Longitudinal stability of MRI for mapping brain change using tensor-based morphometry

Abstract

Measures of brain change can be computed from sequential MRI scans, providing valuable information on disease progression, e.g., for patient monitoring and drug trials. Tensor-based morphometry (TBM) creates maps of these brain changes, visualizing the 3D profile and rates of tissue growth or atrophy, but its sensitivity depends on the contrast and geometric stability of the images. As part of the Alzheimer's Disease Neuroimaging Initiative (ADNI), 17 normal elderly subjects were scanned twice (at a 2-week interval) with several 3D 1.5 T MRI pulse sequences: high and low flip angle SPGR/FLASH (from which Synthetic T1 images were generated), MP-RAGE, IR-SPGR (N = 10) and MEDIC (N = 7) scans. For each subject and scan type, a 3D deformation map aligned baseline and follow-up scans, computed with a nonlinear, inverse-consistent elastic registration algorithm. Voxelwise statistics, in ICBM stereotaxic space, visualized the profile of mean absolute change and its cross-subject variance; these maps were then compared using permutation testing. Image stability depended on: (1) the pulse sequence; (2) the transmit/receive coil type (birdcage versus phased array); (3) spatial distortion corrections (using MEDIC sequence information); (4) B1-field intensity inhomogeneity correction (using N3). SPGR/FLASH images acquired using a birdcage coil had least overall deviation. N3 correction reduced coil type and pulse sequence differences and improved scan reproducibility, except for Synthetic T1 images (which were intrinsically corrected for B1-inhomogeneity). No strong evidence favored B0 correction. Although SPGR/FLASH images showed least deviation here, pulse sequence selection for the ADNI project was based on multiple additional image analyses, to be reported elsewhere.

Fig. 1

This figure shows the axial, coronal and sagittal views of a subject imaged using four of the five MRI sequences studied in this paper (SPGR, IR-SPGR, MP-RAGE and calculated Synthetic T1).

Fig. 2

Illustration of the approach employed in this paper where a color map of the Jacobian determinant (a measure of expansion or compression of time) is generated by registering follow-up scan (source image) to baseline (target image) and computing the Jacobian of the forward mapping (baseline to follow-up). This map is overlaid on the baseline structural image to visualize voxelwise local tissue change. Red colors denote regions where local expansion is detected, blue colors denote compression, and green denotes regions with no detected change. Notice that the 3D local tissue change, encoded using the Jacobian map, could not be easily visualized by inspecting the subtraction map due to difficulties in reconstructing 3D spatial relationships from 2D slices.

Fig. 3

This figure visualizes the repeatability, defined as the voxelwise variance of the log-transformed Jacobian maps, for different sequence types and transmit/receive coils. Note that SPGR in the figures actually refers to SPGR for GE acquisitions and FLASH for Siemens acquisitions, while BC and PA denote bird cage and phased array designs, respectively.

Fig. 4

This figure visualizes the performance, measured by the voxelwise absolute mean deviation from zero of the log-transformed Jacobian maps, for different sequence types. Based on the assumption of no brain structural difference, a sequence is defined to have better performance if it has a lower absolute mean deviation. Visually, SPGR acquired using a birdcage coil with N3 correction performs the best overall. This is confirmed using permutation tests as shown in Tables 2–5.

Fig. 5

This figure visualizes the comparison of B0-corrected images (MEDIC) vs. GW and B1 intensity-corrected MP-RAGE images with N3 correction (N = 12 scans). A permutation test established no statistically significant differences between the two image types, although an effect may be detectable in a larger sample of scans. As in Fig. 4, the performance, coded here in color, is defined as the voxelwise absolute mean deviation from zero of the log-transformed Jacobian maps.