Multivariate and univariate analysis of continuous arterial spin labeling perfusion MRI in Alzheimer's disease

Abstract

Continuous arterial spin labeling (CASL) magnetic resonance imaging (MRI) was combined with multivariate analysis for detection of an Alzheimer's disease (AD)-related cerebral blood flow (CBF) covariance pattern. Whole-brain resting CBF maps were obtained using spin echo, echo planar imaging (SE-EPI) CASL in patients with mild AD (n=12, age=70.7+/-8.7 years, 7 males, modified Mini-Mental State Examination (mMMS)=38.7/57+/-11.1) and age-matched healthy controls (HC) (n=20; age=72.1+/-6.5 years, 8 males). A covariance pattern for which the mean expression was significantly higher (P<0.0005) in AD than in HC was identified containing posterior cingulate, superior temporal, parahippocampal, and fusiform gyri, as well as thalamus, insula, and hippocampus. The results from this analysis were supplemented with those from the more standard, region of interest (ROI) and voxelwise, univariate techniques. All ROIs (17/hemisphere) showed significant decrease in CBF in AD (P<0.001 for all ROIs, alphacorrected=0.05). The area under the ROC curve for discriminating AD versus HC was 0.97 and 0.94 for covariance pattern and gray matter ROI, respectively. Fewer areas of depressed CBF in AD were detected using voxelwise analysis (corrected, P<0.05). These areas were superior temporal, cingulate, middle temporal, fusiform gyri, as well as inferior parietal lobule and precuneus. When tested on extensive split-half analysis to map out the replicability of both multivariate and univariate approaches, the expression of the pattern from multivariate analysis was superior to that of the univariate.

Fig. 1

Group average CASL CBF images for HC and AD are shown in the 1^st and 2^nd row, respectively, at three axial locations from lower, middle, and upper brain. Corresponding difference images (HC-AD) are shown in the 3^rd row. Units in the color bar are in [mL/100g•min].

Fig. 2

A) areas with robust negative loadings (i.e., concomitant decreased flow in AD relative to HC) as ascertained by the bootstrap procedure. Most areas are located around the Parahippocampal Gyrus in the Medial Temporal and Occipital Lobes; Thalamus was identified by our analysis as well. No areas of positive loadings (i.e., concomitant increased flow for AD subjects) were found. B) Subject expression of the discriminant pattern for both AD and HC groups. One can appreciate the very small overlap between the two groups.

Fig. 3

Plot of CBF values [mL/100g*min] vs. ROI for AD patients (red, n = 12) and HC (blue, n = 20). Left (solid) and right (striped) hemispheres are shown separately. Error bars represent ± 1 SE across subjects in each group.

Fig. 4

Voxelwise SPM{T}map for the (HC - AD) contrast [corrected, p < 0.05] overlaid on: A) surface rendering of the brain, B) 3D sections of the SPGR of one of the HC subjects.

Fig. 5

Ratios of t-statistic (left column) and AUC (right column), plotted for both univariate (top row) and multivariate approaches (bottom row). Ratios were defined as the value obtained for the replication sample divided by the value obtained for the derivation sample. Ratios approaching or surpassing unity indicate good replicability.