No impact of attenuation and scatter correction on the interpretation of dopamine transporter SPECT in patients with clinically uncertain parkinsonian syndrome

The picture archiving and communication system of our institution was searched for DAT-SPECT with ¹²³I-FP-CIT that had been performed due to a clinically uncertain suspicion of nigrostriatal degeneration. The only inclusion criteria were that SPECT had been performed with low-energy-high-resolution parallel-hole or fan-beam collimators and that the raw projection data were available for consistent retrospective image reconstruction. This identified 1765 DAT-SPECT. The corresponding DAT-SPECT images were inspected visually to exclude cases with atypical reduction of striatal ¹²³I-FP-CIT uptake most likely caused by structural/vascular lesions. This led to the exclusion of 25 of the 1765 DAT-SPECT. The remaining 1740 DAT-SPECT from 1712 different patients (43.3% females) were included. The mean age at the time of DAT-SPECT was 66.7 ± 11.6 years (range 20.2–90.8 years). There were no further eligibility criteria to make sure that the included data were representative of everyday clinical routine.

DAT-SPECT had been performed between December 2008 and January 2020 according to common procedure guidelines [21‐23] with four different hardware settings: Siemens e.cam (Siemens Healthineers, Erlangen, Germany) dual-head camera equipped with low-energy-high-resolution collimators (n = 704), Siemens Symbia TruePoint dual-head camera with low-energy-high-resolution (n = 147) or fan-beam collimators (n = 457), and Mediso AnyScan Trio (Mediso, Budapest, Hungary) triple-head camera equipped with low-energy-high-resolution-high-sensitivity collimators in dual-head mode (n = 432). Detailed acquisition parameters are given in Supplementary Table 1.

In order to reduce variability of no interest associated with the different hardware settings and/or the use of different reconstruction algorithms over time, all raw projection data were reconstructed retrospectively using iterative ordered-subsets-expectation–maximization [24] with attenuation and simulation-based scatter correction as well as collimator-detector response modeling as implemented in the HybridRecon-Neurology tool of the Hermes SMART workstation v1.6 (Hermes Medical Solutions, Stockholm, Sweden) [25‐28]. All parameter settings were as recommended by Hermes [25] for the EANM/EANM Research Ltd (EARL) ENC-DAT project (European Normal Control Database of DaTSCAN) [10, 16, 29‐31]. More precisely, ordered-subsets-expectation–maximization was performed with 5 iterations and 15/16 subsets for 120/128 views. For noise suppression, reconstructed images were postfiltered by convolution with a 3-dimensional Gaussian kernel of 7 mm full-width-at-half-maximum. The simulator of the SPECT-acquisition process has been described in [26, 27]. In brief, the simulator is composed of two components, a Monte Carlo simulator to track the photons in the patient and a collimator-detector response model. In the current study, the collimator-detector response was modeled by a Gaussian point-spread-function model (rather than using a separate Monte Carlo simulator) [26, 27]. The Gaussian model is based on collimator specifications and on the intrinsic spatial resolution of the detector (both to be provided by the user; Supplementary Table 1). The Monte Carlo simulator modeled scatter in the patient using the convolution-based forced detection algorithm simulating 10⁵ photons with 2 scatter update iterations [26, 27]. Attenuation was modeled using a uniform attenuation map with a narrow-beam attenuation coefficient of 0.146/cm and assuming exponential loss of radiation intensity according to Chang [25, 32]. The outer contour of the head for modeling of attenuation (and scatter) was delineated using the automatic thresholding-based approach implemented in the Hermes SMART workstation. This algorithm smoothes the non-corrected image by convolution with a 3-dimensonal isotropic Gaussian kernel of 25 mm full-width-at-half-maximum and then fits an ellipse to the 30% isocontour, separately in each transaxial plane [25]. The automatic delineation avoids variability of no interest associated with manual delineation. After all photons have been simulated, the subprojection maps are convolved with the collimator–detector response function from the Gaussian point-spread-function model.

Each SPECT scan was reconstructed also without ASC but otherwise exactly the same parameter settings as with ASC. In particular, collimator-detector response modeling was turned on in both cases. The batch utility of the HybridRecon-Neurology tool was used to reconstruct all 3480 (= 2*1740) images in a single batch in order to avoid errors by incorrect manual operation.

The rationale for using the same iterative ordered-subsets-expectation–maximization algorithm for image reconstruction with and without ASC was to avoid “contamination” of ASC effects with algorithm effects associated with the use of different reconstruction algorithms with and without ASC (e.g., filtered backprojection without ASC versus iterative ordered-subsets-expectation–maximization with ASC).

Individual DAT-SPECT images were stereotactically normalized to the anatomical space of the Montreal Neurological Institute (MNI) using the normalize tool of the Statistical Parametric Mapping software package (version SPM12) with the following parameter settings: affine transformation (no nonlinear warping), template weighting by a binary mask of the whole head including the scalp and excluding the cerebellum, source weighting by a uniform cube of 200 mm edge length centered at the center of mass of the individual DAT-SPECT image in patient space, MNI template regularization, preserve concentrations, voxel size 2 × 2 × 2 mm³, trilinear interpolation, and no wrapping. Multiple custom DAT-SPECT templates representative of normal and different levels of Parkinson-typical reduction of striatal ¹²³I-FP-CIT uptake were used as target for stereotactical normalization. SPM then finds the linear combination of these templates that best matches the patient’s image.

For each of the 1740 DAT-SPECT, stereotactical normalization was performed first for the image with ASC. The corresponding image without ASC was coregistered to the non-normalized ASC image and then stereotactically normalized using the same transformation as for the ASC image. The rationale for this was to avoid bias that might have been caused by independent stereotactical normalization of DAT-SPECT images with and without ASC.

Intensity scaling was achieved by dividing the voxel intensities of the stereotatically normalized images by the individual 75th percentile of the voxel intensity in a reference region comprising the whole brain without striata, thalamus, brainstem, cerebellum, and ventricles [33]. A binary mask of the reference region predefined in MNI space was used for this purpose. The scaled images are semi-quantitative images representing the distribution volume ratio (DVR).

A standardized display (Fig. 1) was used for the visual interpretation of the DAT-SPECT images, similar to the display used in everyday clinical routine at our site. The display comprised ten transversal DVR image slices of 4 mm thickness from the superior to the inferior edge of the striatum with the maximum of the color table individually scaled to the maximum intensity in the ten images. In addition, the display presented a transversal DVR image slab of 12 mm thickness through the center of the striatum with the maximum of the color table scaled to a fixed upper DVR threshold [34]. The DVR threshold had been optimized previously, separately for images with ASC (DVR = 4.05) and for images without ASC (DVR = 3.45), as inconsistent color mapping might lead to misinterpretation [19, 35].

Visual interpretation of the SPECT images was performed independently by three readers with different experience in the clinical reading of DAT-SPECT. Two readers had more than 10 years of experience (≥ 2000 cases); the third reader had no previous experience. The readers were blinded for all clinical data. They were asked for binary categorization of the SPECT images with respect to the presence of Parkinson-typical reduction (“reduced”) or the absence of Parkinson-typical reduction of striatal ¹²³I-FP-CIT uptake (“normal”). Two training sessions, each with 48 randomly selected images with ASC and 48 randomly selected images without ASC, were performed independently by the three readers. The results of the training sessions were discussed among the 3 readers in order to harmonize visual interpretation between the readers.

For the main reading, four 1740-page PDF documents were created. Each PDF document presented the 1740 DAT-SPECT cases, one per page. Thus, a total of 6960 images were categorized by each reader. Two of the four PDF documents contained the same 1740 images with ASC, and the other two PDF documents contained the same 1740 images without ASC. The ordering of the 1740 images within a given PDF document was randomized using a random permutation of the integers from 1 to 1740 generated by the MATLAB routine “randperm.” This was performed separately for each of the four PDF documents (with different states of the random number generator to guarantee different randomization for each PDF document). The readers categorized the images by clicking “normal” or “reduced” in the PDF document (Fig. 1).

DAT-SPECT images with discrepant categorization between the two reading sessions were read a third time by the same reader to obtain an intra-readers consensus, separately for each reader.

The ¹²³I-FP-CIT SBR in left and right putamen was obtained by hottest voxels analysis of the stereotactically normalized DVR image using large unilateral putamen masks predefined in MNI space [36, 37]. The putamen masks were much bigger than the actual putamen volume to guarantee that all putaminal counts were included. The number of hottest voxels within a unilateral putamen mask to be averaged was fixed to a total volume of 10 ml. The putamen SBR was computed as the mean DVR in the hottest voxels minus 1. The minimum putamen SBR of both hemispheres was used for the analysis.

Cross tables, Cohen’s kappa and the percentage of discrepant cases were used to characterize.

The impact of ASC on the discriminative power of the putamen SBR was tested as follows. The distribution of the putamen SBR was characterized by a histogram with 0.1 bin width. The resulting histogram was fitted by the sum of two Gaussians:where A₁, A₂ are the amplitudes, M₁, M₂ are the mean values, and SD₁, SD₂ are the standard deviations of the Gaussian functions. The MATLAB routine “fminsearch” with default parameter settings was used for this purpose.

The power of the SBR to differentiate between normal and reduced DAT-SPECT was estimated by the effect size d of the distance between the two Gaussians computed as the differences between the mean values scaled to the pooled standard deviation

The cutoff c for differentiation between normal and reduced SBR was selected halfway between M₁ and M₂ in units of standard deviations, that is

The putamen SBR was considered “reduced” if it was smaller than the cutoff c; it was considered “normal” if it was equal or larger than the cutoff c.

The histogram analysis was performed separately with and without ASC.

In order to assess a possible impact of the hardware setting on the discrimination between normal and reduced cases, the proportion of cases with discrepant categorization without versus with ASC was compared between the 4 different hardware settings, separately for visual interpretation and for the automatic categorization based on the semi-quantitative analyses. The chi-square test of the corresponding 4 × 2 cross tables (4 hardware settings × without versus with ASC) was used to test the observed differences for statistical significance.

Mean intra-readers kappa of the visual categorization of the DAT-SPECT images was 0.957 ± 0.012 (range 0.949–0.971) without ASC versus 0.956 ± 0.010 (0.949–0.967) with ASC (Table 1). The mean proportion of cases with discrepant categorization by the same reader was 2.13% ± 0.60% (range 1.44–2.53%) without ASC versus 2.20% ± 0.47% (1.67–2.53%) with ASC (Table 1).

The mean between-reader kappa of the visual categorization of the DAT-SPECT images was 0.946 ± 0.022 (range 0.924–0.967) without ASC versus 0.936 ± 0.014 (0.921–0.947) with ASC (Table 1). The mean proportion of cases with discrepant categorization between two readers was 2.72% ± 1.06% (range 1.67–3.79%) without ASC versus 3.18% ± 0.70% (2.64–3.97%) with ASC (Table 1).

The mean without-versus-with-ASC kappa of the visual categorization of the DAT-SPECT images by the same reader was 0.967 ± 0.010 (range 0.961–0.978, Table 1). The mean proportion of cases with discrepant categorization of the same DAT-SPECT without and with ASC by the same reader was 1.66% ± 0.50% (range 1.09–1.95%). The without-versus-with-ASC kappa of the majority visual categorization of the DAT-SPECT images was 0.971 (standard error 0.006). A disagreement of the majority visual categorization of the same DAT-SPECT without and with ASC occurred in 25 of the 1740 DAT-SPECT (1.44%), with rather balanced disagreement in both directions (15/10 cases, Table 2).

Among the 1740 DAT-SPECT, there was only one single case (0.06%) with discrepant categorization between without and with ASC by all 3 readers (Supplementary Fig. 1).

The histograms of the putamen SBR without and with ASC and their fit by the sum of two Gaussians are shown in Fig. 2. The parameters obtained by the fit are given in Table 3. The effect size d of the distance between the two Gaussians was slightly larger with ASC than without ASC (3.958 versus 3.760). Cross tables of the automatic binary classification by the putamen SBR with versus without ASC using the cutoffs derived from the histogram analyses are given in Table 2. The without-versus-with-ASC kappa for the binary classification was 0.964 (standard error 0.006).

Cohen’s kappa between the automatic binary categorization based on the putamen SBR versus the visual binary categorization (majority vote) was 0.911 (standard error 0.010) without ASC and 0.918 (0.010) with ASC. The proportion of discrepant cases between the SBR-based categorization and visual categorization was 4.43% without ASC and 4.08% with ASC.

The proportion of cases with discrepant visual categorization (majority vote) between without versus with ASC was 1.70%, 2.04%, 0.67%, and 1.62% for the DAT-SPECT acquired with the Siemens e.cam with LEHR, the Siemens Symbia TruePoint with LEHR, the Siemens Symbia TruePoint with fan-beam, and the Mediso AnyScan Trio with LEHRHS, respectively. The proportion of cases with discrepant automatic categorization between without versus with ASC based on the putamen SBR using the cutoffs derived from the histogram analyses was 1.70%, 1.36%, 1.75%, and 2.08% for the DAT-SPECT acquired with the Siemens e.cam with LEHR, the Siemens Symbia TruePoint with LEHR, the Siemens Symbia TruePoint with fan-beam, and the Mediso AnyScan Trio with LEHRHS, respectively. The differences were not statistically significant, neither for the visual interpretation (p = 0.423) nor for the automatic categorization (p = 0.940).