Intravoxel incoherent motion (IVIM) imaging¹ has gained renewed interest during recent years, even in the brain, in spite of the comparatively low cerebral blood volume². One reason for the increased focus on the IVIM technique is improvements in the model of choice, in particular the incorporation of relaxation effects in the IVIM model³, and Bayesian analysis approaches⁴. The inclusion of relaxation effects becomes even more important at high magnetic field strengths (3-7T), mainly due to the shortening of venous T2⁵. Still, to the best of our knowledge, direct measurements of T1 and T2 to improve the estimation of IVIM parameters, such as the perfusion fraction F_b, have not previously been performed. To even more ameliorate the IVIM analysis, the images employed in this work are based on high-resolution acquisition, both with respect to b-value composition and voxel size. All these improvements, together with an extension of the conventional two-compartment model into a three-compartment model, are introduced to overcome the low precision of brain IVIM, as well as the well-known problem with CSF contamination in the perfusion fraction estimates⁶.

In a proof-of-concept experiment, using a 3T whole-body MRI-scanner (MAGNETOM Prisma, Siemens Healthcare GmbH, Erlangen, Germany), we measured multiple b, TI and TE data to be included in a relaxation compensated (RC) three-compartment IVIM model:

$$ \begin{eqnarray}S &= &S_{000}\sum_{i=\{t,c,b\}}F_i\rho_i[1-(1-cos\theta_1)e^{-TI/T_{1i}}+(1 - 2e^{TE/2T_{1i}})cos\theta_1 e^{-TR/T_{1i}}]e^{-TE/T_{2i}}e^{-bADC_i} = \nonumber \\ &= &S_{0,t} \{F_t \rho_t [1-(1-cos\theta_1) e^{-TI/T_{1t}}+(1-2e^{TE/2T_{1t}})cos\theta_1 e^{-TR/T_{1t}}]e^{-TE/T_{2t}}e^{-bD_t} \} \nonumber \\ & + &S_{0,c} \{F_c \rho_c [1-(1-cos\theta_1)e^{-TI/T_{1c}}+(1-2e^{TE/2T_{1c}})cos\theta_1 e^{-TR/T_{1c}}] e^{-TE/T_{2c}}e^{-bD_c} \} \nonumber \\ & + &S_{0,tb} \{F_b \rho_b [1-(1-cos\theta_1) e^{-TI/T_{1b}}+(1-2e^{TE/2T_{1b}})cos\theta_1 e^{-TR/T_{1b}}]e^{-TE/T_{2b}} e^{-b(D_b+D^*)} \} \nonumber\end{eqnarray}$$

where S₀₀₀ is the non-weighted (b=0, TE=0, TR=∞) signal value, F is the fractional volume, ρ is the water content, ADC is the apparent diffusion coefficient, θ₁ is the inversion flip angle (assumed to be 180°) and i denotes the respective compartment, i.e., i=[tissue(t),CSF(c),blood(b)]. For comparison, we also analyzed multi-b data with a non-relaxation-compensated (nonRC) three-compartment IVIM model according to:

$$ \begin{eqnarray}S &=&S_{000}\sum_{i=\{t,c,b\}} F_i\rho_ie^{-TE/T_{2i}}e^{-bADC_i} \end{eqnarray}$$

A spin-echo EPI sequence with diffusion encoding in six directions was employed for the IVIM data collection, with 45 b-values ranging between 15 and 800 s/mm². Imaging parameters for full brain coverage were TR=4500 ms, TE=67 ms, FOV 250×250 mm², matrix size 192×192, slice thickness 4 mm, and 32 slices.

Multi-TE data were obtained using 5 different TEs (61, 80, 100, 120, 140 ms) with a TR of 6500 ms. Multi-TI data were collected using 10 different TIs (500, 1000, 1500, 2100, 2500, 2900, ms) with a TR of 10000 ms and a TE of 67 ms. Multi-TE as well as multi-TI data were acquired with b-values of 0 and 200 s/mm².

A simplified Bayesian voxel-by-voxel analysis, using Gaussian priors and posterior distributions, was used to construct maps of the model parameters⁷. We used non-informative priors for the fractions, and literature values for CSF (T1_c = 2000 ms, T2_c = 500 ms and D_c = 3 µm²/ms), as well as for the self-diffusion of blood (D_b = 1.7 µm²/ms). The analysis was also applied to the mean whole-brain signal. Empirically assigned priors were used on the remaining parameters (see Table 1).

Figure 1 displays the fit of the RC model to the whole brain signal. The whole-brain analysis yielded F_b estimates of 2.1% and 2.5% for RC and nonRC, respectively. The corresponding values for F_c were 12.0% (RC) and 24.0% (nonRC), and for F_t 86.0% (RC) and 73.6% (nonRC). Figure 2 shows parametric maps from a representative mid-brain slice: (a) F_b, F_c, F_t, D*, D_t, as well as T1 and T2 for the relaxation compensated data, and (b) for or the non-relaxation-compensated data. The RC analysis improved the separation of the three compartments (seen most clearly in the ventricles of the F_b-map) and produced a less noisy F_b-map, compared to the nonRC results.When relaxation was both measured and compensated for, the CSF was more distinctly allocated to the F_c-map, and contributed less to CSF contaminations of the F_b-map, suggesting that the relaxation compensation was able to better distinguish between the CSF compartment and the blood compartment, compared to the approach in which relaxation data were ignored in the model. Furthermore, F_c better agreed with the literature value of 9.1±2.4%⁸ when the RC model was used, making this model even more convincing. Although results seem promising, further efforts are needed to eliminate the risk of bias caused by the supplementary numbers of data points used in the RC-model. Our extended three-compartment model with relaxation-compensated data is a promising tool to overcome CSF contamination of the blood-pool data in IVIM imaging.