Diffusion MRI in the cortex of the brain: Reducing partial volume effects from CSF and white matter in the mean diffusivity using high b-values and spherical b-tensor encoding

Abstract

Purpose: The mean diffusivity (MD) is sensitive to the microstructure of the cortex. However, partial volume effects with CSF and white matter (WM) may obscure pathology-related alterations. This work investigates both existing approaches and a novel approach for reducing partial volume effects.

Theory and methods: A bias in MD arises due to partial volume effects, higher-order terms, and the noise floor in magnitude data. We propose to reduce this bias by using high b-value encoding to limit partial volume effects with CSF, spherical b-tensor encoding to reduce the influence of higher-order terms, and super-resolution acquisition and reconstruction to suppress the noise floor. This approach was investigated, along with established approaches (inversion recovery and free water elimination) for reducing partial volume effects, using simulations and in vivo data.

Results: High b-value diffusion MRI with spherical b-tensor encoding reduced partial volume effects with CSF relative to conventional diffusion MRI. Maximum errors decreased from 0.51 to 0.01 μm²/ms in simulations. In vivo, the median absolute deviation of cortical MD decreased from 0.17 to 0.06 μm²/ms, whereas the median decreased slightly from 0.77 to 0.73 μm²/ms. The other methods yielded bias from either CSF, WM, or model assumptions.

Conclusion: The mean diffusivity of the cortex can be mapped in high precision with reduced influence of partial volume effects with CSF and WM matter using high b-values and spherical b-tensor encoding and super-resolution reconstruction.

Keywords: cortex; diffusion MRI; high b‐values; mean diffusivity; partial volume effects; spherical b‐tensor encoding.

FIGURE 1

Quantification errors in MD from PVE and higher‐order signal terms. (A) From left to right: signal vs. b‐curves in pure CSF, WM, and GM. Signals without noise from conventional DWI (LTE), FLAIR (with LTE), and isotropic diffusion encoding (STE). (B) MD values in a GM voxel with different contaminations of CSF (left) and WM (right) at low b‐values (LTE_low, FLAIR_low, and STE_low; 0 and 1 μm²/ms) and high b‐values (LTE_high, FLAIR_high, and STE_high; 1.5 and 2.5 μm²/ms). The unbiased MD_t value in GM (0.75 μm²/ms) is marked by a red square. Altogether, the results suggest that FLAIR_low and STE_high yield the lowest errors. FLAIR, fluid attenuated inversion recovery; FWE, free‐water elimination; GM, graymatter; LTE, linear b‐tensor encoding; MD, mean diffusivity; PVE, partial volume effects; STE, spherical b‐tensor encoding; WM, white matter.

FIGURE 2

Analysis of accuracy and precision of MD in GM. Maximum errors compared to a noncontaminated signal (top row), SD over noise iterations (middle row), and RMSE compared to noncontaminated signal (bottom row) for DWI‐LTE, DWI‐FLAIR, and DWI‐STE. Points corresponding to the b‐value combinations used in vivo are shown in gray, and isolines marking 0.15 μm²/ms (white) and 0.3 μm²/ms (black) are included. Low b‐values yielded large errors for DWI‐LTE and DWI‐STE but not for DWI‐FLAIR. However, SDs were larger for DWI‐FLAIR than for the other techniques. RMSE, root‐mean‐square error.

FIGURE 3

Schematic analysis of effect size. Effect sizes, in terms of Cohen's d for two types of GM abnormalities: global MD elevation, and cortical thinning for LTE_low, FLAIR_low with perfect inversion pulse, FLAIR_low with inversion pulse mismatch, LTE_high, and STE_high. FLAIR_low and STE_high were able to distinguish the two abnormalities well. A mismatch for the inversion pulse reduced this ability. FWE, free‐water elimination.

FIGURE 4

Signal intensities and MD maps from a healthy participant following coregistration to the T₁‐MPRAGE image. First and second row: diffusion weighted images obtained with b₁ (first row) and b₂ (second row). Third and fourth row: transversal (third row) and coronal (fourth row) slices of MD maps obtained with LTE_low, FLAIR_low, LTE_high, FWE_low, and STE_high. Compared to LTE_low, CSF contributions were reduced in all cases.

FIGURE 5

Examples of features of interest in MD in one healthy participant following coregistration to the T₁‐MPRAGE image. (A) Fifteen transverse slices of an MD map obtained with STE_high in a healthy participant (23 years). The patterns identified were: diffuse hyperintensities adjacent to the posterior part of the ventricles (orange arrows), fat artifacts (yellow arrows), and hypointensities in the basal ganglia (blue arrows). (B) Transverse slices covering the optic radiation in MD maps from LTE_low, FLAIR_low, LTE_high, FWE_low, and STE_high, where all approaches except FWE_low yielded hyperintense areas covering these. Image data from one subject is presented, but similar patterns were seen in all subjects studied.

FIGURE 6

Quantitative analysis of MD values in the cortex of healthy participants. Left: histograms showing the distribution of MD values in the cortex for LTE_low (orange), FLAIR_low (yellow), LTE_high (dark blue), FWE_low (green), and STE_high (light blue). The 25th, 50th, and 75th quantiles of the MD values are presented for each technique. Right: bar graph showing the mean MAD of the cortical MD values in the healthy participants. All approaches reduced the number of high MD voxels and MAD of the MD values compared to LTE_low.