European Journal of Nuclear Medicine and Molecular Imaging

, Volume 34, Issue 4, pp 496–501

Quantification of Parkinson’s disease-related network expression with ECD SPECT

Authors

  • Thomas Eckert
    • Center for NeurosciencesThe Feinstein Institute for Medical Research
    • Department of Neurology IIUniversity of Magdeburg
  • Koen Van Laere
    • Division of Nuclear MedicineLeuven University Hospital
  • Chengke Tang
    • Center for NeurosciencesThe Feinstein Institute for Medical Research
  • Daniel E. Lewis
    • Center for NeurosciencesThe Feinstein Institute for Medical Research
  • Christine Edwards
    • Center for NeurosciencesThe Feinstein Institute for Medical Research
  • Patrick Santens
    • Department of NeurologyGhent University Hospital
    • Center for NeurosciencesThe Feinstein Institute for Medical Research
    • Department of NeurologyNorth Shore University Hospital and New York University School of Medicine
Original article

DOI: 10.1007/s00259-006-0261-9

Cite this article as:
Eckert, T., Van Laere, K., Tang, C. et al. Eur J Nucl Med Mol Imaging (2007) 34: 496. doi:10.1007/s00259-006-0261-9
  • 147 Views

Abstract

Purpose

Spatial covariance analysis has been used with FDG PET to identify a specific metabolic network associated with Parkinson’s disease (PD). In the current study, we utilized a new, fully automated voxel-based method to quantify network expression in ECD SPECT images from patients with classical PD, patients with multiple system atrophy (MSA), and healthy control subjects.

Methods

We applied a previously validated voxel-based PD-related covariance pattern (PDRP) to quantify network expression in the ECD SPECT scans of 35 PD patients, 15 age- and disease severity-matched MSA patients, and 35 age-matched healthy control subjects. PDRP scores were compared across groups using analysis of variance. The sensitivity and specificity of the prospectively computed PDRP scores in the differential diagnosis of individual subjects were assessed by receiver operating characteristic (ROC) analysis.

Results

PDRP scores were significantly increased (p < 0.001) in the PD group relative to the MSA and control groups. ROC analysis indicated that the overall diagnostic accuracy of the PDRP measures was 0.91 (AUC). The optimal cutoff value was consistent with a sensitivity of 0.97 and a specificity of 0.80 and 0.71 for discriminating PD patients from MSA and normal controls, respectively.

Conclusion

Our findings suggest that fully automated voxel-based network assessment techniques can be used to quantify network expression in the ECD SPECT scans of parkinsonian patients.

Keywords

ImagingBrain SPECTPerfusionParkinson’s diseaseMultiple system atrophy

Introduction

Parkinsonian features such as tremor, rigidity, and hypokinesia can be present in several neurodegenerative disorders that clinically resemble idiopathic Parkinson’s disease (PD). Differentiating these conditions from PD is very difficult, especially at the earliest stages of clinical disease. While a combination of asymmetric symptoms, presence of resting tremor, and responsiveness to levodopa support the early diagnosis of PD [1, 2], only 75% of patients with these clinical signs ultimately prove to have this diagnosis at postmortem examination [3].

Atypical parkinsonian disorders are less common than classical PD [4]. Of these syndromes, multiple system atrophy (MSA) is the most common and the most frequently misdiagnosed. A variety of structural and molecular imaging techniques have been investigated for their potential to differentiate objectively between these syndromes [59].

While many MRI-based techniques have been developed exclusively as research tools, radiotracer-based imaging methods have recently been used in routine clinical settings. Nonetheless, these methods, while quite sensitive descriptors of presynaptic nigrostriatal dopaminergic dysfunction, are less suited to differentiate between classical PD and atypical parkinsonian syndromes in which nigral disease occurs in the company of differing patterns of downstream pathology.

By contrast, [18F]fluorodeoxyglucose (FDG) positron emission tomography (PET) has proven to be a valuable tool for the differential diagnosis of parkinsonian disorders, even at the earliest stages of disease [10]. Indeed, we have shown that classical PD is associated with a reproducible disease-related metabolic covariance pattern (PDRP) characterized by increases in pallidothalamic and pontine activity associated with relative reductions in cortical motor regions and parieto-occipital association areas [11, 12]. The expression of this pattern can be assessed prospectively in individual subjects and can differentiate between patients with classical PD, patients with other parkinsonian disorders, and healthy volunteer subjects [1315]. PDRP expression can be suppressed by effective treatment such as dopaminergic therapy [16, 17], as well as a variety of stereotaxic surgical procedures for advanced PD [1719].

Because PET is not a widely accessible modality, use of this technique as a diagnostic tool for PD is limited. Although most PDRP topographies have been identified in patient populations scanned in the rest state with FDG PET, other imaging modalities can be used to quantify network expression in individual subjects. Indeed, we have recently demonstrated that PDRP scores are highly intercorrelated in FDG and 15O-water PET images obtained in the same subjects [12]. We have previously demonstrated that PDRP expression can be quantified in ethylene cysteinate dimer (ECD) single-photon emission computed tomography (SPECT) data utilizing potentially cumbersome region of interest (ROI)-based methods [20]. However, this analysis has the potential to not include all the regions that may be contributing to the network, and also mitigates the effect of potentially important areas in which increases and/or decreases in metabolic activity occur with distinct subregions.

With the development of fully data-driven network quantification methods, we have been able to identify and quantify the expression of disease-related patterns on a voxel-by-voxel basis [12, 15, 17, 19]. Indeed, this method has allowed for accurate group separation on a prospective case basis, without the assumptions and potential bias associated with the ROI-based approach.

In the present study, we used this voxel-based network approach to quantify PDRP expression in ECD SPECT scans from patients with classical PD or MSA, as well as age-matched healthy volunteer subjects. Computed PDRP scores obtained using this blinded, fully automated method were compared across diagnostic groups and assessed for sensitivity and specificity of current diagnosis.

Materials and methods

Subjects

Thirty-five PD patients [18 men, 17 women; age 62.9 ± 9.1 years (mean±SD); Hoehn & Yahr Stage [21] (H&Y) 2.8 ± 0.7; disease duration 9.8 ±4.0 years], 15 MSA patients (9 men, 6 women; age 61.5±9.2 years; H&Y 2.8±1.0; disease duration 3.0±2.2 years), and 35 age-matched healthy control subjects (23 men, 12 women; age 62.0 ±10.0 years) underwent SPECT imaging with 99mTc-ECD. Subjects in this study were recruited and scanned at the University Hospital of Ghent, Belgium. All patients were assessed by the same investigator (P.S.). All subjects were non-demented and the clinical diagnoses of the PD and MSA patients were based on the UK Brain Bank criteria [3] and the Quinn criteria [22], respectively. The local ethics committee approved the study and all subjects gave written informed consent. Limited imaging and clinical data for these subjects have been reported by us previously in the context of a correlation study between symptoms and brain perfusion [9].

SPECT

SPECT imaging was performed 30 min after intravenous injection of 925 MBq (25 mCi) (±5%) of 99mTc-labeled ECD (Neurolite; Bristol-Myers-Squibb/Dupont Pharmaceuticals). Scanning was performed with eyes closed in a dimly lit room with minimal background noise. All patients were scanned while on antiparkinsonian medication.

SPECT scans for the healthy control group and 34 patients were acquired on a dual-head Helix camera (Elscint/General Electric). These scans were acquired over a 20-min period, with 60 angles through 360° in step-and-shoot mode. The remaining 16 patient scans were acquired on a triple-head gamma camera (Toshiba GCA-9300A; Dutoit Medical). These scans were acquired over a 20-min period, with 90 angles through 360° in continuous mode and data rebinned in 4° parallel bins. Both camera systems were equipped with low-energy, high-resolution, parallel-hole collimators.

All acquisitions were done in a 128×128 matrix. Image reconstruction was performed using filtered back-projection (HERMES; Nuclear Diagnostic) with a Butterworth filter (order 8, cutoff 0.40 cm−1). Uniform Sorenson attenuation correction was used with an attenuation coefficient of 0.09 cm−1. A variable outline of the head was utilized for the attenuation correction based on slice-by-slice ellipses defined on the activity contours (threshold 50%) (Hermes, Nuclear Diagnostics, Sweden). For the infratentorial slices, this was adjusted manually slice by slice to include the skull and facial bones. Triple-head data were corrected to the lower-resolution parallel dual-head data using a previously validated transfer procedure based on isotropic resolution correction [23].

Data analysis

Anatomic standardization was performed by non-linear transformation using 3 × 4 × 3 basis functions and five non-linear iterations to the generic SPECT template using SPM99 (Wellcome Department of Cognitive Neurology, London, UK). A resulting voxel size of 2 × 2 × 2 mm was used. The normalized scans were smoothed with an isotropic 14-mm kernel.

Network analysis

In this study, we used the PDRP that we identified previously in a voxel-based principal components analysis (PCA) of FDG PET scans from 20 PD patients and 20 age-matched normal control subjects [15, 17]. The FDG PET images that were used to generate PDRP were standardized to the PET template of SPM99. The expression of this spatial covariance pattern in each of the ECD SPECT scans was computed on a prospective single case basis using an automated algorithm.

The details of this procedure are presented elsewhere [12, 24] and software is available on-line at http://neuroscience-nslij.org/Methods/software.html. The resulting values were Z-transformed and offset so that the control group had a mean PDRP score of zero [15]. PDRP scores were compared across groups using ANOVA with post hoc Bonferroni correction [25]. We performed a power analysis utilizing Gpower software (available at http://www.psycho.uni-duesseldorf.de/aap/projects/gpower/index.html). To assess the accuracy of discrimination for PD and non-PD patients based upon the computed PDRP scores, we used ROC analysis [26]. The area under the curve (AUC) was used as an overall measure of the diagnostic accuracy of this test in distinguishing between PD and MSA patients as well as PD patients and healthy control subjects. Statistical comparisons and correlational analyses were performed using JMP 5.0 software (SAS Institute Inc., Cary, NC, USA).

Results

We found that PDRP expression was significantly different between the groups (ANOVA, p < 0.001). PDRP expression scores were elevated (post hoc test with Bonferroni correction, p < 0.001) in the ECD SPECT scans of the PD group (mean subject score ±SE, 7.8±0.9) relative to the MSA patients and the healthy volunteers (mean scores −3.1±1.8 and 0.0±0.55 for the two groups, Fig. 1). A trend (p = 0.1) was noted toward reduced PDRP expression in the MSA group relative to controls. Considering the group sizes of 35 PD patients, 15 MSA patients, and 35 healthy control subjects, this analysis has a statistical power of 0.98 at p < 0.001.
https://static-content.springer.com/image/art%3A10.1007%2Fs00259-006-0261-9/MediaObjects/259_2006_261_Fig1_HTML.gif
Fig. 1

Bar histogram displaying mean subject scores for the PET-derived PD-related spatial covariance pattern (PDRP) in ECD SPECT perfusion images. Scores were determined prospectively in each individual ECD SPECT scan using an automated routine that was blind to diagnostic category. PDRP expression is increased in the PD group (gray bar) when compared with MSA (dark gray bar) and normal control subjects (white bar). [Error bars represent standard error of the mean. Arrows indicate the level of significance between the groups (ANOVA and post hoc Bonferroni correction).]

ROC analysis demonstrated an AUC of 0.91 for PD versus MSA (Fig. 2) and 0.91 for PD versus normal (Fig. 3). With an optimally chosen cutoff value of 1.5 for PDRP expression, all but one of the PD patients (97.1%) were correctly classified, as were 12 of the 15 MSA patients (80%) (Table 1). The same cutoff value also optimally separated PD patients from healthy control subjects with a sensitivity of 0.97 and a specificity of 0.71 (Table 2).
https://static-content.springer.com/image/art%3A10.1007%2Fs00259-006-0261-9/MediaObjects/259_2006_261_Fig2_HTML.gif
Fig. 2

ROC curves for the discrimination between PD patients and disease severity- and age-matched MSA patients using the expression of the PET-derived PD pattern in the SPECT perfusion data. The AUC was 0.91. The optimally chosen PDRP cutoff value is represented by an asterisk. A diagonal dotted line is included for reference, corresponding to an AUC of 0.5

https://static-content.springer.com/image/art%3A10.1007%2Fs00259-006-0261-9/MediaObjects/259_2006_261_Fig3_HTML.gif
Fig. 3

ROC curves for the discrimination between PD patients and age-matched healthy control subjects using the expression of the PET-derived PD pattern in the SPECT perfusion data. The AUC was 0.91. The optimally chosen PDRP cutoff value is represented by an asterisk. A diagonal dotted line is included for reference, corresponding to an AUC of 0.5

Table 1

PDRP cutoff values and their respective sensitivities, specificities, positive predictive values (PPV) and negative predictive values (NPV) in separating PD patients from MSA patients

PDRP cutoff value

Sensitivity

Specificity

PPV

NPV

0

0.97

0.73

0.89

0.92

1.5

0.97

0.80

0.92

0.92

3

0.77

0.80

0.9

0.6

4.5

0.66

0.93

0.96

0.54

6

0.57

0.93

0.95

0.48

The optimal cutoff value is in bold.

Table 2

PDRP cutoff values and their respective sensitivities, specificities, positive predictive values (PPV) and negative predictive values (NPV) in separating PD patients from healthy control subjects

PDRP cutoff value

Sensitivity

Specificity

PPV

NPV

0

0.97

0.60

0.71

0.95

1.5

0.97

0.71

0.77

0.96

3

0.77

0.74

0.75

0.76

4.5

0.66

0.86

0.82

0.71

6

0.57

0.94

0.91

0.69

The optimal cutoff value is in bold.

Discussion

The measurement of PDRP expression with FDG PET has been shown to be a valuable objective imaging marker of parkinsonism (for review see [11]). In addition to being expressed in multiple PD patient cohorts [13, 14], including early stage patients [27], PDRP scores have also been shown to correlate consistently with disease severity and duration [28, 29], as well as with treatment effects [11, 17].

In previous imaging studies, we have shown that the PD-related spatial covariance patterns can be used to differentiate classical PD and atypical variants [30, 31]. Indeed, utilizing a computer-assisted method, we found that FDG PET can potentially differentiate between patients with PD and atypical parkinsonian conditions with an accuracy of approximately 90% [10]. However, prior studies utilized methods that were investigator-dependent involving either manual ROI placement [30, 31] or a decision tree-guided evaluation of regional changes [10].

In contrast to the ROI-based methods used previously to assess network expression in ECD SPECT data [20], the voxel-based approach used in the current study does not require preliminary assumptions about the involved regions. We also examined the potential of this fully automated, investigator-independent method in the diagnostic assessment of individual subjects. Indeed, employing an optimal cutoff value, we were able to correctly classify 97% of the PD patients and 80% of the MSA patients without user interaction. We also found that this PDRP value separated PD patients from control subjects with the same high level of sensitivity (0.97) but with a slightly lower level of specificity (0.71). The consistent high sensitivity [true positives/(true positives+false negatives)] shows the accuracy of our method for revealing increased PDRP expression in PD patients. Specificity [true negatives/(true negatives+false positives)] reflects the ability of our method to detect low PDRP expression in non-PD subjects. The specificity value in the MSA patients is slightly better than that in the control subjects. We attribute this difference to the trend toward lower PDRP expression in the MSA patients relative to the normal control subjects. Overall, the accuracy of this “black box” technique is reflected by the high AUC of the ROC curve for discriminating patients with classical PD from those with atypical variant conditions like MSA.

The finding that the PET-derived PDRP is expressed in ECD SPECT scans from PD patients suggests that brain perfusion techniques have the potential to be used in conjunction with network quantification algorithms to assess the effects of novel therapies for this disorder. Moreover, it is possible that this approach could be utilized with other imaging techniques such as perfusion-weighted MRI (e.g. [32]). A potential limitation of our study is that SPECT imaging was performed while the patients were on dopaminergic medication. Indeed, we have previously shown that levodopa infusion changes regional metabolism and reduces PDRP expression in PD patients [16, 17]. By contrast, regional brain perfusion abnormalities are thought to be unchanged with levodopa in MSA patients [33]. Therefore, ECD SPECT imaging after withdrawal of dopaminergic medication is likely to result in group differences in PDRP expression that are more profound than those seen in this study.

We note that this study only investigated the expression of the PDRP network and does not provide the basis for further diagnostic inferences. For instance, one cannot differentiate between MSA patients and healthy controls based upon this spatial covariance pattern, although other disease-related patterns may exist for this purpose [34]. Nonetheless, the PDRP performed relatively well in discriminating PD patients from their MSA counterparts when matched for disease stage. The differentiation of these two groups is potentially more challenging when disease duration is short [35], as may be the case in trials of potentially neuroprotective agents. Automated diagnostic procedures incorporating specific disease-related covariance patterns for MSA and other atypical variant conditions, in addition to the PDRP, may be useful in this regard [24].

Copyright information

© Springer-Verlag 2006