Participants and data acquisition
The participants and data acquisition have previously been described . This study was approved by the Regional Board of Medical Ethics in Uppsala as well as the Radiation Ethics Committee of Uppsala University Hospital, and all participants consented in writing to take part in the study before inclusion. The study comprised nine patients with Parkinsonism (5 females, 4 males; median age 72 years, range 49–82). Previously, each participant had undergone a 10 min transmission scan using three rotating 68Ge rod sources on an ECAT Exact HR+ scanner (Siemens, Knoxville), prior to injection of any radioactivity. Participants then underwent a dynamic brain PET scan using 11C-PE2I on a 3T time-of-flight (TOF) PET/MR (SIGNA PET/MR, GE Healthcare, Waukesha, USA) system.
An 80-min PET scan was acquired in list mode, starting simultaneously with intravenous bolus administration of 5 MBq/kg [11C]PE2I using an infusion pump, and divided into 22 time frames of increasing duration (4 × 60s, 2 × 120s, 4 × 180s, 12 × 300s). Relevant MR sequences were acquired during the first five minutes of the [11C]PE2I PET scan to avoid misalignment due to head movements. Three MR sequences were acquired during the [11C]PE2I PET scan: 1) a T1w 3D LAVA Flex that was later used for Atlas MRAC (duration 18s, 1 NEX, FOV 500 mm, slice thickness 5.2 mm, overlap 2.6 mm, matrix 256 × 256, and 5° flip angle); 2) a proton-density ZTE sequence (duration 153s, 4 NEX, FOV 260 mm, slice thickness 1.4 mm, no slice gap, matrix 192
Generation of MR attenuation correction maps
For each participant, the following MR attenuation correction maps (MRAC) were produced:
Atlas-MRAC – The single-atlas based method [11,20] consisted of three steps: (1) application of a Hessian matrix to enhance bone structures in the T1w image, (2) pseudo-CT generation by registration of the enhanced image to a head atlas based on CT scans of 50 participants, (3) standard energy conversion and resampling of the pseudo-CT, resulting in an AC map with dimensions of 128 × 128 × 89 voxels (4.68 × 4.68 × 2.78 mm). This method was vendor-provided (software version MP26) as a standard application for our PET/MR system when the study was initiated.
ZTE-MRAC – In this procedure, an intensity equalization to a ZTE image was followed by logarithmic image rescaling to enhance bone tissue. Next, a mask was used to isolate the brain data, thus removing bed and coil information. A sequence of thresholding operations was then applied to the subject’s brain image to segment bone and air regions by fitting a Gaussian to the main image histogram peaks. Remaining internal air compartments were further segmented by means of histogram thresholding. ZTE bone voxels are assigned continuous attenuation coefficients using a ZTE-intensity HU calibration curve as described in . Finally, the resulting MRAC image was co-registered and resampled to the subjects’ individual Atlas-MRAC by applying 6-parameter rigid-body registration creating the ZTE-MRAC map. This method is also a vendor-implemented process (GE PET Reconstruction Toolbox MP26), but during data acquisition still under development [12,26,27].
MaxProb-MRAC – This method utilized a database of 40 corresponding MR and CT images [43,51]. Processing steps were executed using an in-house MATLAB pipeline (MATLAB R2017a, Mathworks Inc., Natick, MA). Each MR image from the database was paired with the participant’s high-resolution T1w image and masking was applied to both images to reduce extraneous image information. In the following registration steps, the masked MR-image pair was aligned by means of affine registration followed by non-rigid registration using normalized mutual information as the similarity measure (register github.com/BioMedIA/MIRTK). Subsequently, the transformation matrices acquired from the non-rigid registration of MR-images were applied to the corresponding CT images, yielding a database of co-registered MR/CT image pairs. In the next step, the co-registered CTs were segmented into three tissue classes defined by intensity thresholds (air: < -500 HU, soft tissue: -500-300 HU, bone: > 300 HU . Each voxel tissue class of the target subject space was then assigned to a tissue class by majority voting of tissue class labels across the registered CT database. Finally, a voxel-wise pseudo-CT was generated by averaging CT Hounsfield values of atlases belonging to the majority class for the corresponding voxel. A bilinear energy conversion . The resulting MRAC image was rigidly co-registered and resampled to the participant’s individual Atlas-MRAC. Thereafter, the MaxProb-MRAC image was completed by adding, from the Atlas-MRAC map, the neck information missing due to differences in axial FOV between the images retrieved from the MR/CT database (216 mm) and the PET/MR scanner (250 mm).
68– Attenuation maps were reconstructed from the 68Ge transmission scan using ordered subset expectation maximisation (OSEM) with 6 iterations, 8 subsets and a 4 mm Hanning post filter, dimensions 128 × 128 × 63 voxels (5.15 × 5.15 × 2.43 mm). Thereafter, a mask was applied to strip the AC map from bed and head support as well as noise surrounding the head, followed by rigid co-registration and resampling to the participant’s individual Atlas-MRAC map. As the axial FOV differed between the stand-alone PET (155 mm) and PET/MR scanner (250 mm), the 68Ge-AC map was completed with the corresponding information concerning the neck and top of the head from the Atlas-MRAC map.
As a common step for all generated MRAC maps, coil and bed AC map templates were then incorporated, applying an in-house MATLAB pipeline (MATLAB R2017a, Mathworks Inc., Natick, MA).
The dynamic [11C]PE2I PET data were reconstructed using time-of-flight OSEM with 2 iterations, 28 subsets, 5 mm Gaussian post-filter, 128 × 128 reconstruction matrix and 300 mm FOV. MRAC was performed in four ways as described in the previous section. Further, all appropriate corrections for quantitative image reconstruction were applied.
The methodology and validation for voxel-level analysis of dynamic [11C]PE2I scans have been previously reported [48,49].
Following this approach, the reconstructed [11C]PE2I PET images were realigned to correct for interframe patient movements using an early (0-3 min) [11C]PE2I summed image as reference using an in-house MATLAB 2018 script. Interframe motion was estimated in this way for the PET images corrected for AC using 68Ge-AC and the same transformation were applied to the other three dynamic data sets. Subsequently, the same summed image was used for co-registration of the 3D T1w MRI scan based on a 6-parameter rigid transformation, to achieve positional alignment. Then MR images were segmented into grey matter, white matter, and cerebrospinal fluid using Statistical Parametric Mapping (SPM12; Wellcome Trust Center for Neuroimaging, University College London, UK). Grey matter volumes of interest (VOI) were established on the T1w structural MR images using an automated probabilistic template as implemented in the PVElab software  for cortical and limbic regions. For striatal regions, MAPER , a technique optimized for segmentation of atrophic brains, was used to achieve a more accurate delineation. All VOIs were projected over the dynamic scans to generate time-activity curves (TACs).
Parametric images were generated from the [11C]PE2I scan using receptor parametric mapping (RPM) with cerebellar grey matter as a reference region . RPM is an implementation of the simplified reference tissue model (SRTM; ) with a set of predefined basis functions to linearize the model and estimate voxel-wise R1 and BPND that can be applied to [11C]PE2I. The parametric [11C]PE2I BPND images demonstrate specific binding of [11C]PE2I to DAT directly proportional to DAT density (availability) and therefore mainly show the deep grey matter of the striatum. The parametric [11C]PE2I R1 images display relative cerebral blood flow, which reveals overall brain functional activity. This procedure was repeated for all four [11C]PE2I PET data sets. Finally, the grey matter VOIs on the co-registered MR-images were projected on the various parametric [11C]PE2I R1 and BPND images from all four datasets.
Volumes of interest
Quantification of BPND was evaluated separately for the caudate nucleus and putamen as well as for the whole dorsal striatum (volume-weighted average of putamen and caudate). These regions normally have high DAT availability, but a pronounced reduction is often observed in persons with parkinsonism .
Evaluation of the quantification of R1 was based on four clusters of VOIs across the brain: anterior cortical regions (cingulate, frontal gyrus), posterior cortical regions (occipital cortex, parietal cortex, and somatosensory-motor cortex), dorsal striatal regions (caudate nucleus, putamen) and limbic regions (amygdala, hippocampus, hypothalamus, and thalamus) . In addition, whole-brain grey matter (WB) was determined as an overall measure.
Evaluation of MRAC maps
All participants’ MRAC maps were converted to MNI space using SPM12. Mean MRAC images for all four methods were calculated and bias compared to 68Ge-AC was calculated at the voxel level as follows: see formula 1 in the supplementary files.
where i refers to the three MRAC methods being compared to 68Ge-AC.
Attenuation maps were assessed on soft tissue and bone compartments. A large VOI was drawn over the 68Ge-AC map soft tissue and copied to the other AC methods. A bone mask was created by simple segmentation of the ZTE map followed by 2-pixel erosion to reduce the probability of coinciding with other tissues. The bone mask was then transferred to all other AC methods. Mean bias values were calculated for both soft tissue and bone.
Evaluation of [11C]PE2I BPND and R1 estimates
[11C]PE2I BPND and R1 images for each subject were converted to MNI space using SPM12. Mean [11C]PE2I BPND and R1 images for all four methods were calculated and bias (eq.1) compared to 68Ge was calculated at the voxel level for the defined VOIs.
For quantitative assessment, we used relative differences (% bias; see eq. 2) in BPND and R1 as a measure of accuracy, while standard deviation of the bias was taken as a measure of precision.
See formula 2 in the supplementary files.
with i referring to the three evaluated MRAC methods.
Correlation analysis (Spearman) and Deming regression were used to assess the degree of agreement between BPND and R1 values obtained using each of the evaluated MRAC methods and the reference method (68Ge-AC).
Statistical analysis was performed using GraphPad Prism 6 (GraphPad Software, La Jolla, California). Significant differences in bias (p<0.05) between each evaluated MRAC method and 68Ge-AC as reference method were assessed using a Friedman test with post-hoc tests (comparison to 68Ge-AC) and Dunn’s correction.
Evaluation of time-activity curves
Previous reports indicate that effects of attenuation correction on uptake over time can be inhomogeneous [29,43,44,57] and may potentially affect the accuracy and precision of the parameters estimated by kinetic modelling. We therefore examined relative bias and its SD in [11C]PE2I standardized uptake values (SUV) over time. This was done for striatal regions with a high DAT density and the reference region (cerebellum) used after implementation of Atlas-, ZTE- and MaxProb-MRAC.