### 3542

Double diffusion encoding enables unique parameter estimation of the Standard Model in diffusion MRI
Santiago Coelho1,2, Jose M. Pozo1,2, Sune N. Jespersen3,4, Derek K. Jones5,6, and Alejandro F. Frangi1,2

1Centre for Computational Imaging & Simulation Technologies in Biomedicine (CISTIB), School of Computing, University of Leeds, Leeds, United Kingdom, 2Leeds Institute for Cardiac and Metabolic Medicine (LICAMM), School of Medicine, University of Leeds, Leeds, United Kingdom, 3Center of Functionally Integrative Neuroscience (CFIN) and MINDLab, Department of Clinical Medicine, Aarhus University, Aarhus, Denmark, 4Department of Physics and Astronomy, Aarhus University, Aarhus, Denmark, 5Cardiff University Brain Research Imaging Centre (CUBRIC), Cardiff University, Cardiff, United Kingdom, 6School of Psychology, Australian Catholic University, Melbourne, Australia

### Synopsis

The widely adopted Standard Model (SM) for diffusion in white matter tissue has been shown to possess intrinsic degeneracies, making parameter estimation from single diffusion encoding (SDE) data ill-conditioned. We extend the SM to a multidimensional diffusion MRI acquisition and analyse the case where the fibre orientation distribution function (fODF) is a Watson distribution. The information contained in the kurtosis tensor, accessed by SDE, is insufficient to recover biophysical model parameters from the MR signal. We prove that Double Diffusion Encoding (DDE), letting us additionally access the full diffusion tensor covariance, makes parameter estimation well-posed.

### Introduction

Diffusion MRI (dMRI) probes tissue architecture at scales much smaller than typical MRI resolution. Biophysical modelling can extract microstructural information from the dMRI signal at the expense of making assumptions about main tissue features. For brain white matter (WM), the standard model (SM)1 has been widely adopted to interpret the dMRI signal. It comprises two impermeable compartments, resembling intra-axonal and extra-axonal environments. Recent work showed that parameter estimation of the SM from single diffusion encoding (SDE) measurements is ill-conditioned2,3. It has been proposed that complementary information provided by orthogonal measurements, such as echo time dependence4, diffusion time dependence5, or double diffusion encoding (DDE), might help resolve this degeneracy6,7,8. We focus on the case where the fibre orientation distribution function (fODF) is a Watson distribution (Watson SM), extend it to a general b-tensor acquisition9 and analyse its cumulant expansion. We prove that adding complementary information provided by DDE is sufficient to make parameter estimation well-posed.

### Theory

According to the SM, the dMRI signal is the convolution over the unit sphere $\mathbb{S}^2$

$$S(\boldsymbol{B})=S_0~\int_{\mathbb{S}^2}~\mathcal{P}(\hat{\mathbf{u}})\mathcal{K}(\boldsymbol{B},\hat{\mathbf{u}})\,dS_{\hat{\mathbf{u}}},$$

where $S_0=S(\boldsymbol{B}\equiv0)$ is the non-weighted diffusion signal, $\mathcal{P}(\hat{\mathbf{u}})$ the fODF (Watson distribution here), and

$$\mathcal{K}(\boldsymbol{B},\hat{\mathbf{u}})=f\exp\bigl[-D_{\text{a}}B_{ij}u_iu_j\bigr]+(1-f)\exp\bigl[-bD_\text{e}^\perp-\Delta_\text{e}B_{ij} u_i u_j\bigr],$$

is the response signal (kernel) from an individual fibre segment oriented along direction $\hat{\mathbf{u}}$, acquired with the b-tensor $\boldsymbol{B}$9. $f$ is the $T_2$-weighted intra-axonal volume fraction, $D_{\text{a}}$ the axial intra-axonal, $D_\text{e}^\parallel$ the axial extra-axonal and $D_\text{e}^\perp$ the radial extra-axonal diffusivity ($\Delta_\text{e}=D_\text{e}^\parallel-D_\text{e}^\perp$).

In SDE, $\boldsymbol{B}=b\,\hat{\mathbf{n}}\otimes\hat{\mathbf{n}}$, thus, $\boldsymbol{B}$ has one non-zero eigenvalue, viz. linear tensor encoding (LTE)9. In DDE, $\boldsymbol{B}=b_1\,\hat{\mathbf{n}}_{1}\otimes\hat{\mathbf{n}}_{1}+b_2\,\hat{\mathbf{n}}_{2}\otimes\hat{\mathbf{n}}_{2}$, with gradient directions $\hat{\mathbf{n}}_1$, $\hat{\mathbf{n}}_2$, and b-values, $b_1$, $b_2$ (generally equal). If $\hat{\mathbf{n}}_1=\hat{\mathbf{n}}_2$ (parallel direction pair) then $\boldsymbol{B}_{\text{DDE}}=\boldsymbol{B}_{\text{SDE}}$. However, if $\hat{\mathbf{n}}_1\perp\hat{\mathbf{n}}_2$ (perpendicular direction pair) and $b_1=b_2$, $\boldsymbol{B}_{\text{DDE}}$ has two equal non-zero eigenvalues and equals planar tensor encoding (PTE).

### Methods

The fourth order cumulant expansion of the SM dMRI signal for a general b-tensor has the form

$$\log(S)=\log(S_0)-B_{ij}D_{ij}+\frac12B_{ij}B_{k\ell}Z_{ijk\ell}+{\cal~O}(\boldsymbol{B}^3),$$

where $\boldsymbol{D}$ is the diffusion tensor and $\boldsymbol{Z}$, for multicomponent Gaussian compartments (MGC) as SM, is the diffusion tensor covariance10. In our case,

$$D_{ij}=\bigl\langle~D_{ij}\bigr\rangle=\sum_\alpha~f_\alpha~D_{ij}^{(\alpha)},\quad~Z_{ijk\ell}=\Bigl\langle\bigl(D_{ij}-\langle~D_{ij}\rangle\bigr)\bigl(D_{k\ell}-\langle~D_{k\ell}\rangle\bigr)\Bigr\rangle=\sum_\alpha~f_\alpha~D_{ij}^{(\alpha)}D_{k\ell}^{(\alpha)}-D_{ij}D_{k\ell},$$

where $f_\alpha$ and $D_{ij}^{(\alpha)}$ denote the fraction and diffusion tensor of compartment $\alpha$. LTE accesses $\boldsymbol{D}$ and the totally symmetric part of $\boldsymbol{Z}$, identical to $\boldsymbol{W}$, the kurtosis tensor in MGCs: $\bar{D}^2 W_{ijk\ell}=3Z_{(ijk\ell)}$. Jespersen et al. showed $\boldsymbol{D}$ and $\boldsymbol{W}$ are insufficient to compute the kernel parameters due to an intrinsic degeneracy of the model5. The purpose of this work is to assess whether adding the remaining information in $\boldsymbol{Z}$ resolves this degeneracy.

Owing to the axial symmetry of the Watson distribution, $\boldsymbol{D}$ has 2 independent components, $\boldsymbol{W}$ has 3 and $\boldsymbol{Z}$ has 5. We can write them as

$$\boldsymbol{D}=D^\perp\delta_{ij}+\Delta\mu_i\mu_j,\qquad\boldsymbol{W}=\omega_1\boldsymbol{P}+\omega_2\boldsymbol{Q}+\omega_3\boldsymbol{I}\qquad\text{and}\qquad\boldsymbol{Z}=\frac13\bar{D}^2\boldsymbol{W}+\zeta_1\boldsymbol{R}+\zeta_2\boldsymbol{J},$$

with

$$P_{ijk\ell}=\mu_i\mu_j\mu_k\mu_\ell,\\Q_{ijk\ell}=\mu_{(ij}\delta_{k\ell)},\\I_{ijk\ell}=\delta_{(ij}\delta_{k\ell)},\\R_{ijk\ell}=\tfrac12\bigl(\mu_i\mu_j\delta_{k\ell}+\mu_k\mu_\ell\delta_{ij}\bigr)-\tfrac14\bigl(\mu_i\mu_k\delta_{j\ell}+\mu_j\mu_k\delta_{i\ell}+\mu_i\mu_\ell\delta_{jk}+\mu_j\mu_\ell\delta_{ik}\bigr),\\J_{ijk\ell}=\delta_{ij}\delta_{k\ell}-\tfrac12\bigl(\delta_{ik}\delta_{j\ell}+\delta_{i\ell}\delta_{jk}\bigr),$$

where $\delta_{ij}$ is the Kronecker delta and $\hat{\pmb{\mu}}$ the Watson distribution main direction. $D^\parallel$ and $D^\perp$ are the axial and radial diffusivities ($\bar{D} = \frac13 D^\parallel+\frac23 D^\perp$, $\Delta=D^\parallel - D^\perp$), and $\omega_1$, $\omega_2$ and $\omega_3$ are functions of axial, radial, and mean kurtosis ($W_\parallel$, $W_\perp$, $\bar{W}$)11. Information from the non-symmetric part of $\boldsymbol{Z}$ is parameterised by $\zeta_1$ and $\zeta_2$.

Three datasets considering all combinations of $f=[0.1,0.3,0.5,0.7,0.9]$, $D_\text{a}=[0.3,0.8,1.3,1.8,2.3]\mu\!\,m^2/ms$, $D_{\text{e}}^{\parallel}=[0.8,1.3,1.8]\mu\!\,m^2/ms$, $D_{\text{e}}^{\perp}=[0.5,1,1.5]\mu\!\,m^2/ms$, and $\kappa=[0.84,2.58,4.75,9.27,15.53,33.70]$ were simulated. Acquisition protocols (LTE, LTE+PTE, PTE) had equal number of measurements (60) and b-values ($1-2 ms/\mu\!\,m^2$). Rician noise was added ($\text{SNR}=\text{50}$) and nonlinear least squares fitting of Eq. 1 was performed to the data.

### Results

Expressing $\zeta_1$ and $\zeta_2$ as a function of the model parameters gives two new equations independent to those in $\boldsymbol{D}$ and $\boldsymbol{W}$:

$$\tfrac34\zeta_1=(1-f)D_\text{e}^{\perp}\Delta_{\text{e}}\,p_2-D^\perp\Delta,\qquad\text{and}\qquad\tfrac32\zeta_2=(1-f)\Bigl[\tfrac23D_\text{e}^{\perp}\Delta_{\text{e}}(1-p_2)+{D_\text{e}^{\perp}}^2\Bigr]-{D^\perp}^2,$$

where $p_2=\frac{1}{4}\bigl[\frac{3}{\sqrt{\kappa}F(\sqrt{\kappa})}-2-\frac{3}{\kappa}\bigr]$ and $F$ denotes Dawson's function12. Combining these equations with those of $\boldsymbol{D}$ and $\boldsymbol{W}$5, and solving for $p_2$ (if $f\,D_\text{a}^2+(1-f)\,\Delta_\text{e}^2\neq0$) gives a unique solution

$$p_2=\frac{\frac13(W_\parallel-W_\perp)\bar{D}^2-\frac32\zeta_1+\Delta^2}{\frac13(W_\parallel+2W_\perp)\bar{D}^2-\frac32(\zeta_1+3\zeta_2)+\Delta^2}.$$

From this, we uniquely recover the kernel parameters3 (if $p_2\neq0$). See 13 for a similar result. Figure 1 shows LTE+PTE has the lowest root mean squared error in the nonlinear estimation of all model parameters.

Combining LTE and PTE provides full access to $\boldsymbol{D}$ and $\boldsymbol{Z}$. This enables unique mapping between the dMRI signal sensitive up to $\mathcal{O}(b^2)$ and the model parameters. In contrast, LTE or PTE data alone are insufficient to get full access to $\boldsymbol{Z}$.

### Discussion and Conclusion

Combining DDE parallel (LTE) and perpendicular (PTE) direction pairs provides access to $\boldsymbol{D}$ and $\boldsymbol{Z}$. This is sufficient to make the parameter estimation of the Watson SM well-posed without high diffusion weighting. Combining LTE and PTE data is one of the possible ways to access all independent components in $\boldsymbol{D}$ and $\boldsymbol{Z}$. The impact of noise on their estimation needs further analysis. We are currently exploring the selection of the optimal set of b-tensors for estimating $\boldsymbol{D}$ and $\boldsymbol{Z}$ from noisy data.

### Acknowledgements

This work has been supported by the OCEAN project (EP/M006328/1) and MedIAN Network (EP/N026993/1) both funded by the Engineering and Physical Sciences Research Council (EPSRC) and the European Commission FP7 Project VPH-DARE@IT (FP7-ICT-2011-9-601055). DKJ is supported by a Wellcome Trust Investigator Award (096646/Z/11/Z) and a Wellcome Trust Strategic Award (104943/Z/14/Z).

### References

1. Novikov D S, Jespersen S N, Kiselev V G, Fieremans E. Quantifying brain microstructure with diffusion MRI: Theory and parameter estimation. NMR Biomed. 2018; page e3998.

2. Jelescu I O, Veraart J, Fieremans E, Novikov D S. Degeneracy in model parameter estimation for multi-compartmental diffusion in neuronal tissue. NMR Biomed. 2016; 29:33-47.

3. Novikov D S, Veraart J, Jelescu I O, Fieremans E. Rotationally-invariant mapping of scalar and orientational metrics of neuronal microstructure with diffusion MRI. NeuroImage. 2018; 174:518-538.

4. Veraart J, Novikov D S, Fieremans E. TE dependent Diffusion Imaging (TEdDI) distinguishes between compartmental T2 relaxation times. NeuroImage. 2017; 182:360-369.

5. Jespersen S N, Olesen J L, Hansen B, Shemesh N. Diffusion time dependence of microstructural parameters in fixed spinal cord. NeuroImage. 2017; 182:329-342.

6. Coelho S, Beltrachini L, Pozo J M, Frangi A F. Double Diffusion Encoding vs Single Diffusion Encoding in parameter estimation of biophysical models in Diffusion-Weighted MRI. In Proc. Annual Meeting of ISMRM, 2017, page 3383.

7. Fieremans E, Veraart J, Ades-Aron B, Szczepankiewicz F, Nilsson M, Novikov D S. Effects of combining linear with spherical tensor encoding on estimating brain microstructural parameters. In Proc. Annual Meeting of ISMRM, 2018, page 254.

8. Dhital B, Reisert M, Kellner E, Kiselev V G. Diffusion weighting with linear and planar encoding solves degeneracy in parameter estimation In Proc. Annual Meeting of ISMRM, 2018, page 5239.

9. Westin C-F, Knutsson H, Pasternak O, Szczepankiewicz F, Özarslan E,Van Westen D, Mattisson C, Bogren M, O'Donnell L J, Kubicki M, Topgaard D, Nilsson M. q-space trajectory imaging for multidimensional diffusion MRI of the human brain. NeuroImage. 2016; 135:345-362.

10. Jespersen S N. Equivalence of double and single wave vector diffusion contrast at low diffusion weighting. NMR Biomed. 2011; 25:813-818.

11. Hansen B, Khan A R, Shemesh N, Lund T E, Sangill R, Eskildsen S F, Østergaard L, Jespersen S N. White matter biomarkers from fast protocols using axially symmetric diffusion kurtosis imaging. NMR Biomed. 2017; 30:1-17.

12. Abramowitz M, Stegun I. Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables. National Bureau of Standards Applied mathematics series 55, 1972.

13. Reisert M, Kiselev V G, Dhital B. A unique analytical solution of the white matter standard model using linear and planar encodings. Preprint arXiv. 2018; vol. 1808.04389v1.

14. Hintze J L, Nelson R D. Violin plots: A box plot-density trace synergism. The American Statistician. 1998; 52(2):181-184.

15. Kay S M. Fundamentals of statistical signal processing: estimation theory. Englewood Cliffs, N.J: PTR Prentice-Hall, 1993.

### Figures

Violin plots14 of root mean square error (RMSE)15 for all model parameters for all simulated sets. Black dots and red lines denote mean and median RMSE. RMSE corresponding to each model parameter set was computed by repeating the estimation on 50 independent noise realisations of the measurements for each voxel. For clarity, we show the RMSE in the mean-squared-cosine angle instead of $\kappa$, $\langle\cos^2\varphi\rangle=c_2=\langle(\hat{\mathbf{u}}\cdot\hat{\pmb{\mu}})^2\rangle=\bigl(2\sqrt{\kappa}F(\sqrt{\kappa})\bigr)^{-1}-(2\kappa)^{-1}$.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3542