E ects of normalization approach and global covariates on voxel-based morphometry: Comparing DARTEL and standard SPM approaches using age-related cortical change
Jonathan E. Peelle, Rhodri Cusack, Richard N. A. Henson MRC Cognition and Brain Sciences Unit, Cambridge, UK
Background
• Voxel-based morphometry (VBM) is routinely used to assess regional gray matter volume in both healthy participants and patient populations. Despite wide usage, methodological decisions which have a potentially large impact on the results are seldom subjected to systematic comparison. • Here we assess the effects of two methodological decisions: the normalization approach used to register a group of participants’ brains into a standard space (cf. Klein et al., 2009), and the various ways global measures such as total gray matter (TGM) or total intracranial volume (TIV) are accounted for in the model. • We examined these issues in a large (N=330) sample of healthy adults ranging in age from 18–78, equally distributed across decades (cf. Good et al., 2001) , allowing us to draw inferences regarding cortical volume change in healthy aging.
T-statistics: Linear decrease with age
Standard SPM
No covariates
All age effects
DARTEL-MNI
• Using a consistent range for resulting t maps, it is apparent that when not accounting for TGM changes, significant age-related decline is found throughout much of the brain. Proportional and ANCOVA approaches reduce this effect. • Standard SPM VBM appears to result in more edge effects, possibly due to registration errors. • Large global GM decreases suggest a more helpful question is the regionally-differentiated rate of decline, as opposed to merely which voxels show a significant decline.
Proportional TGM ANCOVA
Method
• 330 participants, 55 per decile between the ages of 18–78. All scans conducted on a Siemens 3T Tim Trio scanner at MRC CBU using MPRAGE sequence, 1mm isotropic voxels. • Standard SPM (SSPM, comparable to “optimized” VBM) and DARTEL (Ashburner, 2007) analyses were conducted using comparable parameters (voxel size = 2mm3, smoothing = 8 mm FWHM, identical explicit mask constructed from thresholding the mean GM image from the Standard SPM analysis). • SPM5 was used for all processing. Analysis was performed on modulated images only.
Template
Removing age effects matching any slope related to TGM
Removing age effects matching slope of TGM
6 t
16
Standard SPM
Unified segmentation
* input into DARTEL
DARTEL
Template
ROI Analysis
• To further assess the effect of global values on results we used regions defined in the AAL atlas to perform a region of interest analysis, and ranked regions by the linear parameter estimate obtained. • Notable differences in ranking were present, illustrated for superior frontal gyrus and amygdala.
Cerebelum_3_L Pallidum_L Cerebelum_10_L Thalamus_L Putamen_L Caudate_L Paracentral_Lobule_L Cerebelum_4_5_L Hippocampus_L Olfactory_L Occipital_Sup_L Parietal_Sup_L Cerebelum_9_L Heschl_L Cerebelum_Crus2_L Postcentral_L ParaHippocampal_L Supp_Motor_Area_L Precentral_L Frontal_Inf_Orb_L Frontal_Sup_Medial_L Occipital_Inf_L Temporal_Inf_L Frontal_Sup_Orb_L Frontal_Sup_L Temporal_Pole_Sup_L Cerebelum_7b_L Temporal_Sup_L Lingual_L Insula_L Occipital_Mid_L Precuneus_L Cerebelum_8_L Rolandic_Oper_L Frontal_Inf_Tri_L Frontal_Mid_Orb_L Cerebelum_Crus1_L Rectus_L Calcarine_L Frontal_Mid_L Temporal_Mid_L Cuneus_L SupraMarginal_L Frontal_Inf_Oper_L Frontal_Med_Orb_L Temporal_Pole_Mid_L Angular_L Parietal_Inf_L Fusiform_L Amygdala_L Cingulum_Ant_L Cingulum_Post_L Cingulum_Mid_L Cerebelum_6_L −3.5
Template Mean GM
Mean GM
Warp to template Affine transform to MNI Smooth (8mm FWHM)
No covariates
Smooth (8mm FWHM)
Global Effects
2 1.5 20
M F
0.8 0.6 20
M F
TGM/TIV (liters)
2.5 TGM (liters) TIV (liters)
1
0.5 0.4 0.3
M F
Cerebelum_3_L Olfactory_L Occipital_Inf_L Paracentral_Lobule_L Temporal_Inf_L Frontal_Sup_Medial_L Caudate_L Frontal_Sup_Orb_L Rolandic_Oper_L Precuneus_L Cerebelum_9_L Pallidum_L Cerebelum_4_5_L Supp_Motor_Area_L Occipital_Sup_L Hippocampus_L Frontal_Med_Orb_L Thalamus_L ParaHippocampal_L Calcarine_L Heschl_L Frontal_Mid_Orb_L Parietal_Sup_L Lingual_L Frontal_Sup_L Temporal_Mid_L Postcentral_L Fusiform_L Frontal_Mid_L Cerebelum_10_L Cuneus_L Temporal_Sup_L Frontal_Inf_Orb_L Occipital_Mid_L Cerebelum_8_L Precentral_L Putamen_L SupraMarginal_L Rectus_L Angular_L Insula_L Temporal_Pole_Sup_L Amygdala_L Cerebelum_7b_L Cerebelum_Crus2_L Cingulum_Post_L Cerebelum_Crus1_L Temporal_Pole_Mid_L Frontal_Inf_Tri_L Parietal_Inf_L Cingulum_Ant_L Cingulum_Mid_L Frontal_Inf_Oper_L Cerebelum_6_L x 10
−3
Proportional
Cerebelum_3_L Cerebelum_4_5_L Hippocampus_L Thalamus_L Olfactory_L Cerebelum_9_L Caudate_L Putamen_L ParaHippocampal_L Pallidum_L Cerebelum_Crus2_L Amygdala_L Temporal_Inf_L Cerebelum_10_L Cerebelum_8_L Lingual_L Occipital_Sup_L Insula_L Paracentral_Lobule_L Occipital_Inf_L Cerebelum_7b_L Cerebelum_Crus1_L Frontal_Inf_Orb_L Rolandic_Oper_L Heschl_L Supp_Motor_Area_L Temporal_Sup_L Parietal_Sup_L Temporal_Pole_Sup_L Fusiform_L Calcarine_L Frontal_Sup_Orb_L Postcentral_L Precuneus_L Frontal_Sup_Medial_L Temporal_Mid_L Rectus_L Frontal_Inf_Tri_L Precentral_L SupraMarginal_L Occipital_Mid_L Frontal_Mid_Orb_L Frontal_Sup_L Temporal_Pole_Mid_L Cuneus_L Frontal_Med_Orb_L Frontal_Mid_L Frontal_Inf_Oper_L Cingulum_Ant_L Angular_L Cingulum_Mid_L Parietal_Inf_L Cingulum_Post_L Cerebelum_6_L 1 x 10
−3
TGM ANCOVA
Single subject
Single subject
−3
−2.5
−2
−1.5
−1
−0.5
−1.5
−1
−0.5
0
0.5
−1
−0.5
0
0.5
1
1.5
2
40
Age (years)
60
80
40
Age (years)
60
80
20
40
Age (years)
60
80
TGM TGM ANCOVA Full model fit
x 10
−9
Left mSFG Gray Matter Volume Gray Matter Volume
Left Amygdala Gray Matter Volume
Left temporal pole
• Values for tissue class partitions were determined using each individual’s segmented structural image; TIV was estimated by summing these values. • TIV remains relatively stable throughout the lifespan, but TGM shows a significant decline (~2025%). Scaling TGM by TIV (top right) appears to remove some variability but does not ameliorate the dominant age effects. • Also notable is that when TGM is entered as a covariate (ANCOVA approach), the parameter estimate varies across voxels. This corresponds to different slopes of linear effects which are accounted for (right).
Total GM (collapsed M/F)
1 0.9 0.8 0.7 0.6 0.5 20 40 60 Age (years)
4
Across all individuals TGM shows large age-related declines dominated by a strong linear component (black).
• Shown at right are plots from 3 voxels showing GM for that voxel, TGM slope, TGM ANCOVA fit, and the full model fit.
80
Age
Age
Age
5
x 10
Whole Brain Results: Proportional
• Whole brain analysis performed using DARTEL-normalized data and proportinal scaling, shown at right, shows spatial variation in age-related GM decrease (effect size combined across linear and quadratic components).
decreases slower than TGM decreases faster than TGM
When TGM is entered as a covariate, voxelwise parameter estimates show considerable variation.
4 Frequency 3 2 1 0 0 0.5 1 TGM Beta value x 10−6
Statistical Analyses
• Second order polynomial expansion of age was entered seperately for males and females. • Two models had no covariates: the “no covariate” approach and a “proportional” scaling approach, in which each individual’s GM image was divided by its sum over voxels. Both of these statistical models were identical (top right). • For the TGM ANCOVA model, TGM was added as an additional regressor (bottom right). • Analyses presented here are collapsed across sex. • In addition to whole brain analyses we conducted an ROI analysis using regions from the AAL template. For each ROI we computed declines based on the average value of all voxels within that ROI. This allowed rank ordering of regions by slope of GM change.
average linear decrease with age TGM ANCOVA
Conclusions
• Results from the DARTEL-normalized VBM analysis were consistent with increased registration accuracy. • Treatment of global effects in the model significantly influences magnitude and spatial distribution of results, and must be carefully considered in interpreting the data. • Because linear global effects dominate age-related cortical change, whole brain proportional scaling may be a reasonable approach to identifying regional differences in age-related GM loss.
References
Ashburner J (2007) A fast diffeomorphic image registration algorithm. NeuroImage 38, 95–113. Good CD, Johnsrude IS, Ashburner J, Henson RNA, Friston KJ, Frackowiak RSJ (2001) A voxel-based morphometric study of ageing in 465 normal adult human brains. NeuroImage 14, 21–36. Klein A, Andersson J, Ardekani BA, Ashburner J, Avants B, Chiang M-C, Christensen GE, Collins DL, Hellier P, Song JH, Jenkinson M, Lepage C, Rueckert D, Thompson P, Vercauteren T, Woods RP, Mann JJ, Parsey RV (2009) Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage 46, 786–802.