Spectral Diffusion MRI

• Spectral Diffusion MRI is a technique that represents diffusion signals as continuous spectra over diffusivity, moving beyond single b-value models.
• It utilizes frequency-domain encoding with varied gradient waveforms to distinguish contributions from restricted diffusion and exchange phenomena in tissues.
• Advanced spatial-spectral inversion methods enable voxel-wise recovery of multidimensional diffusion and relaxation spectra, aiding compartment mapping in organs.

Spectral diffusion MRI denotes a family of diffusion MRI formulations in which the signal is represented through a spectrum rather than through a single effective diffusion time or a fixed low-order compartment model. In the literature considered here, this includes frequency-domain probing of the diffusion spectrum $D(\omega)$ with time-varying gradients, voxel-wise recovery of distributions over diffusivity and relaxation parameters, and spectral analyses of Bloch–Torrey dynamics in periodic media (Chakwizira et al., 2023, Liu et al., 2024, Moutal et al., 2020, Boito et al., 2021). In renal multi-$b$-value applications, the same term is also used for model-free decomposition of diffusion-weighted signal decay into continuous spectra of exponential components, with peaks assigned to vascular perfusion, tubular filtrate motion, and tissue parenchyma diffusion (Liu et al., 8 Aug 2025, Liu et al., 2024).

1. Conceptual scope and defining formulations

A central premise of spectral diffusion MRI is that diffusion encoding is not exhaustively characterized by a single scalar $b$-value or by a single “diffusion time.” For arbitrary gradient waveforms $g(t)$, the dephasing vector is

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$

with Fourier transform $q(\omega)$ and encoding power spectrum

$W(\omega)=|q(\omega)|^2.$

Under the Gaussian phase approximation, a frequency-domain signal representation takes the form

$$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$

This makes the encoding a filter over the frequency-dependent diffusivity $D(\omega)$ rather than a measurement at a single time scale (Chakwizira et al., 2023).

The same spectral logic appears in confined-diffusion models. In the confinement tensor model, the signal under arbitrary time-varying gradients is written as

$$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$

with $b$0. In this formulation, the kernel $b$1 is the spectral object carrying confinement-dependent time-scale information (Boito et al., 2021).

A different but related usage appears in spatial-spectral compartment mapping, where each voxel is modeled as a distribution over latent diffusion or diffusion-relaxation parameters. The forward model is linear in the spectral coefficients,

$b$2

and the inversion seeks a nonnegative spectroscopic image rather than a single fitted parameter per voxel (Liu et al., 2024).

In renal applications, spectral diffusion MRI is formulated as a non-negative sum of exponentials on a discretized diffusivity grid,

$b$3

with $b$4, so that the recovered spectrum can be partitioned into tissue, tubular, and vascular components (Liu et al., 8 Aug 2025). Across these variants, the unifying feature is spectral representation: either of encoding, of the medium, or of the operator governing diffusion-driven signal formation.

2. Frequency-domain encoding of restriction and exchange

The most explicit in vivo realization of spectral diffusion MRI in the frequency domain uses free gradient waveforms designed to vary sensitivity to restricted diffusion and exchange independently. In this framework, equal $b$5-values do not imply equal sensitivity, because the signal depends on the distribution of encoding power over frequency and on higher-order correlations of $b$6 (Chakwizira et al., 2023).

Two waveform-dependent quantities are central. Restriction-weighting is summarized by

$b$7

while exchange-weighting is defined from the fourth-order autocorrelation

$b$8

The primary fitting model used for in vivo exchange estimation is

$b$9

with

$b$0

In practice, the terms involving $b$1 and $b$2 were set to zero on the basis that $b$3 dominated in all tested protocols (Chakwizira et al., 2023).

The human brain study examined six healthy volunteers using both strong and ultra-strong gradients at 80, 200, and 300 mT/m. Discovery protocols at 200 and 300 mT/m used 150 unique free gradient waveforms each, spanning a wide range of $b$4 space. Application protocols used 12 waveforms, divided into six “exchange-encoding” waveforms with similar $b$5 and varying $b$6, and six “restriction-encoding” waveforms with similar $b$7 and varying $b$8. At 80 mT/m, the application protocol used four waveforms with two varying $b$9 at fixed $g(t)$0 and two varying $g(t)$1 at fixed $g(t)$2 (Chakwizira et al., 2023).

The results showed distinct time-dependence signatures in grey and white matter. Grey matter, including cerebellar cortex and thalamus, showed signal variations along both $g(t)$3 and $g(t)$4, consistent with both restricted diffusion and exchange. White matter, including internal capsule, showed signal variations predominantly along $g(t)$5, consistent with restricted diffusion dominating and negligible exchange on these timescales. Across subjects and gradient strengths, exchange in grey matter was at least twice as fast as in white matter, and the shortest exchange times observed were in the cerebellar cortex at approximately 115 ms. The grey-white matter exchange contrast from a 25-minute 300 mT/m protocol was preserved in a 4-minute 300 mT/m protocol and in a 10-minute 80 mT/m protocol, although the 4-minute subset showed more noise and lower absolute $g(t)$6 (Chakwizira et al., 2023).

These observations place spectral diffusion MRI beyond a simple PGSE-versus-OGSE distinction. PGSE largely weights low frequencies, OGSE emphasizes higher frequencies, whereas free waveforms were designed to shape both the encoding power spectrum and the exchange-sensitive higher-order filter. This suggests that the principal gain is not merely shorter effective diffusion time, but experimentally controllable non-colinearity between restriction and exchange sensitivities.

3. Confinement, Bloch–Torrey spectra, and periodic media

A second major line of work frames spectral diffusion MRI in terms of confined diffusion and operator spectra. In the confinement tensor model, diffusion in a Hookean potential is characterized by a confinement tensor $g(t)$7 and an effective diffusivity $g(t)$8, with $g(t)$9 in the scalar-diffusivity version used for fitting. A computationally convenient closed-form signal is

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$0

where

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$1

This model was proposed as a replacement for the unrestricted diffusion tensor distribution in multidimensional diffusion MRI methods with confined subdomains (Boito et al., 2021).

The corresponding distributional extension is the confinement tensor distribution,

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$2

which retains the non-exchanging compartment assumption but gives each compartment restricted-like time dependence. Simulations showed that the confinement tensor model captured both free and restricted diffusion signatures, including STE-specific restricted diffusion modulations that DTI missed. In a mixed protocol comprising 102 LTE/PTE/STE measurements, 102 TC-OGSE measurements at four distinct oscillation frequencies, and 13 $q(t)=\gamma\int_0^t g(\tau)\,d\tau,$3 images, the addition of OGSE improved diffusivity recovery in systems with identical geometry but different diffusivities (Boito et al., 2021).

A more operator-theoretic formulation appears in periodic media, where the transverse magnetization obeys the Bloch–Torrey equation

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$4

with Robin boundary condition $q(t)=\gamma\int_0^t g(\tau)\,d\tau,$5 on $q(t)=\gamma\int_0^t g(\tau)\,d\tau,$6. The Bloch–Torrey operator is

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$7

Because the gradient term is not invariant under lattice translations, direct spectral analysis on a single unit cell is not available in the naive form. The periodic-medium approach uses narrow pulses and pseudo-periodic boundary conditions so that the problem can be reduced to a single unit cell, yielding efficient computation and theoretical insight into Bloch bands, branching points, and mode localization (Moutal et al., 2020).

In this setting, the PGSE-like narrow-pulse signal in a periodic medium can be written as

$q(t)=\gamma\int_0^t g(\tau)\,d\tau,$8

At long $q(t)=\gamma\int_0^t g(\tau)\,d\tau,$9, the $q(\omega)$0 mode dominates, giving diffusion-diffraction peaks at $q(\omega)$1. At high gradient, eigenmodes localize near boundary points where the gradient is normal to the boundary, with localization scale $q(\omega)$2. The paper showed that branching points in the non-Hermitian spectrum mark the onset of localization, and that the low-gradient limit is singular in the sense that even as $q(\omega)$3 the Bloch–Torrey spectrum collapses to band averages rather than following ordinary Hermitian perturbation theory (Moutal et al., 2020).

Taken together, these formulations establish that “spectral” in spectral diffusion MRI can refer not only to the spectrum of diffusivities or the frequency content of encoding waveforms, but also to the spectral structure of the governing operator itself.

4. Spatial-spectral inversion and compartment mapping

A third usage of spectral diffusion MRI concerns voxel-wise recovery of multidimensional parameter distributions, especially diffusivity–relaxation spectra. The starting point is the linear mixture model

$q(\omega)$4

which, after stacking measurements, becomes $q(\omega)$5. The inverse problem is regularized spatially through

$q(\omega)$6

and estimated by solving

$q(\omega)$7

This is a convex spatial-spectral partial volume mapping problem with a linear-mixture forward model, spatial regularization, and nonnegativity constraints (Liu et al., 2024).

The computational contribution was a tailored implementation of LADMM. With a single split $q(\omega)$8, the $q(\omega)$9-subproblem was made voxelwise by setting $W(\omega)=|q(\omega)|^2.$0, and the $W(\omega)=|q(\omega)|^2.$1-subproblem was simplified by choosing

$W(\omega)=|q(\omega)|^2.$2

The resulting method avoids the more expensive spatially coupled linear solves that arise in prior ADMM formulations. A truncated SVD of $W(\omega)=|q(\omega)|^2.$3 further reduces the cost of applying $W(\omega)=|q(\omega)|^2.$4 from $W(\omega)=|q(\omega)|^2.$5 to $W(\omega)=|q(\omega)|^2.$6 (Liu et al., 2024).

In diffusion–relaxation spectral mapping, the dictionary for DR-CSI used a $W(\omega)=|q(\omega)|^2.$7 grid over $W(\omega)=|q(\omega)|^2.$8, giving $W(\omega)=|q(\omega)|^2.$9. On ex vivo spinal cord data with $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$0 measurements and $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$1 voxels, LADMM achieved the DFCS reached by ADMM after 1.5 minutes in 0.48 minutes, approximately $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$2 faster, with peak memory 2.5 MB versus 22.9 MB. Across additional MRI settings, the reported speed improvements ranged from $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$3 to $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$4, with reduced memory use in each case. The method also produced nearly identical spatially averaged spectra and component maps to ADMM while improving feasibility for large-scale spatially regularized inversions (Liu et al., 2024).

This spatial-spectral line of work broadens the meaning of spectral diffusion MRI from waveform design to inverse-problem formulation. Rather than tailoring the encoding spectrum alone, it reconstructs a spectrum of latent compartments per voxel and couples neighboring voxels through regularization to improve identifiability and robustness.

5. Multi-$$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$5-value spectral diffusion in renal MRI

In renal applications, spectral diffusion MRI is implemented as a model-free multicomponent analysis of diffusion-weighted decay curves. The voxel-wise normalized signal is fitted by non-negative least squares with Tikhonov smoothness regularization,

$$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$6

In the 2025 kidney allograft study, $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$7 and $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$8, with the diffusion grid logarithmically spaced over $$\ln\frac{S}{S_0}\approx-\frac{1}{2\pi}\int_{-\infty}^{\infty} D(\omega)\,|q(\omega)|^2\,d\omega.$$9–$D(\omega)$0 in units of $D(\omega)$1. Peaks were grouped into tissue ($D(\omega)$2), tubular ($D(\omega)$3), and vascular ($D(\omega)$4) compartments, while peaks below $D(\omega)$5 were excluded because $D(\omega)$6 was insufficient to reliably capture very slow diffusion (Liu et al., 8 Aug 2025).

Compartment metrics were defined by

$D(\omega)$7

The same feature family was used as a basis for statistical classification of fibrosis and renal dysfunction. In a prospective two-center study of 99 kidney allograft recipients scanned at 3T with identical protocols on Siemens Skyra and Siemens Prisma systems, spectral diffusion detected fibrosis in patients with normal/stable $D(\omega)$8 with $D(\omega)$9, $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$0. It detected mild/moderate fibrosis with $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$1, $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$2. The study also reported that interobserver correlation exceeded 0.50 in 24 out of 40 diffusion parameters, with tissue diffusion components showing better reliability than tubular components (Liu et al., 8 Aug 2025).

A related 2024 study formulated quantitative intravoxel flow estimation from the same spectral decomposition by introducing water transport time and converting $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$3 to conventional flow units. For kidneys, with $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$4 and $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$5, the paper gave

$$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$6

and for a 120 g kidney,

$$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$7

In simulations, spectral diffusion correlated strongly with truth for three-compartment anisotropic diffusion, with $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$8, $$E=\exp\!\left(- \int \frac{d\omega}{2\pi}\; \hat{\mathbf{G}}^\dagger(\omega)\,\mathbf{K}(\omega)\,\hat{\mathbf{G}}(\omega)\right), \qquad \mathbf{K}(\omega)=2\gamma^2\,\mathbf{D}\,(\omega^2\mathbf{I}+\mathbf{\Omega}^2)^{-1},$$9, and for two-compartment anisotropic diffusion, with $b$00, $b$01. A fixed regularization parameter $b$02 increased computation up to 208-fold and agreed with voxel-wise cross-validated regularization with concordance correlation coefficient $b$03. In renal allografts, tissue parenchyma compartment $b$04 increased significantly with fibrosis, and tubular $b$05 decreased significantly in allografts with impaired function (Liu et al., 2024).

These renal studies make a distinct conceptual move. Rather than fixing a bi-exponential or tri-exponential model a priori, they treat the decay curve as a nonparametric spectrum and let physiologic compartment counts vary across voxels. This suggests that, in organs with variable compartment number, spectral diffusion MRI functions as a flexible alternative to rigid IVIM parameterization.

6. Limitations, interpretive issues, and ongoing directions

Several limitations recur across the spectral diffusion MRI literature. In frequency-domain waveform methods, the modeling rests on Gaussian phase approximation or cumulant expansions, and simplified parameterizations may neglect anisotropy and intra-compartmental kurtosis. In the human brain exchange study, the Kärger-type exchange assumptions were noted to be often violated in gray matter, exchange was described as likely multiexponential, and protocols with shorter minimum $b$06 were said to be more sensitive to faster components. Structural disorder can mimic exchange in the sense of “loss of diffusional heterogeneity,” producing apparent exchange. At the acquisition level, peripheral nerve stimulation limits slew rate, and oscillatory restriction-weighted waveforms are $b$07-inefficient at low $b$08, which limited the range of $b$09 at 80 mT/m (Chakwizira et al., 2023).

In confinement-tensor and operator-spectral approaches, the inverse problem remains ill-posed and noise-sensitive. The confinement tensor distribution framework assumes non-exchanging compartments and commuting $b$10 and $b$11, with axisymmetry constraints used in practice. The periodic-medium Bloch–Torrey analysis further showed that the low-gradient limit is singular and that non-analytic branching points delimit the validity of low-gradient perturbative intuition. A plausible implication is that spectral diffusion MRI should not be reduced to a single asymptotic regime: low-gradient band averaging, motional narrowing, and high-gradient localization each emphasize different geometric observables (Boito et al., 2021, Moutal et al., 2020).

In spatial-spectral compartment mapping, computational feasibility has improved, but the usual inverse-problem trade-offs persist. Accuracy depends on the physics encoded in the dictionary and on grid resolution; too coarse a grid biases spectra, while too fine a grid increases $b$12 and ill-posedness. Spatial smoothing improves robustness but can bias edges and small compartments. The nonnegativity-only constraint used in LADMM was explicit, and no simplex or sum-to-one constraint was imposed (Liu et al., 2024).

In renal spectral diffusion, compartment boundaries were empirically defined, the slow-diffusion cutoff depended on $b$13, and external validation was still needed. The 2025 fibrosis study noted that inflammation and rejection could act as confounders, anisotropic tubules and $b$14 effects were beyond scope, tubular features were less reproducible than tissue features, and longitudinal studies were needed to establish prognostic value for early chronic kidney disease progression and antifibrotic therapy monitoring (Liu et al., 8 Aug 2025). The 2024 flow study similarly emphasized moderate anisotropy, Gaussian diffusion per compartment, and the need for further validation of peak-to-compartment assignment and for broader assessment across scanners and conditions (Liu et al., 2024).

One common misconception addressed directly by this literature is that matching $b$15-value is sufficient to match microstructural sensitivity. Spectral diffusion MRI rejects that equivalence: two waveforms with the same $b$16 can have different $b$17 and $b$18, two voxels with the same ADC can have different diffusivity spectra, and two geometries with similar short-time behavior can differ in Bloch–Torrey spectral structure. The field therefore remains defined less by a single pulse sequence than by a spectral viewpoint on encoding, inversion, and interpretation.