Gregory Klein, Mehul Sampat, Davis Staewen, David Scott, Joyce Suhy BioClinica Inc, Newark, CA, USA
- We compare results of standard uptake value ratio (SUVR) analyses of Alzheimer’s Disease Neuroimaging Initiative (ADNI) florbetapir PET scans using a native space compared to SPM template methods and a variety of possible SUVR reference regions.
- The objective is to find a method with highest longitudinal effect size to allow a sufficiently powered clinical trial efficacy measure using the smallest number of subjects.
- Quantitative analysis of longitudinal florbetapir data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study were completed for 476 subjects clinically diagnosed as:
- 179 Normal
- 160 Early Mild Cognitive Impairment (EMCI)
- 93 Late Mild Cognitive Impairment (LMCI)
- 44 Alzheimer's Disease (AD)
- The focus of the analysis for this work are SUVR’s for AD vs Normal subjects. Characteristics of these two groups are below:
- Input data were ADNI “level 4 data”: uniformly smoothed, co-registered, averaged dynamic PET data that have been reoriented into a standard voxel grid1
SUVR Reference Regions
- Numerous reference regions were evaluated:
- Brainstem (BS)
- Cerebellar Grey (CG)
- Cerebellar White Matter (WMcereb)
- Subcortical White Matter (WM)
- Eroded Subcortical White Matter (WMeroded)
- WM + Wmcereb (WMall)
- Whole Brain (WB)
- Whole Cerebellum (WC)
- Average of BS, WMeroded, WC (AvgRef)
SUVR Methods (Fig 1)
- Three different quantification methods each with specific reference regions were evaluated for computation of the SUVR:
Freesurfer (FS) SUVR Method
- Four cortical regions were obtained by Freesurfer on the T1-MRI data in native patient space.
- PET data were co-registered to the MRI, and a global SUVR was obtained by averaging the SUVRs of four cortical regions.2
Clark PET-Only (PO) SPM SUVR Method
- PET data were aligned to a common template using SPM8.
- A single PET image representing the average of a population of negative and positive florbetapir scans was used as a reference dataset in the SPM alignment algorithm.
- A global SUVR was obtained as the average of six cortical regions normalized to reference region activity.3
AAL Grey-Masked SPM Method (AAL)
- PET data were normalized to Talairach space using SPM.
- Activity from white matter regions are excluded by masking the PET with a grey matter mask obtained
by an SPM segmentation of the subject’s coregistered
- A global SUVR was obtained as the average of 5 bilateral cortical areas defined on the AAL ROI atlas. 4,5
Note: Coregistered SPM white matter segmentations were also used to define a “WMall” reference region for the PO and AAL methods. Not all reference regions were possible for all SUVR methods.
Comparison Metric – Effect Size
- Cohen’s d effect size is the metric used for comparison of methods:
- For the cross-sectional comparison of AD vs Normal groups, the numerator was the difference of mean SUVRs between the two groups and the denominator was the average standard error of each group SUVR.
- For the longitudinal analysis of AD SUVR change, the numerator was the mean SUVR difference between time points, and the denominator was the SUVR std error.
- Effect size in the cross-sectional analysis showed similar results for most methods, with these general trends:
- Higher effect size using WM references
- Higher effect size using Freesurfer method
- Lowest effect size with cerebellar grey (CG)
- Longitudinal effect size showed large differences between methods, with these general trends:
- Much higher effect with WM references
- Freesurfer method superior across all reference regions
DISCUSSION / CONCLUSIONS
- Results indicate that while effect size in a cross-sectional analysis does not vary greatly across different SUVR methods and reference regions, there is a very large difference seen in the longitudinal analysis. This is an important consideration in the selection of methods for use as an efficacy endpoint in clinical trials.
- A reference region including white matter consistently performs better in both the cross-sectional and longitudinal analyses. Possible reasons for this compared to pons, brainstem or cerebellar references include increased noise of the latter reference regions due to their position at the outer extremes of the axial PET scanner field of view, methodological difficulties in correct segmentation of these regions compared to white matter, or possibly increased biological variability in these regions compared to white matter.
- In both the cross-sectional and longitudinal analyses, and across all reference regions, the native-space Freesurfer method produced the greater effect size. We hypothesize this is most likely due to possible increased accuracy of regional segmentation for individual scans using the Freesurfer method, and to the use of smaller regions less likely to be biased by partial volume effects of uptake seen in white matter and other non-grey matter regions.
- Though the Freesurfer method requires co-registered T1-MRI and extensive computation, it appears that the method can offer improved sensitivity and reduced sample size in longitudinal clinical trial efficacy endpoints.
1. Koeppe et al. ADNI PET Preprocessing 2009 http://adni.loni.ucla.edu/methods/pet-analysis/pre-processing/
2. Landau et al. J Nucl Med 2013; 54:1–8.
3. Clark et al. JAMA. 2011;305:275–283.
4. Adamczuk et al. NeuroImage: Clinical 2 (2013)
5. Barthel et al. Lancet Neurol 2011:10:424-35.
The authors thank Susan Landau and Bill Jagust for extremely helpful discussions about this analysis.