Development of a protocol to assess within-subject, regional white matter hyperintensity changes in aging and dementia

Abstract

Background: White matter hyperintensities (WMH), associated with both dementia risk and progression, can individually progress, remain stable, or even regress influencing cognitive decline related to specific cerebrovascular-risks. This study details the development and validation of a registration protocol to assess regional, within-subject, longitudinal WMH changes (ΔWMH) that is currently lacking in the field.

New method: 3D-FLAIR images (baseline and one-year-visit) were used for protocol development and validation. The method was validated by assessing the correlation between forward and reverse longitudinal registration, and between summated regional progression-regression volumes and Global ΔWMH. The clinical relevance of growth-regression ΔWMH were explored in relation to an executive function test.

Results: MRI scans for 79 participants (73.5 ± 8.8 years) were used in this study. Global ΔWMH vs. summated regional progression-regression volumes were highly associated (r2 = 0.90; p-value < 0.001). Bi-directional registration validated the registration method (r2 = 0.999; p-value < 0.001). Growth and regression, but not overall ΔWMH, were associated with one-year declines in performance on Trial-Making-Test-B.

Comparison with existing method(s): This method presents a unique registration protocol for maximum tissue alignment, demonstrating three distinct patterns of longitudinal within-subject ΔWMH (stable, growth and regression).

Conclusions: These data detail the development and validation of a registration protocol for use in assessing within-subject, voxel-level alterations in WMH volume. The methods developed for registration and intensity correction of longitudinal within-subject FLAIR images allow regional and within-lesion characterization of longitudinal ΔWMH. Assessing the impact of associated cerebrovascular-risks and longitudinal clinical changes in relation to dynamic regional ΔWMH is needed in future studies.

Keywords: Aging; Cerebrovascular disease; Dementia; Longitudinal; Small vessel ischemic disease; White matter hyperintensity.

Fig. 1:

Sample of the excluded T2-weighted fluid-attenuated inversion recovery (FLAIR) images. A) A FLAIR image with a motion artifact leads to undefined ventricles and white matter hyperintensities edges (see the circle). Also, the motion artifact could lead to a variation in the intensity that may appear as a WMH and may cause an overestimation volume (see the rectangle). (B) A FLAIR image shows an irregular brain shape and ventricles that can cause a segmentation error. Note: The motion artifact in panel A can be compared to panel B, and the asymmetry in panel B can be compared to panel A.

Fig. 2:

Pre-longitudinal registration step. T2-BL, -Y1 (T2-weighted image, FLAIR, baseline, year one visit). T1-BL, -Y1 (T1-weighted, MPRAGE image, baseline, year one visit). This step divided into two parts: First, intensity correction (N3) for all the images and co-registering the T1 and T2 for each time point separately. Second, registering each one-year visit imaging sequence (T1 or T2) to the baseline.

Fig. 3:

The main output steps of the penumbra protocol. A and B are the FLAIR images of the twotime-points before the longitudinal step. C is the midpoint image of the two FLAIR images of the longitudinal step. D and E are the longitudinal registered images of the two-time-points after stripping the nonbrain tissue. F is the white matter mask after segmentation step (normal appearing WM and WMH). G and H are the WMH masks of the two-time-points. I is the regional pattern penumbra (white) and regression (black) masks after subtracting WMH masks at time-2 and time-1 from each other.

Fig. 4:

The z-score of the FLAIR images from baseline and year 1 scans. A and B are the normalized FLAIR images of a two-time-points before longitudinal registration. C is the z-score difference of the two-time-points before registration. D and E are the normalized FLAIR images of a two-time-points after longitudinal registration and F is z-score difference image. Red arrows show the result of subtraction of the two normalized images before longitudinal registration due to the poor alignment.

Fig. 5:

Longitudinal registration step. T2-BL, -Y1 (T2-weighted image, FLAIR, baseline, one-yearvisit). T1-BL, -Y1 (T1-weighted image, baseline, one-year-visit). DF, deformation field. LR, longitudinal registration. The T2-weighted FLAIR images (T2-BL and rT2-Y1) from the preregistration step registered longitudinally to generate the midpoint image (T2-Midpoint) and the deformation field of each time point (DF-Y1 and DF-BL). Applying the deformation filed of each time point to register all images to the midpoint image, to get the longitudinal registered images (LR-T2-BL, LR-T1-BL, LR-T2-Y1, and LR-T2-Y1).

Fig. 6:

WMH masks from a single representative subject overlaid on their FLAIR image. Red: WMH static, Blue: WMH regression and, Green: WMH growth.

Fig. 7:

An example of year 1 WMH voxels overlaid on a baseline FLAIR image. A: Baseline. B: One-year. The narrow arrows show the edge of the WMH bordering the ventricles. A: Baseline. B: One-year. The narrow arrows show the edge of the WMH bordering the ventricles. The full yellow arrows indicate the regression of the WMH at one-year visit (changes inside the WMH cluster). The arrowheads show the changes of the WMH toward the deep WM. The rectangles show stable WMH. The circle shows the appearance of a new WMH (Penumbra). The ellipsoids show the penumbra inside the WMH cluster.

Fig. 8:

Longitudinal WMH for 79 participants. A) Participants are sorted greatest to least (left to right) by total WMH at first scan (i.e., WMH1 / TIV1). The top row of panels shows the growth, stable, and regression segmentations of WMH, as a proportion of mean TIV. The bottom row of panels shows the total WMH change as a proportion of mean TIV. B) Three different cases with WMH2-WMH1 compared to the longitudinal WMH results that demonstrate both growth and regression in WMH whereas the WMH difference shows only a single volume.