype enables multi‐step processing of entire sets of neuroimaging data. In addition, the processing approaches can be simply adapted to individual needs. The relevant NiPype dependencies for this work are Statistical Parametric Mapping (SPM12; Penny et al., 2011), ANTs (Avants et al., 2011), FMRIBs Software Library (FSL; Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012), and PETPVC (Thomas et al., 2016). It should be noted here that a set of NiPype image processing workflows, named Pypes (Savio, Schutte, Graña, & Yakushev, 2017), is already publicly available. This collection of workflows also comprises a workflow for the MR‐based registration of PET images into a standardized space. The available workflows in Pypes, however, do not use implementations for PET parameter estimation, such as binding potential, or volume of distribution, to draw conclusions based on quantitative data. The performance of parameter estimation in transformed spaces is not recommended as it could potentially bias the results. Thus, one cannot use the outputs from the Pypes workflows as a basis for parameter estimation. In the approach proposed here, emphasis was placed on an optimized selection of pre‐processing 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.

A graphical representation of the constructed NiPype preprocessing pipeline. ANTs, Advanced Normalization Tools; FSL, FMRIB's Software Library; SPM, Statistical Parametric Mapping

To be able to apply the pipeline, the user is required to install the software requirements mentioned above and must simply provide the following.

The raw structural MR scans in addition to the respective binary brain masks.

The reconstructed and TAC evaluated PET data.

A target template file (i.e., in MNI Space).

A list of tuples, defining the frames of interest for each subject (starting frame no., no. of frames).

(Optional) For V _T estimation: metabolite corrected mean venous blood plasma activity at the time of interest as a text file.

(Optional) For BP _ND estimation: a binary NifTi file masking the reference region in template space.

Currently, the pipeline expects all neuroimaging data in NifTi format. 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.

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 co‐registered 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.

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) step to match each voxel to the corresponding voxel in the template image. The ICBM152 2009c nonlinear asymmetric, 1 × 1 × 1 mm resolution brain template, available on the website of the McConnel Brain Imaging Center, Montreal Neurological Institute of the McGill University Montreal (Fonov, Evans, McKinstry, Almli, & Collins, 2009) was chosen as the registration target for this work. Unless explicitly stated, MNI space refers to the above‐mentioned MNI template space in the following work. The resulting transformation into MNI space (forward transform) and from MNI space back into subject space (inverse transform) were output from this pipeline segment for later purposes.

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 [¹¹C]ABP (Burger et al., 2010) and [¹¹C]FMZ (Mauler et al., 2020). 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 concentration C _ND

The estimations of BP _ND and V _T were performed in subject space before the transformation into MNI space was applied. For BP _ND calculation, the MNI space reference region was warped into subject space using the inverse transform. For the estimation of BP _ND in the [¹¹C]ABP subjects, the warped mask was multiplied with the individual SPM12 GM segment to exclude voxels unlikely to contain GM. Due to the interpolation steps during the transformation, voxel values at the border of the mask were changed so that the resulting mask was no longer binary. To calculate the mean value (BP _ND, V _T, or normalized activity concentration) fslstats was applied. By default, fslstats thresholds and binarizes the input mask at 0.5 and then calculates the mean value from all non‐zero voxels. This threshold is helpful 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).