Diffusion MRI: Disentangling Micro- from Mesostructure and Bayesian Parameter Evaluation
Marco Reisert1, Elias Kellner1, Bibek Dhital1, J├╝rgen Hennig1, and Valerij G. Kiselev1

1Department of Radiology, Medical Physics, University Medical Center Freiburg, Freiburg, Germany


Diffusion-sensitized MRI probes the cellular structure of the human brain, but the primary microstructural information gets lost in averaging over higher-level, mesoscopic tissue organization such as different orientations of neuronal fibers. While such averaging is inevitable due to the limited imaging resolution, we propose a method for disentangling the microscopic cell properties from the effects of mesoscopic structure. The proposed method finds detectable parameters of a given microstructural model and calculates them within seconds, which makes it suitable for a broad range of applications.


Assessing the tissue microstructure (in particular of brain white matter) using diffusion MRI (dMRI) is a great challenge. It requires solution of an ill-posed inverse problem. It is known that different models map well on the same data [1] and eventually the fitting quality cannot validate the underlying model even in the simplest case [2]. This deep problem is further exacerbated by the averaging of the cellular signal on the mesoscopic subvoxel level due to multiple orientations of neuronal fibers as described by the orientation distribution function (ODF), which makes the dMRI signal less specific to the microstructure.

The inverse problem is commonly (with a few exceptions, e.g. [3,4]) treated as finding the microstructural parameters $$$x$$$, that have the maximal a-posteriori probability (MAP) given the signal or its derived `features', $$$f$$$, aka fitting. Doing so, one has to introduce numerous constraints or priors, one has to account for the real noise statistics, and, last but not least, nonlinear fitting is very slow (lasting from a few hours to days). Here we propose a simultaneous solution to the above problems on the way to finding parameters of a given microstructural model. Not all of them are detectable, recognition of those is a constituent part of the proposed method.


We use derived features of the signal that are exclusively sensitive to the microstructure being invariant with respect to the ODF. Using the decomposition in spherical harmonics, $$$Y_{l,m}$$$, the signal takes the form of a product $$ S^b_{l,m}=M^b_l\,p_{l,m}\quad,$$where $$$M$$$ is the model, which is assumed to be axially symmetric, $$$p$$$ is the ODF and $$$b$$$ represents the dependence on a radial variable in q-space (e.g. the b-shells). A known ODF-invariant feature is given by $$$l=0$$$ [5,6]: $$ f_0^b:=S^b_{0,0}=M_0^b\,.$$We propose a larger set of ODF-independent features which are ratios of $$$l$$$-wise squared norms for $$$l>0$$$: $$ f^b_l := \frac{\sum_m |S^b_{l,m}|^2}{\sum_{m,b'} |S^{b'}_{l,m}|^2} = \frac{|M_l^b|^2}{\sum_{b'} |M_l^{b'}|^2} \label{eq:features}$$This infinite set is practically limited by the noise (Fig.1).

Instead of the commonly used fitting paradigm, we engage a Bayesian estimator to find the expectation values of the unknown parameters $$$\widetilde{x}_B(f)$$$. The fitting of the Bayesian estimator is closely related to machine learning. Therein, the full signal formation process is simulated for many parameter constellation of a specific tissue model (Fig.2) including the ODF and measurement noise. If the mapping $$$\widetilde x_B (f)$$$ is applied to the simulated data the fidelity of the parameter determination can be easily monitored (Fig.3).

We applied the method to data from a healthy subject from the HCP data set [7] and to two healthy subjects measured in-house using a two-shell scheme (2-shell60) and the most parsimonious sampling of q-space with only 28 points with $$$b>0$$$ (hex28).


A training of the Baesian estimator is performed on a regular desktop computer within minutes. The resulting estimator $$$\widetilde{x}_B(f)$$$ is applied to all data obtained with a given measurement scheme. The computation time for one brain is in the range of seconds. The number of acquired $$$b$$$-shells determines the number of features. We found the restriction $$$l\leq 2$$$ (Fig.1), which results in 5 features for HCP: $$$f=\{f^1_0,f^2_0,f^3_0,f^1_2,f^2_2\}$$$, and 3 for 2-shell60 and hex28: $$$f=\{f^1_0,f^2_0,f^1_2\}$$$. Figures 4 and 5 show maps of the obtained microstructural parameters. According to Fig.3, the volume fractions are reliable while the diffusivities are not.


The proposed method is 2--3 order of magnitude faster than the commonly used MAP (nonlinear fitting) approach. A further difference with MAP is how the two methods behave when data are insufficient for determination of all model parameters. MAP needs to be stabilized by a priori constraints, while the present approach can work with the unconstrained microstructural tissue models while monitoring the detectability of model parameters as illustrated in Fig.3. Effectively, constraints are replaced by the prior distributions of parameters.

Refraining from constraints in any form is not possible for the common acquisition schemes. According to our results shown in Fig.3 only two independent microstructural parameters, the volume fractions can be found reliably. In particular, this explains the surprisingly high quality of the diffusivity maps for the fastest hex28 measurement (Fig.5): The output values are determined by the prior distribution with some influence of gray or white matter is considered. One more detectable mesoscopic parameter is the width of the FOD, so the total number of detectable parameters is three.

The robustness and efficiency of the present method offers it as a working horse for development of better acquisition schemes, more adequate tissue models and clinically oriented research.


We are grateful to Els Fieremans, Dmitry Novikov and José Solera for stimulative discussions. This study was supported by German Research Foundation (DFG) via grants RE 3286/2-1 and DFG KI 1089/3-2.


[1] Ileana O. Jelescu, Jelle Veraart, Els Fieremans, and Dmitry S. Novikov. Degeneracy in model parameter estimation for multicompartmental diffusion in neuronal tissue. NMR in Biomedicine, In print, 2015.

[2] Valerij G Kiselev and Kamil A Il’yasov. Is the “biexponential diffusion” biexponential? Magn Reson Med, 57(3):464–9, Mar 2007.

[3] Els Fieremans, Jens H Jensen, and Joseph A Helpern. White matter characterization with diffusional kurtosis imaging. Neuroimage, 58(1):177–88, Sep 2011.

[4] Dmitry S. Novikov, Ileana O. Jelescu, and Els Fieremans. From diffusion signal moments to neurite diffusivities, volume fraction and orientation dis- tribution: An exact solution. In Proc. 23st Annual Meeting of ISMRM, page 469, Toronto, Canada, June 2015.

[5] Sune N Jespersen, Christopher D Kroenke, Leif Østergaard, Joseph J H Ackerman, and Dmitriy A Yablonskiy. Modeling dendrite density from magnetic resonance diffusion measurements. Neuroimage, 34(4):1473–86, Feb 2007.

[6] Enrico Kaden, Frithjof Kruggel, and Daniel C Alexander. Quantitative mapping of the per-axon diffusion coefficients in brain white matter. Magn Reson Med, May 2015.

[7] David C Van Essen, Stephen M Smith, Deanna M Barch, Timothy E J Behrens, Essa Yacoub, Kamil Ugurbil, and WU-Minn HCP Consortium. The WU-Minn human connectome project: an overview. Neuroimage, 80:62–79, Oct 2013.

[8] Hui Zhang, Torben Schneider, Claudia A Wheeler-Kingshott, and Daniel C Alexander. NODDI: practical in vivo neurite orientation dispersion and density imaging of the human brain. Neuroimage, 61(4):1000–16, Jul 2012.


Upper row: Examples of feature maps. Bottom row: The quantities $$$c_l^{b_1,b_2}$$$ are used for selecting reliable features. In the absence of noise $$$c_l^{b_1,b_2}=1$$$. Deviation from unity indicates a too high noise contamination as suggested by the histogram for $$$c_4^{b_1,b_2}$$$.

The microstructural model includes an axonal compartment, an extra-axonal compartment and, optionally, a free water compartment (CSF) with a fixed diffusion coefficient $$$D_f=3\,\rm\mu m^2/ms$$$.

Correlation statistics of determined microscopic parameters with the ground truth on the training set for all measurement protocols with the focus on the most demanding HCP scheme (a--d). This analysis indicates insufficiency of all protocols for detecting the microscopic diffusivities for realistic noise level while the volume fractions are detectable.

Results for the HCP data: (a) Maps of microscopic parameters, (b) the distribution of these parameters within white matter and (c) the distribution of the normalized log-likelihood within white matter for two dispersion models considered here and NODDI [8].The narrow distribution of diffusivities follows from insufficient data (Fig.3).

Results as in Fig.4 for the hex28 protocol with (a) a single measurement and (b) with three repetitions. Parameter distributions (c) are shown for the case (a). The narrow distribution of diffusivities should be interpreted as insufficient data provided by this acquisition scheme (Fig.3).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)