Stereotactical normalization with multiple templates representative of normal and Parkinson-typical reduction of striatal uptake improves the discriminative power of automatic semi-quantitative analysis in dopamine transporter SPECT

Patients

The PACS of the Department of Nuclear Medicine of the University Medical Center Hamburg Eppendorf was searched for DAT-SPECT that had been performed to support the diagnosis of a clinically uncertain parkinsonian syndrome. The only inclusion criteria were that DAT-SPECT had been performed with low-energy-high-resolution parallel-hole or fan-beam collimators and that the raw projection data were digitally available for consistent retrospective image reconstruction. There were no further eligibility criteria to make sure that the included data were representative of everyday clinical routine at our site. This resulted in 1740 DAT-SPECT.

About two thirds of the DAT-SPECT had been included in previous studies on deep learning-based classification of DAT-SPECT [17] and data-driven identification of diagnostically useful extrastriatal signal in DAT-SPECT [18].

SPECT imaging

SPECT had been performed between December 2008 and January 2020 according to common procedures guidelines [19, 20] with four different cameras: Siemens e.cam dual head camera equipped with low-energy-high-resolution collimators, Siemens Symbia TruePoint dual head camera with low-energy-high-resolution collimators, Siemens Symbia TruePoint with fan-beam collimators, and Mediso AnyScan Trio triple head camera equipped with low-energy-high-resolution-high-sensitivity collimators in dual head mode. Detailed acquisition parameters are given in Table 1.

Table 1 SPECT acquisition parameters

All SPECT images were reconstructed retrospectively using the iterative ordered-subsets-expectation–maximization algorithm with resolution recovery implemented in the HybridRecon-Neurology tool of the Hermes SMART workstation v1.6 with parameter settings recommended for FP-CIT SPECT by Hermes (5 iterations, 15/16 subsets for 120/128 views, postfiltering with 3-dimensional Gaussian kernel of 7 mm full-width-at-half-maximum, uniform attenuation correction with narrow-beam attenuation coefficient 0.146/cm, simulation-based scatter correction, resolution recovery with a Gaussian model).

123I-FP-CIT templates

An 123I-FP-CIT template representative of normal striatal 123I-FP-CIT uptake was generated as follows: twelve DAT-SPECT images with normal striatal signal according to visual inspection were selected randomly from the 1740 DAT-SPECT included in this study. More precisely, the 1740 DAT-SPECT images were sorted in alphabetical order (according to the file names) and then the first 12 images with normal striatal signal according to visual inspection were used. Each individual image was stereotactically normalized into the anatomical space of the Montreal Neurological Institute (MNI) using the Normalize tool of the statistical parametric mapping software package (version SPM12), a custom 123I-FP-CIT template in MNI space generated previously [16] as target, and 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 predefined in MNI space, source weighting by a binary 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, no wrapping. Intensity normalization of the stereotactically normalized images was achieved by voxelwise scaling to the individual 75th percentile of the voxel intensity in a reference region comprising the whole brain without striata, thalamus, brainstem, cerebellum, and ventricles [15]. The twelve resulting stereotactically normalized distribution volume ratio (DVR) images were averaged (soft mean). A preliminary, left–right-symmetric template was obtained by flipping the mean image at the midsagittal plane and averaging the mean and the flipped image. The final left–right-symmetric template representative of normal striatal 123I-FP-CIT uptake was obtained by repeating these steps two times using the preliminary template from the previous iteration as target image for stereotactical normalization.

123I-FP-CIT templates representative of moderate and of strong Parkinson-typical reduction of striatal uptake were generated from twelve randomly selected DAT-SPECT with moderate reduction and twelve randomly selected DAT-SPECT with strong reduction of the striatal signal. More precisely, the 1740 DAT-SPECT images were sorted in alphabetical order (according to the file names) and then the first 12 images with moderate reduction and the first 12 images with strong reduction of the striatal signal according to visual inspection were selected for the generation of the corresponding templates. The further procedure was analogous to the generation of the ‘normal’ template except that these templates were not made left–right-symmetric. For the moderate reduction template, the stereotactically normalized DVR images were left–right-flipped prior to averaging such that the striatal deficit was more pronounced in the right hemisphere in all cases. The final template representative of moderate Parkinson-typical reduction in the right hemisphere was left–right flipped at the midsagittal plane to generate a template representative of moderate Parkinson-typical reduction in the left hemisphere. This resulted in four different 123I-FP-CIT templates in MNI space representative of normal uptake and different stages of Parkinson-typical reduction of 123I-FP-CIT uptake in the striatum with attenuation and scatter correction.

Four 123I-FP-CIT templates representative of DAT-SPECT without attenuation and scatter correction were generated from the DAT SPECT images of the same patients reconstructed from the same projection data using the iterative ordered-subsets-expectation–maximization algorithm on the Hermes SMART workstation without attenuation and scatter correction [21]. The final set of eight 123I-FP-CIT templates is shown in Fig. 1.

Fig. 1figure 1

Templates of the 123I-FP-CIT distribution volume ratio (DVR) in MNI space. The templates are representative of normal striatal signal (left column), moderate (middle columns) and strong (last column) Parkinson-typical reduction of striatal uptake with and without attenuation and scatter correction (ACSC). Each template was generated from twelve randomly selected DAT-SPECT images. The 123I-FP-CIT template representative of normal striatal signal with attenuation and scatter correction (upper left) was used for single template stereotactical normalization

The rationale for using 12 SPECT images for each of the 123I-FP-CIT templates was that the SPM 15O-water template was created from 12 normal 15O-water PET images [22]. That 12 is an adequate number of images for template generation also in DAT-SPECT with 123I-FP-CIT was shown by Kas and co-workers who compared 4 different 123I-FP-CIT templates generated from 5 to 15 normal scans with respect to their impact on semi-quantitative analysis and voxel-based statistical testing [21]. They found that stereotactical normalization with these templates “provided results close enough to consider that the templates can be used interchangeably without altering the clinical interpretation” [21].

Preprocessing of individual DAT-SPECT images

Each of the 1704 individual DAT-SPECT images (with attenuation and scatter correction) that was not used for the generation of the 123I-FP-CIT templates was stereotactically normalized to MNI space using either the single template representative of normal striatal 123I-FP-CIT uptake (with attenuation and scatter correction) as target or the set of the eight different templates. In the latter case, SPM tries to find the linear combination of these templates that best matches the intensities in the patient’s image. All other settings for stereotactical normalization and subsequent intensity scaling to obtain DVR images were as described in subsection “123I-FP-CIT templates”.

Stereotactical normalization was checked for major failures by visual inspection of each of the 1704 stereotactically normalized DAT-SPECT images using a display with three orthogonal views. This was done separately for the normalization results with the single template and with multiple templates.

Semi-quantitative analysis

Two different methods were used for semi-quantitative analysis. First, the 123I-FP-CIT DVR in left and right putamen was estimated by hottest voxels (HV) analysis of the stereotactically normalized DVR image using large unilateral putamen masks predefined in MNI space as described previously [16]. The putamen masks were much bigger than the actual putamen volume in order 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 10 ml.

For comparison, the 123I-FP-CIT DVR in left and right putamen was estimated by conventional analysis, that is, by the mean of the voxel intensities in anatomical ROIs for the unilateral putamen predefined in MNI space. The unilateral putamen masks of the Automated Anatomical Labelling (AAL) atlas were used for this purpose [23].

The 123I-FP-CIT SBR in left and right putamen was obtained from the corresponding DVR according to SBR = DVR − 1, separately for both methods of semi-quantitative analysis. The minimum of the putamen SBR of both hemispheres was used for the further analyses. The mean putamen SBR of both hemispheres was used for comparison.

Statistical analysis

The general linear model for repeated measures was used to test the impact of the templates (multiple templates versus single template), the ROI method (hottest voxels analysis versus anatomical AAL ROIs), and the characteristic (minimum versus mean of both hemispheres) on the SBR. The camera (e.cam with LEHR versus TruePoint with LEHR versus TruePoint with fan-beam versus AnyScan Trio with LEHRHS, Table 1) was included in the model as between-subjects factor.

The distribution of the putamen SBR was characterized by a histogram with bin width of 0.1. The resulting histogram was fitted by the sum of two Gaussians:

$$}\,}) = A_ \exp \left( } - M_ } \right)^ }}}_^ }}} \right) + A_ \exp \left( } - M_ } \right)^ }}}_^ }}} \right),$$

(1)

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, that is,

$$d = \left( - M_ } \right)/\sqrt }_^ + }_^ }}} .$$

(2)

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

$$c = \left( }_ M_ + }_ M_ } \right)/\left( }_ + }_ } \right).$$

(3)

The histogram analysis was performed separately for each combination of templates (multiple templates or single template), ROI method (hottest voxels analysis or anatomical AAL ROIs), and characteristic (minimum or mean of both hemispheres).

In order to assess the robustness of the effect size estimates, the histogram analysis was performed on 1000 random subsamples each comprising 90% of the whole DAT-SPECT sample.

The amount of stretching of individual DAT SPECT images required for stereotactical normalization was characterized by the determinant (DET) of the affine normalization transformation, separately for stereotactical normalization with multiple templates and with the single normal template. The putamen SBR was tested for association with the corresponding DET using linear regression (with constant). The DET of stereotactical normalization with multiple templates was used for the four multiple templates settings, the DET of stereotactical normalization with the single template was used for the four single template settings. The regression analysis was performed separately for DAT-SPECT with normal and reduced SBR, where the setting-specific cutoff according to Eq. (3) was used to categorize SBR as ‘normal’ or ‘reduced’.

The association between the single template DET and the multiple templates DET was tested by linear regression (without constant). The regression analysis was performed separately for DAT-SPECT with normal and reduced SBR. The regression analysis was restricted to cases in which the categorization as normal or reduced by the minimum hottest voxels SBR agreed between stereotactical normalization with the single template and stereotactical normalization with multiple templates.

Statistical analyses were conducted using SPSS version 27 (SPSS Inc., Chicago, Illinois). All p-values are given two-sided. Statistical significance was defined as p < 0.05.

留言 (0)

沒有登入
gif