Analytical phase kurtosis of the constant gradient spin echo

TL;DR

This work addresses the validity of the Gaussian phase approximation (GPA) in diffusion MR by analytically deriving the excess phase kurtosis κ4/κ2^2 for the constant gradient spin echo (CGSE) in three 1D models: Poisson pore-hopping, trapped-release, and restricted diffusion. It develops a cumulant framework for phase statistics, derives exact expressions for κ2 and κ4, and validates them with Monte Carlo simulations, showing that GPA often fails under moderate gradient strengths and encoding times. The key contributions are the explicit, non-Gaussian phase predictions across disparate microstructures, a critical critique of diffusion-based displacement kurtosis models, and practical guidelines for when GPA-based interpretations remain valid. The results have broad implications for diffusion MR microstructure modeling, suggesting that phase-based analyses and non-perturbative approaches are needed to accurately capture signal behavior in realistic gradient encodings.

Abstract

The Gaussian phase approximation (GPA) underlies many standard diffusion magnetic resonance (MR) signal models, yet its validity is rarely scrutinized. Here, we assess the validity of the GPA by analytically deriving the excess phase kurtosis $κ_4/κ_2^2$, where $κ_n$ is the $n^{\text{th}}$ cumulant of the accumulated phase distribution due to motion. We consider the signal behavior of the spin echo with constant gradient amplitude $g$ and echo time $T$ in several one-dimensional model systems: (1) a stationary Poisson pore-hopping model with uniform pore spacing $Δx$ and mean inter-hop time $τ_{\text{hop}}$; (2) a trapped-release model in which spin isochromats are initially immobilized and then released with diffusivity $D$ following an exponentially-distributed release time, $τ_{\text{rel}}$; and (3) restricted diffusion in a domain of length $L$. To our knowledge, this is among the first systematic analytical treatments of spin echo phase kurtosis without assuming Gaussian compartments or infinitesimally short gradient pulses. In the pore-hopping system, $κ_4/κ^2_2 = (9/5)τ_{\text{hop}}/T$, inversely proportional to the mean hop number, $T/τ_{\text{hop}}$. In the trapped-release system, $κ_4/κ_2^2$ is positive and decreases roughly log-linearly with $T/\langleτ_{\text{rel}}\rangle$, where $\langleτ_{\text{rel}}\rangle$ is the average release time. For restriction, $κ_4/κ_2^2$ vanishes at small and large $L/\sqrt{DT}$, but has complicated intermediate behavior. There is a negative peak at $L/\sqrt{DT}\approx 1.2$ and a positive peak at $L/\sqrt{DT}\approx 4.4$. Monte Carlo simulations are included to validate the analytical findings. Overall, we find that the GPA does not generally hold for these systems under moderate experimental conditions, i.e., $T=10\;\mathrm{ms}$, $g\approx 0.2-0.6\;\mathrm{T/m}$.

Figures (8)

• Figure 1: Exact signal (solid, black) and $n^{\text{th}}$-order approximations for the Poisson pore-hopping model. (a) Signal vs. $T/\tau_{\text{hop}}$, where $T = 10\;\mathrm{ms}$, $g = 0.25\;\mathrm{T/m}$ were fixed, and $\tau_{\text{hop}}$ was varied between $0.01 - 100\;\mathrm{ms}$, linearly spaced by $0.01\;\mathrm{ms}$. The overall mobility was also fixed at $D_{\text{eff}} = 2\;\mathrm{\mu m^2/ms}$, with the spacing scaled according to Eq. \ref{['eq: Deff poisson']}: $\Delta x = \sqrt{2D_{\text{eff}}\tau_{\text{hop}}}$. The GPA corresponds to $S^{(2)}$ (solid, light green). Higher-order approximations (dashed) are also shown. (b) Signal vs. $g^2$, with $g$ varied linearly between $0 - 0.6\;\mathrm{T/m}$ by $0.01$, and $T/\tau_{\text{hop}} = 4$ fixed. Other parameters were the same as part (a), and the same legend applies. A secondary axis with $b = \gamma^2g^2T^3/12$ is included for reference.
• Figure 2: $\kappa_4/\kappa_2^2$ vs. $\alpha = kT$ for the trapped-release model with mean release time $1/k$. This behavior does not depend on $D$ or $g$.
• Figure 3: Signal approximations and simulated signal for the trapped-release model. (a) $S^{(2)}$ (solid) and $S^{(4)}$ (dashed) vs. $\alpha = kT$, with $\alpha$ varied log-linearly between $10^{-2} - 10^2$ at $200$ points. The echo time $T = 10\;\mathrm{ms}$ and $D = 2\;\mathrm{\mu m^2/ms}$ were fixed, while $g =$ [0.15, 0.25, 0.35] $\mathrm{T/m}$ was varied (colors). (b) $S^{(2)}$ (solid), $S^{(4)}$ (dashed), and MC simulated signal (circles) vs. $g^2$ for fixed values of $\alpha =[2,\,5]$ (colors), with $g$ varied linearly between $0-0.6$ by $0.01\;\mathrm{T/m}$. Other parameters were kept the same as part (a). A secondary axis with $b = \gamma^2 g^2 T^3/12$ is included for reference. Note that $S^{(2)}$ scales with $g^2$ while $S^{(4)}$ scales sub-linearly.
• Figure 4: $\kappa_4/\kappa_2^2$ vs. $L/\sqrt{DT}$ for a restricted diffusion in domain length $L$. Parameters $T = 10\;\mathrm{ms}$ and $D = 2\;\mathrm{\mu m^2/ms}$ were fixed, while $L$ was varied. The kurtosis was estimated using $\langle\phi^4 \rangle$ in Eq. \ref{['eq: phi4 restriction']} and $\langle \phi^2 \rangle$ in Eq. \ref{['eq: phi2 restriction']}, truncated at $a,b,c,m \le 101$.
• Figure 5: Phase distribution $P(\phi)$ at several values of $L/\sqrt{DT} = [1.2,\,4.4,\,25]$ for fixed $T = 10\;\mathrm{ms}$, $D = 2\;\mathrm{\mu m^2/ms}$, and $g = 0.35\;\mathrm{T/m}$. MC simulated phase (blue) are compared to a Gaussian (dashed), obtained from the mean and standard deviation of simulated data. See main text for simulation methods. From left-to-right, the first plot lies near to the negative peak in Fig. \ref{['fig: restriction kurt']}, the second to the positive peak, while the third approaches the Gaussian limit. The simulated $\kappa_4/\kappa_2^2$ values are $\approx -0.42,\,0.29, \,0.08$, respectively, in agreement with Fig. \ref{['fig: restriction kurt']}. Insets highlight behavior at the left tail of the distributions. Note the truncated tails in the first plot vs. approximately Gaussian tails in the second.