Introduction

High-gradient diffusion MRI has demonstrated promise for mapping tissue microstructural properties¹ in a variety of neurological disorders, including multiple sclerosis^2,3 and stroke. Many factors can affect the reliability of these metrics, including motion artifact, gradient nonlinearity, susceptibility, eddy current distortion, and/or patient positioning. Previously, robustness of AxCaliber diffusion measurements has been tested through scan-rescan experiments in the spinal cord⁴. Recently, we devised a spherical mean technique (SMT) for estimating axon diameter index and restricted volume fraction in white matter independent of the underlying fiber configuration⁵. In this work, we sought to assess the scan-rescan reliability of quantitative MR metrics provided by SMT-based microstructural imaging.

Methods

Data acquisition
In total, seven healthy, cognitively normal adults (ages 22-72, 4F) participated in this study. Participants were scanned twice in the same day, with a rest period in which they were removed from the scanner for approximately one hour between scans. Imaging data were acquired on the 3T Connectome scanner with a maximum gradient strength of 300mT/m^6,7 using a custom-built 64 channel head coil⁸. Sagittal 2-mm isotropic resolution diffusion-weight spin-echo EPI images were acquired with whole-brain coverage. Other parameters include: TR/TE=4000/77ms, d=8ms, D=19/49ms, 8 diffusion gradient strengths linearly spaced from 30-290mT/m per D, 32-64 diffusion directions, parallel imaging (R=2) and slice acceleration factor of 2.

Image processing
Diffusion data were corrected for susceptibility and eddy current distortions using the EDDY and TOPUP tools in FSL. To achieve voxel-wise correspondence between scan/rescan images and preserve the original 2 mm isotropic voxel size, the average of interleaved b=0 images was registered to the T1w image within each scan session using boundary-based registration (BBR). The T1w images of the two sessions were also registered using BBR. These transformations were then concatenated to obtain a single transformation between the diffusion space of scan/rescan sessions. To ensure that the process of registration exerts identical influences on the two scan sessions, a halfway transformation was calculated using the concatenated transformation obtained in the last step, which was finally applied to the calculated maps in their native space to bring them to the mid-way point between scan/rescan. FreeSurfer reconstruction (version 6.0) was performed to obtain white matter masks in each subject’s native space, which were transformed to the mid-way point between scan/rescan, and voxels within the white matter mask of both scan/rescan sessions were included to report voxel-wise repeatability measures. The JHU probabilistic atlas of white matter tracts⁹ was used to generate ROIs to report tract-wise repeatability measures, where NiftyReg was used to perform the nonlinear registration between native diffusion space and the template space

Data analysis
Voxel-wise estimates of axon diameter index, restricted water fraction, hindered water fraction and hindered diffusivity were fitted for using Markov Chain Monte Carlo sampling as described before⁵. In addition, DTI fitting was performed using the b = 950 s/mm2 data. Voxel-wise correlation coefficient and absolute errors were calculated to measure the voxel-wise repeatability. On the tract level, intraclass correlation coefficients (ICC) were calculated to assess the capability of each metric to reveal the between-subject variation given the intra-subject variability. Specifically,

ICC= (σ_inter² )/(σ_inter² + σ_intra²)

Where σ_inter² = variance_i(m_i,avg) and σ_intra² = mean(1/2((m_i,scan-m_i,avg)² + (m_i,rescan-m_i,avg)²)), m_i,scan is the metric value within a particular tract for a subject i. m_i,avg is the average value of said metric between scan and rescan sessions¹⁰. According to this expression, an ICC close to 1 reveals a much higher between-subject variation than intra-subject scan-rescan error.

Results

Figure 1 shows representative maps of scan and rescan sessions in a healthy subject that demonstrates the visual consistency in diffusion metrics between scan/rescan sessions. Assessment of registration between scan/rescan sessions yielded a high cross-correlation coefficient for all subjects (Fig.2). Voxel-wise estimates yielded a high Pearson’s correlation coefficient and low absolute deviation of error across all subjects (Fig.3). Figure 4 provides mean correlation coefficients and absolute errors across subjects; the correlation coefficient was calculated within white matter voxels per subject, and the average was taken across all subjects for the represented diffusion metrics (same calculation applied for absolute error). Intraclass correlation coefficient was also calculated for selected tracts across all subjects. ICC values for restricted volume fraction were high across the majority of represented tracts. ICC values for axon diameter are higher for well-defined fiber bundles, such as the corticospinal tract (Fig.5).

Discussion and Conclusion

We demonstrate scan-rescan reliability across microstructural metrics derived from high-gradient diffusion MRI, including spherical mean estimates of axon diameter index and restricted volume fraction. The correlation coefficients for FA and restricted volume fraction were high and somewhat lower for the axon diameter index, which appeared more susceptible to image noise and artifacts (e.g., eddy current distortions using higher b-value data). Nevertheless, the correlation coefficients and absolute errors were comparable for those reported for scan-rescan repeatability of AxCaliber metrics in the spinal cord obtained with high-gradient diffusion MR and suggest these metrics can be used to glean relative trends in microstructural measures between individuals. Overall, our preliminary findings support the robustness and reliability of high-gradient diffusion MR metrics for use in clinical research studies. Future work will focus on assessing the scan-rescan reliability of these metrics in more subjects and patients.

Acknowledgements

This work was supported by NIH K23NS096056 and U01EB026996.