Intravoxel incoherent motion (IVIM) is a proposed method of measuring blood volume and perfusion¹. It relies only on the motion of water within a voxel, and thus should be capable of measurements throughout the brain, including white matter. In this study high-resolution IVIM data were acquired to isolate and compare signals arising from grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF).

6 subjects were scanned on a 3T Siemens Prisma with a 32-channel head coil and RESOLVE: a readout-segmented EPI diffusion sequence (rs-EPI) with 1mm isotropic resolution. To achieve sufficient SNR the 6:14-minute protocol was repeated 6 times. Due to potential motion, images were not averaged. The rs-EPI sequence was implemented with TE=69ms, TR=1.6s, BW=766 Hz/pixel, echo-spacing=0.4ms, GRAPPA=2. b-values were 0, 10, 20, 40, 80, 110, 140, 170, 200, 300, 400, 500, 600, 700, 800 and 900s/m², and were acquired with Stejskal-Tanner encoding. Images were corrected for motion and eddy-current distortions. An MPRAGE was acquired at the same resolution, and segmented using FAST² to estimate the fractions of GM, WM and CSF per voxel. Masks were created from voxels with 100% of each tissue type.

The diffusion coefficient $$$D$$$, the pseudo-diffusion coefficient $$$D^*$$$ and the fraction of fast-diffusing spins $$$f_v$$$ , were estimated using a biexponential model, which was fitted to each voxel in two steps to mitigate overfitting.³ This assumes that $$$D^*$$$ is substantially larger than $$$D$$$ and may be neglected for large b-values. First, a monoexponential model was fit to b-values >200 to estimate $$$D$$$:

$$S=S_0e^{-bD}$$

Then the biexponential model was fit to all b-values, with fixed $$$D$$$, to estimate $$$f_v$$$ and $$$D^∗$$$:

$$S=S_0\{(1-f_v)e^{-bD}+f_ve^{-bD^*}\}$$

$$$S_0$$$ was unconstrained in both cases and IVIM perfusion was defined as $$$f_vD^*$$$.

Voxelwise fitting was performed with no smoothing or averaging, and the IVIM-derived parameter maps were compared to structural images. Correlations between blood volume fraction $$$f_v$$$ and GM, WM and CSF fraction of partially-volumed voxels were investigated by plotting 2D histograms. Box-plots were created to compare the distributions of $$$f_v$$$ and perfusion in pure GM, WM and CSF.

Fig.1 shows maps of IVIM parameters from a single subject, alongside a structural image and the CSF partial-volume estimate map derived from it. Areas with the highest $$$f_v$$$ correspond to those with the highest CSF fraction, and no obvious contrast between GM and WM is seen in IVIM maps. Fig.2 shows 2D histograms exploring the relationship between GM, WM and CSF partial volume estimates (pve) and $$$f_v$$$. No correlation is seen for GM or WM, but a positive trend is observed for CSF.

Fig.3 compares the distribution of voxelwise $$$f_v$$$ and $$$f_vD^∗$$$ for a single subject, derived only from voxels with no partial voluming. These show considerably higher $$$f_v$$$ and $$$f_vD^∗$$$ in CSF compared to GM or WM. Extreme outliers were observed for all tissue types and were excluded from the plots. $$$f_vD^∗$$$ was significantly different in CSF compared to WM (p<0.001 in all subjects) and GM (p<0.05 in all subjects, and p<0.001 in 4). $$$f_vD^∗$$$ in GM and WM differed significantly in only 3 subjects (p<0.05), but in all cases the median values were higher in WM than GM.

$$$f_v$$$ and $$$f_vD^∗$$$ are elevated in voxels containing CSF, but no contrast is visible between GM and WM. There appears to be some consistent correlation between $$$f_v$$$ and CSF fraction, but not with GM or WM fraction. This strongly suggests that $$$f_v$$$ is primarily sensitive to voxelwise CSF content and not true capillary blood volume. The box-plots confirm that $$$f_v$$$ and $$$f_vD_∗$$$ are highest in the CSF compartment, and also show that the median $$$f_vD^∗$$$ value is marginally higher in WM than GM. The same observations were made for all subjects. This result is unphysiological for a perfusion contrast.

Overfitting may explain the high WM results. Data were analysed using a biexponential model for all voxels, when the monoexponential fit may suffice. However, by fitting a biexponential to a noisy monoexponential curve, large values for $$$f_v$$$ can be estimated despite a true $$$f_v=0$$$. The anisotropy in WM leads to greater variance in signal across diffusion directions compared to GM, which can masquerade as lower SNR and lead to poorer fitting.

The idea that the biexponential form of the diffusion signal arises from CSF contamination is not new.⁴ However, the fact that biexponential behaviour was observed even in 100% CSF voxels, with resulting $$$f_v$$$ of ~0.4, suggests that IVIM may in fact be measuring CSF motion. In larger voxels, the mixing of different tissue types is likely an additional contributing factor.