Head-to-Head Comparison of Image Analysis Methods in Longitudinal Alzheimer's Disease FDG Pet Studies

Gregory Klein¹, Mehul Sampat¹, Davis Staewen¹, David Scott¹, Joyce Suhy¹, Eric M Reiman², Kewei Chen²

¹Bioclinica, Newark, CA, USA; ²Banner Alzheimer's Institute, Banner Health, Phoenix, AZ, USA

OBJECTIVES

The Alzheimers Disease Neuroimaging Initiative (ADNI) PET Core has analyzed FDG images using ROI, summary index and voxel-based approaches1.We compare longitudinal effect sizes for these methodologies along with Freesurfer ROI approaches using the same ADNI datasets for all methods.

METHODS

Population

• Quantitative analyses of baseline and 24-m FDG data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study were completed for 399 subjects who had at least two FDG with corresponding MRI. . All subjects also had Florbetapir scans for classification of Aβ status. Subject demographics are shown in Table 1. The dataset used for comparison clinically diagnosed as:
  • 158 Normal (N) (52 Aβ+, 106 Aβ-)
  • 208 Mild Cognitive Impairment (MCI) (108 Aβ+, 100 Aβ-)
  • 33 Alzheimer's Disease (AD) (27 Aβ+, 6 Aβ-)
• Clinical status determined from Baseline Dx at visit one, adjusted if Dx at visit two showed improved status.
• Aβ status (determined via baseline Florbetapir PET, 1.1 threshold Freesurfer Whole Cerebellar reference) was used to stratify the analysis into “likely decliner” (N Aβ+, MCI Aβ+ & AD Aβ+) vs “likely stable” (N Aβ-) groups.

Indices of cortical hypometabolism were computed using five different methods:

1. SPM “MetaROI” 2
2. Statistical ROI (sROI) 3
3. Longitudinal Hypermetabolic Convergence Index (L-HCI) 4
4. Freesurfer ROI (FS-WB, FS-WM)
5. LASSO-optimized Freesurfer (LASSO-FS-WB, LASSO-FS,WM)

Methods 1-3 have been described previously (Fig 1). Methods 4 and 5 used a Freesurfer (FS) MRI segmentation to obtain SUVR indices from co-registered PET data. Both Freesurfer methods evaluated whole brain (WB) and subcortical white matter (WM) reference regions and target regions known to be affected by AD (Figs 1,2). Longitudinal Cohen’s effect size was evaluated for each method, as well as the absolute difference in measures between time points.

RESULTS

Longitudinal Effect Size

• All analysis methods show increasing absolute longitudinal difference and effect sizes for the Aβ+ Normals, Aβ+ MCIs, and Aβ+ AD decliner groups respectively compared to the Aβ- Normals (Fig 3). While the MetaROI method showed the largest absolute difference (Fig 4), effect size for the LASSO-FS and L-HCI showed the highest effect sizes. The sROI, FS and MetaROI methods showed a smaller effect size. Comparing among the Freesurfer methods, the LASSO-optimized ROIs were superior and the greatest effect size was seen using the WM reference.

Sample Size Considerations

• Sample size estimates per arm were calculated to detect a 25% treatment effect with 80% power and 5% type I error, assuming two-sample, two-sided tests (Table 3). An optimal analysis can reduce required sample size dramatically.image

CONCLUSION

• Though the MetaRoi method shows the greatest longitudinal difference, it actually has the smallest effect size of all methods. Considerable reduction in sample sizes to adequately power clinical trials may be possible using alternate methods, such as the LASSO-FS-WM or L-HCI techniques.

ACKNOWLEDGMENTS

The authors are extremely thankful for the work of Asha George for her statistical analysis of the data.

REFERENCE

• Jagust, et al. Alz&Dementia 6 (2010):221-229.
• Landau, et al. Neurology 2010;75:230-238
• Chen, et al. Neuroimage 2010;51(2):654-664.
• Chen, et al. HAI 2016. PO-49, p 151.
• Bauer, et al. J Alz Dis Parkinson. 2013,3:1
• Tibshirani. J Royal Stat Soc, Series B, 1996,267-288