The primary finding of this study was that the use of multiple templates representative of normal and Parkinson-typical reduction of striatal 123I-FP-CIT uptake resulted in increased effect size of the difference between normal and reduced putamen SBR obtained by automatic semi-quantitative analysis compared to a single template representative of normal striatal 123I-FP-CIT uptake. This suggests that the power for the detection of nigrostriatal degeneration by automatic semi-quantitative analysis of DAT-SPECT might be improved by the use of multiple templates. The benefit of multiple templates versus a single template was observed for both methods of semi-quantitative analysis, hottest voxels analysis and conventional analysis with anatomical ROIs, and for both methods of combining left and right hemisphere, minimum and mean value (Fig. 3).
To some extent this can be explained by the fact that multiple templates removed the trend of a positive correlation between the putamen SBR and the determinant DET of the affine transformation for stereotactical normalization that was observed with the single normal template as target in cases with reduced SBR (standardized regression coefficient ß = -0.027, p = 0.457, with multiple templates versus ß = 0.070, p = 0.056, with the single template; Fig. 4). The DET characterizes the amount of stretching that is required to map individual DAT-SPECT to MNI space (representative of a rather large brain [22]). Thus, positive correlation of the putamen SBR with the DET suggests overestimation of the putamen SBR by overestimation of the amount of stretching resulting in artificially enlarged putamen size. This was confirmed by the finding of almost 10% larger DET in reduced cases with the single template compared to multiple templates: DET(single template) = 1.096 ⋅ DET(multiple templates).
The scatter plot of the minimum hottest voxels putamen SBR obtained with the single normal template as target for stereotactical normalization versus stereotactical normalization with multiple templates identified one clear outlier (Figs. 6). A transversal image of the DAT-SPECT of the outlier after stereotactical normalization with the single normal template is shown in Fig. 7. The 80%-isodensity contour of the outlier’s left striatum copied to the normal template suggests that the outlier’s striatum with reduced 123I-FP-CIT uptake in the putamen was slightly shifted towards the posterior brain. As a consequence, part of the caudate nucleus with normal 123I-FP-CIT uptake ended up in the putamen mask in MNI space and resulted in strong overestimation of the putamen SBR. This effect was largely avoided by using multiple templates for stereotactical normalization including one with similar striatal signal as this patient (Fig. 7). This probably also contributed to the benefit from the multiple templates with respect to the differentiation of normal and reduced putamen SBR.
However, the impact of the stereotactical normalization on binary categorization of DAT-SPECT as normal or reduced was rather small. Binary categorization by the minimum hottest voxels putamen SBR with multiple templates or the single normal template for stereotactical normalization were concordant in 1665 of 1702 cases (97.8%, Table 3). Amongst the 37 cases with discrepant categorization (2.2%), there were almost four times more false negative cases (n = 29) with the single template (multiple templates as standard-of-truth) than false positive cases (n = 8). This suggests that potential improvement of discriminative power by multiple templates might be mainly driven by improved sensitivity for the detection of nigrostriatal degeneration.
The fact that the impact of multiple templates on the binary categorization of DAT-SPECT was rather small, to some extent might be due to a ceiling effect: any reasonable method for semi-quantitative DAT-SPECT analysis provides relatively high diagnostic accuracy that leaves room for small improvement only [8]. This is explained by the rather high symptom threshold in neurodegenerative Parkinsonian syndromes. Post-mortem studies have shown that motor symptoms in Parkinson’s disease start at rather advanced stages of nigrostriatal degeneration when loss of DAT in the unilateral putamen has reached about 50% [23]. However, small improvement of the discriminative power of semi-quantitative analysis in DAT-SPECT that might be achieved by the use of multiple templates for stereotactical normalization comes practically at no costs. In particular, additional procedures (e.g., blood sampling) or prolonged scan duration (e.g., for dynamic imaging) are not required. The multiple templates have to be generated only once. The multiple templates generated for this study are freely available on request.
An additional (small) benefit from the use of multiple templates was the complete lack of normalization failures with multiple templates. There was a very small proportion (2 of 1704, 0.1%) of normalization failures with the single template. These failures most likely were associated with rather prominent 123I-FP-CIT uptake in the salivary glands (supplementary Figure S1), since the failure could be avoided by manual cropping of the salivary glands prior to stereotactical normalization in both cases. However, visual quality check of stereotactical normalization is mandatory in automatic semi-quantitative analysis independent of the template(s) used as target for stereotactical normalization.
Amongst the 1665 cases with concordant categorization, 55.9% were categorized as normal, 44.1% as reduced. This is in line with the recommendation to restrict DAT-SPECT to clinically uncertain cases, as this recommendation implies an about 50% pre-test probability of nigrostriatal degeneration.
There was a rather strong positive correlation between the putamen SBR and the determinant DET of the affine transformation for stereotactical normalization in normal DAT-SPECT. This was observed with the single template and with multiple templates as target for stereotactical normalization (Fig. 3), although the correlation might be slightly weaker with multiple templates (ß = 0.291 versus ß = 0.330, both p < 0.0005). This was an unexpected finding. We rather expected a negative correlation between the putamen SBR and DET, based on the following rationale: larger DET indicates a smaller brain in native patient space and, therefore, more pronounced underestimation of the actual putamen SBR due to partial volume effects. The observed positive correlation between the putamen SBR and the DET in the normal DAT-SPECT is not specific for the hottest voxels method used for semi-quantitative analysis, since the same correlation was observed with the conventional semi-quantitative analysis using AAL putamen masks (results not shown). We hypothesize that the positive correlation between the putamen SBR and the DET was driven by residual variability of spatial resolution in the DAT-SPECT images despite harmonized image reconstruction, at least to some extent (supplementary Figure S2).
The use of multiple templates for stereotactical normalization has proven beneficial also in other nuclear brain imaging procedures, e.g. positron emission tomography with FDG [24–26] or ligands for senile amyloid plaques [27].
The set of templates used for the multiple templates approach in the current study combined templates with attenuation and scatter correction and templates without attenuation and scatter correction. The rationale for this was that there is considerable between-subjects variability of the 123I-FP-CIT uptake in the scalp. DAT-SPECT images without attenuation and scatter correction show more prominent signal in the scalp compared to DAT-SPECT images with attenuation and scatter correction (Fig. 1). Thus, inclusion of the templates without attenuation and scatter correction resulted in increased variability of the scalp signal amongst the multiple templates. This was expected to allow better modeling of the DAT-SPECT images of individual patients by linear combinations of the multiple templates. Furthermore, there is no risk of overfitting here so that additional templates in general are expected to improve the performance of the multiple templates approach.
The 123I-FP-CIT template representative of normal striatal 123I-FP-CIT uptake was made left-right-symmetric by left-right flipping and then averaging the original and the flipped image. The 123I-FP-CIT template representative of moderate Parkinson-typical reduction in the left hemisphere was obtained by left-right flipping of the 123I-FP-CIT template representative of moderate Parkinson-typical reduction in the right hemisphere. The rationale for this was to eliminate left-right asymmetry in the template(s) that might cause artificial left-right-asymmetry in SBR estimates.
Secondary findings of this study were that (i) hottest voxels analysis outperformed conventional semi-quantitative analysis with anatomical putamen ROIs predefined in template space and (ii) the minimum of the putamen SBR of both hemispheres outperformed left and right SBR as well as the mean of both hemispheres with respect to the differentiation of normal and reduced SBR (Table 2, Fig. 3), in line with previous studies [14, 28].
Finally, this study proposes the effect size of the distance between two Gaussians fitted to the histogram of the putamen SBR in a patient sample as clinically relevant quality metric. This metric requires a sufficiently large sample size so that the shape of the histogram is sufficiently smooth and stable with respect to the choice of histogram bins. In this study, 1702 DAT-SPECT were included in the analysis, which proved sufficient.
The following limitations of this study should be noted. First, there was no standard-of-truth diagnosis available, e.g., clinical diagnosis by a movement disorder specialist ≥ 2 years after DAT-SPECT. Thus, the impact of single versus multiple templates for stereotactical normalization on the performance of the putamen SBR to differentiate between neurodegenerative and non-neurodegenerative parkinsonian syndromes could not be assessed by conventional performance measures such as sensitivity and specificity. Second, the templates were generated from images reconstructed with the same iterative algorithm as the individual DAT-SPECT used for the performance testing. The impact of different reconstruction algorithms in non-harmonized settings was not tested. Third, the impact of single versus multiple templates was not tested in DAT-SPECT with atypical reduction of the striatal signal, e.g., most pronounced reduction in the caudate nucleus (e.g., due to a vascular/structural lesion).