Mehul Sampat1, Florent Roche1, Peng Yu2, Adam J Schwarz2, Joë l Schaerer1, Gennan Chan1, Boubakeur Belaroussi1, Luc Bracoud1, Joyce Suhy1, Joonmi Oh1, and the Alzheimer's Disease Neuroimaging Initiative
1 BioClinica Newark, CA, USA and Lyon, France 2 Eli Lilly and Company, Indianapolis, IN, USA
Accurate and reliable quantification of structural brain changes, within a regulatory compliant framework, is important in clinical trials for Alzheimer's disease (AD).
Cortical changes relevant to AD pathology have been reported in various regions including isthmus cingulate, precuneus, inferior parietal, temporal pole and prefrontal cortex [1,2].
We report the performance of atrophy measurements for these structures using tensor-based morphometry (TBM) , and compare to changes in cortical volumes and thickness measured with longitudinal FreeSurfer (LFS and LFS_Th, respectively)  on the standardized ADNI dataset .
Population and MRI Data
Baseline and Month-3 data from 20 ADNI-2 normal controls (NCs) were used for test-retest purposes (mean age of 77y with range of 67-89y).
Baseline, Month-12 and Month-24 ADNI-1 data were used to quantify percent changes for 493 subjects (99 ADs, 115 LMCI-converters, 115 LMCI-non-converters and 164 NCs).
3D T1-weighted MRI scans from the ADNI-1 and ADNI-2 database (http://adni.loni.ucla.edu) were analyzed in this study.
Whole brain segmentation for both subcortical and cortical structures were done using FreeSurfer (FS) v5.3 analysis suits.
Pair-wise Tensor-Based Morphometry (TBM):
Assessing local expansion or compression using deformable registration and aggregate volume change over a regions of interest.
Novel implementation of log-symmetric demons algorithm with cross-correlation image similarity metric yielding deformation field that is diffeomorphic (computed Jacobians are always positive) and symmetric (reduces the likeliness of any bias in change analysis).
Each follow up visit change was measured relative to Baseline visit.
2) Global longitudinal FS:
Individual subject atlas was built using all available visits.
Absolute symmetrized percent change (ASPC)  was calculated to assess variability.
Generalized areas under curve (AUCs) were calculated and the DeLong test  was performed to compare receiver operating characteristic (ROC) curves among TBM, LFS and LFS_Th methods.
Generalized AUCs were calculated by computing probability that two random subjects are properly ranked with respect to ordinal outcome with two or more levels.
Linear regression was performed to test any bias in change analysis.
Variability using Baseline and Month-3 scans from NCs
Mean ASPC (%) were smaller for TBM than LFS for all regions.
Algorithm comparison results using Baseline, Month-12 and Month-24 scans
DeLong test results
Linear regression results
The TBM method successfully and reliably quantified changes over time for various cortical regions with no bias, and showed improved sensitivity to differentiate subgroups compared to LFS and LFS_Th.
A pair-wise analysis approach to quantify changes using TBM may be a useful option for clinical trial applications, considering the intrinsic limitation of LFS which needs to be run after scans from all required visits are acquired for the best outputs.
 McEvoy, et al., Radiology, 2009;251(1):195-205
 Greene, et al., Neurobiol Aging, 2010;31(8):1304-1311
 Vercauteren et al., Med Image Comput Comput Assist Interv, 2008;11:754-761
 Fischl, et al., Cerebral Cortex, 2004;14:11-22.
 Wyman, et al., Alzheimers Dement. 2013;9(3): 332–337
 Reuter, et al., Neuroimage, 2012;61(4):1402-1418
 DeLong, et al., Biometrics, 1988;44(3):837-845