Dependence of the Extra‐Cellular Diffusion Coefficient on the Fractions of Neurites and Cell Bodies in Gray Matter Authors: Hong‐Hsi Lee, Ali Abdollahzadeh, Hansol Lee, Ricardo Coronado‐Leija, Els Fieremans, Susie Y. Huang, Dmitry S. Novikov Published: 2026 ABSTRACT PurposeThe dependence of the long‐time (tortuosity) limit of the extra‐cellular diffusivity on the intra‐cellular volume fraction is of fundamental importance for microstructure modeling. While such dependencies have been explored for the white matter, the tortuosity limit in gray matter is unknown due to complex cell composition and geometry. Here we rationalize and validate numerically the analytical relation between the extra‐cellular diffusivity and intra‐cellular fractions of cell bodies (somas) and neurites.MethodsThe tortuosity relation for extra‐cellular diffusivity qualitatively follows from effective medium theory, coarse‐grained by diffusion outside somas (spheres) and neurites (cylinders), respectively. This problem is equivalent to finding the overall conductivity in a medium of grains in a matrix, with methodology dating back to the 19th century. We extend the effective medium methodology to populations of impermeable spheres and randomly oriented cylinders with various volume fractions, yielding closed‐form expressions corroborated by Monte Carlo simulations.ResultsWe establish the power‐law scaling of the extra‐cellular diffusivity with the volume fractions of the extra‐soma and extra‐neurite spaces. We further evaluate the proposed framework using simulations in realistic tissue geometries, and by applying it to in vivo MRI data.ConclusionTheory and simulations relate extra‐cellular tortuosity to soma and neurite fractions, thereby offering a diffusion MRI protocol design optimized for in vivo assessment of soma size and soma/neurite fractions within clinical scan times. Such in vivo measurements can be used to study development, aging, and neurodegenerative disorders. Introduction Gray matter in the human brain is composed of neuronal cell bodies (somas), dendrites and axons (neurites), along with glial cells and the extracellular space [1]. Alterations in soma size and soma/neurite density have been observed in normal aging and neurodegenerative diseases [2, 3, 4, 5]. These findings call for an in vivo histological imaging technique to evaluate soma and neurite microstructure that could be used for earlier detection of neurodegeneration and monitoring of treatment response. Diffusion MRI (dMRI) is an imaging technique sensitive to tissue micro‐geometries at a length scale commensurate with the cell size [6], three orders smaller than the image voxel size, allowing in vivo estimation of tissue histopathology. Compared with conventional diffusion signal representations, such as diffusion tensor imaging [7] and diffusion kurtosis imaging [8], biophysical modeling of dMRI enables characterization and interpretation of tissue microstructure in terms of specific parameters, such as cell size [9, 10, 11, 12, 13, 14, 15, 16], intra‐ and extra‐cellular diffusivities and water fractions [17, 18, 19, 20]. Relating the long‐time limit of extra‐cellular diffusivity D to packing geometry of impermeable cells, with the extra‐cellular volume fraction ϕ and intrinsic diffusivity D0, maps onto a classic problem of finding electrical dc conductivity σ(ϕ) of a mixture made of non‐conducting grains with volume fraction 1−ϕ, immersed in a medium with conductivity σ0 [21]. As the diffusion coefficient is normalized per diffusing molecule, while the conductivity grows with the number of charge carriers, the ratio by which the diffusion coefficient is reduced, (1) 1Λ(ϕ)≡D(ϕ)D0=σ(ϕ)σ0·1ϕ, involves the dimensionless conductivity σ/σ0 further divided by the extra‐grain fraction ϕ [22]. It is conventional to introduce the dimensionless tortuosity factor Λ(ϕ)≥1 that tells by how much the extra‐cellular diffusivity is suppressed relative to free D0. As the effective medium theory (EMT) is naturally developed for conductivity, rather than for tortuosity, in what follows we will consider scalings for σ(ϕ)/σ0 and translate to tortuosity via Equation (1). The mapping (1) onto conductivity relates the tortuosity problem to a test‐bed of effective medium theories starting from Maxwell's work in the 19th century [21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]. The EMT‐derived tortuosity relation works well when the grains are loosely packed with low volume fraction, or when the electrical (or equivalently diffusion) properties of grains and matrix are similar. However, neurons and glial cells in the brain are tightly packed with high intra‐cellular volume fractions ∼ 80% [35], making it more challenging to derive the accurate tortuosity relation. In white matter, where axons are tightly packed and effectively impermeable (ϕ∼0.2–0.5), traditional effective medium models lose predictive power, with classical bounds [27] and self‐consistent schemes [25, 26] erring by as much as 100% for the diffusion coefficient D transverse to the tracts [34]. The difficulty arises since in this 2‐dimensional (2d) setting, each axon acts as a strong hindrance to diffusion, reshaping the trajectories of water molecules that are further complicated by the close proximity of neighboring axons, which eliminate any obvious small parameter to guide perturbative analysis. Interestingly, the 2d tortuosity relation of densely packed axons (disks) in white matter was solved by accounting for the axon diameter distribution [34], approximated by a small population for the bulk distribution and another large population for the tail of the distribution. The result is notably different from the 2d mean‐field tortuosity relation Λ=1/ϕ [25, 26]. To achieve a better agreement with simulations in tightly packed geometries, diffusion coarse‐graining was modeled in two stages [34], by first averaging the effect of small axons over scales larger than their size, and subsequently treating the rare large axons as an additional restriction to the reduced diffusivity. However, due to the complexity of gray matter microstructure, the tortuosity relation for gray matter—an essentially 3d problem—has remained unknown. Under a simplified biophysical picture (Figure 1A), gray matter can be considered as a collection of somas (spheres) and neurites (sticks) embedded into an extra‐cellular space (isotropic hindered diffusion) [16]. For gray matter modeling, a tortuosity relation of highly aligned (2d) cylinders along with randomly packed spheres has been applied [36]. However, this model was originally designed for the bovine optic nerve [37], where the mean‐field tortuosity relation did not apply well due to densely packed axons, and the highly aligned cylinders did not match the biological picture of highly dispersed neurites in gray matter [1]. (A) Gray matter in the human brain consists of complex micro‐geometries with abundant cell bodies (somas, >10μm in diameter) and highly dispersed thin neurites (∼1μm in diameter) [1], which can be considered as a collection of randomly packed spheres (blue) and randomly oriented thin cylinders (magenta) with an extra‐cellular space (green). (B) Sequential coarse‐graining solution: To derive the extra‐cellular tortousity of this gray matter model, we consider a two‐step diffusion coarse‐graining process that has a specific order due to the length scale separation of soma and neurite (Section 2.3). (C) Simultaneous coarse‐graining solution: Alternatively, we propose a two‐step diffusion coarse‐graining process that has no specific order (Section 2.4), insensitive to the length scale separation of soma and neurite. In this work, to elucidate the tortuosity relation of extra‐cellular space in gray matter, we start from a picture of diffusion coarse‐graining in a medium composed of randomly packed spheres (somas) and randomly oriented, randomly positioned cylinders (neurites) using EMT (Figure 1). Furthermore, we perform Monte Carlo (MC) simulations of diffusion in the extra‐cellular space of media composed of randomly packed spheres and dispersed cylinders, and validate the derived tortuosity relation across varying sphere and cylinder volume fractions, including the extreme case of densely packed geometries. We additionally perform simulations in a realistic extra‐cellular space geometry segmented from electron microscopy data of the mouse olfactory bulb [35]. To demonstrate the practical relevance of the proposed framework, we apply it to in vivo MRI data. The significance of the developed tortuosity relation lies in reducing the data requirements for dMRI modeling in gray matter, thereby facilitating clinical translation. Theory Tissue micro‐geometry in brain gray matter is complex [1], with an abundance of cell bodies and highly dispersed dendrites and axons. To resolve the tortuosity relation of extra‐cellular diffusivity in gray matter, we simplify the gray matter microstructure, representing it as a collection of randomly packed spheres and randomly oriented cylinders. The spheres do not overlap with each other, while the cylinders may overlap with each other, mimicking dendritic arborization (branching) [38], spines and leaflets [39]. First, we discuss two simple cases, a medium composed of either spheres or cylinders; then we consider the more complex case of a medium composed of both spheres and cylinders. Differential Addition The idea of effective‐medium conductivity originates from the following 19th century observation: For a small fraction f≪1 of spheres of conductivity σ1 embedded in the matrix with conductivity σ0, the conductivity σ(f) of this composite medium follows the following generalized Clausius–Mossotti relation [23, 24] (2) σ−σ0σ+2σ0=f·σ1−σ0σ1+2σ0. (The original relation was formulated for the dielectric permittivity of a mixture, as a solution to the same Laplace equation.) Essentially, a large sphere of volume V and conductivity σ embedded in a matrix of σ0 distorts the current (or is polarized) approximately in the same way as N small spheres (of volume v1) with conductivity σ1 and total volume v=Nv1 within the same V, where f=v/V≪1; this is a mean‐field view that the “polarizabilities” (σ1−σ0)/(σ1+2σ0) add up. Bruggeman [25] was first to realize that Equation (2) can be written in a differential form (from a modern perspective, it appears as a precursor to the real‐space renormalization group equation): Suppose we have already added spheres with total volume v into a matrix with volume V−v, such that the spheres have a finite volume fraction f=v/V; the whole composite medium has an effective conductivity σ0→σ(v). Adding a bit more spheres of v→v+dv would change the conductivity σ→σ+dσ; according to Equation (2): (3) dσ3σ=−12dvV, where we set σ1≡0 (non‐conducting spheres) in the right‐hand side of Equation (2) for our case, and used the infinitesimal fraction dv/V in lieu of f. Since, df=v+dvV+dv−vV≃dvV(1−f), we can write Equation (3) as (4) dlnσ=−32df1−f=32dlnϕ,ϕ=1−f. The solution of this differential equation, with the initial condition σ|ϕ=1=σ0, is σ/σ0=ϕ3/2 [25, 26], yielding the mean‐field 3d tortuosity Λs(ϕ)=ϕ−1/2 for the spheres, compare Equation (1). This so‐called Archie's law [26] generalizes to 2‐dimensional disks, for which Equation (2) has σ+σ0 in the denominator instead of σ+2σ0, such that the corresponding scaling σ/σ0=ϕ2, Λdisk(ϕ)=1/ϕ. Generalizing for a mixture of randomly oriented ellipsoids, one obtains [32] (5) dlnσ=νdϕ,ν=13∑j=1311−Lj where Lj are the demagnetizing factors. For the spheres, all Lj=1/3. For randomly oriented cylinders, L1=L2=1/2, and L3=0. This yields σ=σ0ϕν with ν=νc,s for randomly oriented cylinders (c) and spheres (s): (6) νs=32,νc=53, Compare Equation (4), such that the tortuosity Λc(ϕ)=ϕ−2/3 for randomly oriented cylinders. Mixture of Cylinders and Spheres Imagine now composing a medium of two kinds of grains—non‐conducting cylinders and spheres with fractions fc,s=vc,s/V. By the above mean‐field logic, we have [40] (7) dlnσ=−νcdvcV−νsdvsV. To solve this differential equation for σ(vc,vs), we need to further specify the order of adding different grains to the matrix of conductivity σ0—that is, to provide the relation between vc and vs. The simplest way is to keep a constant ratio λ=vc/vs=fc/fs at all times; when dvc/dvs=vc/vs, dfc,s1−fc−fs=dvc,sV which turns the right‐hand side of Equation (7) into a full derivative: dlnσ=−νcdfc1−fc−fs−νsdfs1−fc−fs=ν·dlnϕ, where ϕ is the overall extra‐cellular volume fraction, (8) ϕ=fe=1−fc−fs. The solution for σ|ϕ=1=σ0 is (9) σmixσ0=ϕν,DmixD0=ϕν−1, where the exponent (10) ν=λ1+λνc+11+λνs=fcνc+fsνsfc+fs is the volume‐weighted average of the exponents (6) for both kinds of grains. The above solution nicely interpolates between either kind of grains, f=fc or f=fs in Section 2.1, via its exponent (10). Experience tells [34] that such mean‐field theory can work down to moderate extra‐cellular fractions ϕ, but would eventually break down for dense packings, ϕ≪1. In what follows, we will explore this solution (9) and (10) numerically and identify two other empirical solutions that may perform better for low ϕ. Sequential Coarse‐Graining In the gray matter, soma size (>10μm in diameter) is much larger than neurite thickness (∼1μm in diameter) [1]. To account for the size disparity between soma and neurites, we propose a sequential two‐step diffusion coarse‐graining process, noting that swapping the order of coarse‐graining over different structures changes the results [34] (Figure 1B). The diffusion length increases with the square root of the time elapsed [41]. Given a strong separation of scales, at shorter times diffusion coarse‐grains the finer objects (neurites/cylinders), yielding an effective medium with a reduced conductivity σc=σ0ϕcνc in the background of the larger objects (soma/spheres). Here, the extra‐cylinder fraction (in the space not occupied by spheres) is given by (11) ϕc=fefc+fe=1−fc1−fs=ϕ1−fs, with ϕ in Equation (8). The reduced conductivity σc is then used as that of a uniform matrix, as an input to the second coarse‐graining to account for the hindrance induced by the spheres/somas at longer times, yielding (12) σseq=σc·(1−fs)νs=σ0ϕcνc·(1−fs)νs. Equivalently, we could start from the effective diffusivity Dc/D0=(σc/σ0)/ϕc=ϕcνc−1 of a uniform matrix, and further correct it by the factor (1−fs)νs−1: (13) DseqD0=1−fc1−fs2/3·(1−fs)1/2. Note that the conductivity (12) and diffusivity (13) satisfy Equation (1), since using Equation (11) leads to σseqσ0·1ϕ=ϕcνc−1(1−fs)νs−1=DseqD0. As we will see below, this solution empirically works for loosely packed cells (e.g., fc+fs≤0.3) with the size disparity between large spheres and thin cylinders [34]. To lift the requirement of size disparity, next we propose an alternative, empirical solution, with simultaneous diffusion coarse‐graining over cylinders and spheres, different from Equations (9) and (10). Simultaneous Coarse‐Graining Given that the order of coarse‐graining processes originates from the size disparity between cylinders and spheres, we hypothesize that a simultaneous coarse‐graining process for cylinders and spheres would potentially yields a solution applicable to the case without size disparity. In other words, diffusion homogenizes the fine details of the cylinders and spheres on a similar time scale (Figure 1C), such that neither compartment can be treated as effectively coarse‐grained first. This suggests the following guess for the diffusivity of a simultaneously coarse‐grained mixture: (14) DsimD0=ϕcνc−1ϕsνs−1=1−fc1−fs2/3·1−fs1−fc1/2, where analogously to Equation (11), we formally introduced the extra‐sphere fraction (15) ϕs=fefs+fe=1−fs1−fc in the space not occupied by the cylinders. This solution is an empirical guess and does not strictly follow the EMT, since there is no justifiable conductivity that is related to Equation (14) via Equation (1). Therefore, it should be viewed as a phenomenological interpolation that captures the combined hindrance effects of cylinders and spheres in regimes where their contributions are not hierarchically separable. This formulation is expected to be most applicable in densely packed cylinders and spheres with comparable characteristic length scales, while in dilute limits with the presence of strong size disparity, the sequential coarse‐graining solution (13) remains more physically justified. Surprisingly, despite its heuristic nature, we find that Equation (14) provides accurate predictions across a broad range of simulated configurations as shown below. Methods Micro‐Geometry Generation Our goal was to generate a gray matter‐mimicking micro‐geometry composed of spheres and cylinders, representing cellular components such as somas and neurites, respectively. The simulation domain was a cube of side length L, within which we sequentially placed two populations of structures: non‐overlapping, randomly packed spheres and randomly oriented, partially overlapping cylinders. The process and associated volume fraction definitions were detailed below. First, Ns spheres of identical diameter ds=2rs were placed, assuming a random close‐packing scheme where no two spheres overlapped, and all spheres were entirely contained within the cube [42]. The resulting packing achieved a target intra‐spherical volume fraction fs, defined as fs=VsL3, where Vs=Ns·4π3rs3 is the total volume of all non‐overlapping spheres, including any overlapping volume between spheres and cylinders. For the given sphere radius, cube size, and intra‐spherical volume fraction, we calculated the number of spheres accordingly. After spheres were placed, Nc cylinders of identical diameter dc=2rc were added to the geometry. Each cylinder was generated using a homogeneous Poisson line process (mean free path randomness) [43] to ensure uniform sampling of orientations and positions: We defined a spherical domain of radius R=5L, centered at the cube's center.A direction vector n was sampled uniformly over a unit sphere.A random point x was selected within the spherical domain, and a line segment was constructed from the point of intersection on one side of the spherical domain to the opposite side, resulting in a finite‐length segment enclosed in the spherical domain.Each line was checked for intersection with the cube. Only lines that intersected the cube were retained.The segment of each intersecting line contained within the cube was used as the cylinder axis.A solid cylinder of fixed diameter dc was constructed along each line segment. We defined a spherical domain of radius R=5L, centered at the cube's center. We defined a spherical domain of radius R=5L, centered at the cube's center. A direction vector n was sampled uniformly over a unit sphere. A direction vector n was sampled uniformly over a unit sphere. A random point x was selected within the spherical domain, and a line segment was constructed from the point of intersection on one side of the spherical domain to the opposite side, resulting in a finite‐length segment enclosed in the spherical domain. A random point x was selected within the spherical domain, and a line segment was constructed from the point of intersection on one side of the spherical domain to the opposite side, resulting in a finite‐length segment enclosed in the spherical domain. Each line was checked for intersection with the cube. Only lines that intersected the cube were retained. Each line was checked for intersection with the cube. Only lines that intersected the cube were retained. The segment of each intersecting line contained within the cube was used as the cylinder axis. The segment of each intersecting line contained within the cube was used as the cylinder axis. A solid cylinder of fixed diameter dc was constructed along each line segment. A solid cylinder of fixed diameter dc was constructed along each line segment. To achieve the target intra‐cylindrical volume fraction fc, the number of cylinders was determined by [44] (Supporting Information S1) (16) Nc=−L3⟨vc⟩·ln1−fc1−fs, where (17) ⟨vc⟩=23πrc2L was the average cylinder volume inside the cube. This approach ensured that the cylinders collectively occupied the correct fraction of the free volume not already filled by spheres, while preserving realistic microstructural overlap and orientation statistics. We created multiple micro‐geometries composed of identical cylinders of 1μm in diameter with varying volume fractions fc=0,0.1,0.2,0.3,0.4,0.5 and identical spheres of either 10μm or 2μm in diameter with varying volume fractions fs=0,0.1,0.2,0.3,0.4,0.5, leading to 46 different micro‐geometries (Figure 2A), excluding the cases of fc=fs=0 and fc=fs=0.5. (A) Gray matter‐mimicking micro‐geometry composed of randomly packed spheres and randomly oriented cylinders with intra‐cylindrical and intra‐spherical volume fractions (fc, fs) and sphere diameter ds. (B) A 3d SBEM dataset of extra‐cellular‐space‐preserved tissue, sampled from a mouse olfactory bulb [35] (Section 3.2). After generating micro‐geometries in a 3‐dimensional continuous space enclosed within the cube of side length L=30μm, we pixelated each geometry into a 600×600×600 matrix with an isotropic pixel size of 50 nm. The fc and fs were calculated again by counting the number of pixels in each compartment. The relative error of fc (fs) was less than 3.5% (0.005%). The relative error of sphere diameter was about 1% or less. Mouse Electron Microscopy To demonstrate our theory in realistic extra‐cellular space of brain gray matter, we analyzed a publicly available three‐dimensional serial block‐face scanning electron microscopy (SBEM) dataset of extra‐cellular‐space‐preserved tissue, sampled from a mouse olfactory bulb (Figure 2B) [35]. The original sample had a volume of 10×10×12.8μm3 with a voxel resolution of 9.8×9.8×25nm3. We segmented all neurites in the sample using DeepACSON [45, 46], followed by instance segmentation initiated from the neurite skeletons provided in the original dataset. We considered the space outside segmented neurites as the extra‐cellular space. The binary mask of the extra‐cellular space was smoothed by applying a signed distance transform, Gaussian smoothing with a width of 0.72 voxel, and thresholding at 0. We cropped the edge of each slice by 250 nm, where segmentation was relatively challenging and less precise, resulting in a final volume of 9.53×9.53×12.8μm3. For diffusion simulations, we downsampled the extra‐cellular binary mask to an isotropic voxel resolution of 25×25×25nm3. The extra‐cellular fraction was fe=0.353. The neurites were cylinder‐like structures and had a fraction of fc=0.647. The sample contained no cell bodies [35], such that sphere‐like structures had a fraction fs= 0. Monte Carlo Simulations We performed Monte Carlo (MC) simulations in (i) pixelated micro‐geometries composed of impermeable spheres and cylinders, created in Section 3.1, and (ii) a binary mask of realistic extra‐cellular space from mouse SBEM (Section 3.2) using the RMS toolbox [47, 48, 49, 50] implemented in CUDA C++ for diffusion in extra‐cellular space. In each simulation, we applied 1×105 random walkers, each diffusing with a step size of δx=45nm (gray matter‐mimicking substrates) and 22.5 nm (mouse SBEM), a time step of δt=δx2/(6D0), and an intrinsic diffusivity D0= 1, 2, or 3μm2/ms in extra‐cellular space. We calculated the apparent diffusivity based on the displacement cumulant and averaged them over 100 directions to calculate the mean diffusivity (MD) at diffusion times t=1–100 ms. To estimate the extra‐cellular diffusivity D in the long time limit, we fitted the following functional form to MD(t) to factor out the diffusivity time‐dependence [51]: (18) MD(t)=D+ct, where c is the strength of hindrance due to the structural disorder. (Note that, technically, one should utilize the ln(t/tc)/t tail in Equation (18) arising from random‐rods disorder with correlation time tc [51]. However, within the time range considered here, the simpler 1/t form yielded a numerically similar tortuosity limit D due to the slowness of lnt, while also providing a more robust fit with fewer free parameters.) We calculated the extra‐cellular tortuosity (1) using the simulated D and compared its values with the mixture solution (9) and (10), sequential (13) and simultaneous coarse‐graining solutions (14). To evaluate the effect of diffusivity time‐dependence of wide pulse‐gradient sequences, we calculated diffusion signals based on diffusional phase in each simulation at four b values b=0.5,1,1.5,2 ms/μ m in 20 directions per b‐shell with inter‐pulse interval Δ and pulse width δ at (Δ, δ)=(13,6), (19, 8), and (43.4, 13.7) ms, corresponding to protocols on Connectome 2.0 scanner (500 mT/m maximum gradient strength) [52], Connectome 1.0 scanner (300 mT/m) [53], and clinical scanner (HCP Lifespan, 80 mT/m) [54, 55]. We computed the extra‐cellular MD(Δ,δ) from simulated signals and compared its values with the simultaneous coarse‐graining solution (14). Furthermore, we evaluated the effect of diffusivity frequency‐dependence of trapezoidal‐cosine oscillating‐gradient sequences (Supporting Information S2, Figure S1) [56, 57, 58, 59, 60]. We calculated diffusion signals based on diffusional phase in each simulation at four b values b=0.25,0.5,0.75, 1 ms/μ m in 20 directions per b‐shell with total waveform time of 92.3 ms, mixing time of 10.5 ms, rise time of 0.9 ms, number of cycles over each side of refocusing pulse N=1 and 2 for nominal frequency ω/2π=25 and 50 Hz, respectively, corresponding to the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Similarly, we computed the extra‐cellular MD(ω) from simulated signals and compared its values with the simultaneous coarse‐graining solution (14). Note that the IMPULSED protocol refers to the biophysical model for cell size imaging [58], rather than the high‐gradient IMPULSE system [61]. In Vivo MRI Seven healthy subjects (27–72 years; 37.1±18.4 years; four females) from a publicly available dataset were included [53]. Data were scanned on the 3T Connectome 1.0 scanner (MAGNETOM CONNECTOM, Siemens Healthineers, Germany; 300 mT/m maximum gradient strength) using a 64‐channel head coil [62]. Diffusion‐weighted images were obtained using pulsed‐gradient spin‐echo echo‐planar‐imaging sequence (Δ=19, δ=8 ms) with eight b values (b=0.05 to 6 ms/μm). Thirty‐two gradient directions per b‐shell were used for b<2.4 ms/μm and 64 otherwise. Imaging parameters included TR = 4000 ms, TE = 77 ms, 2 mm isotropic resolution, partial Fourier = 6/8, GRAPPA = 2, and SMS = 2. Parameter Estimation To demonstrate in vivo applications of the proposed extra‐cellular tortuosity relation, we introduced a constrained version of soma and neurite density imaging (SANDI) model [16] to reduce the number of degrees of freedom and related data requirements. The standard SANDI, employing spherical means per shell, requires at least five b values to estimate five parameters: intra‐soma fraction and radius, intra‐neurite fraction and diffusivity, and extra‐cellular diffusivity. In the constrained SANDI, intra‐neurite diffusivity was fixed at 1.85 μm2/ms [5, 63], and extra‐cellular diffusivity was expressed as a function of intra‐soma and intra‐neurite fractions using the tortuosity relation (14), with extra‐cellular intrinsic diffusivity D0 fixed at 3μm2/ms. Intra‐soma and intra‐neurite fractions were fitted as free parameters, while soma radius was treated as a free hyper‐parameter without biological interpretation. For both models, intra‐soma intrinsic diffusivity was fixed at 2μm2/ms to stabilize fitting with random forest regression [64]. The standard SANDI was fitted to the full eight b‐shell data, serving as the reference, while the constrained SANDI was fitted to a subset of the same data using only two b‐shells (0.8 and 3.45 ms/μm). Group‐averaged parametric maps were generated in MNI space using nonlinear registration for qualitative analysis. Application of the constrained (clinical) SANDI on clinical scanners will be investigated in future work. Results Monte Carlo Simulations In diffusion simulations (Figure 3, Figures S2–S6), the mixture solution (9) and (10) and sequential coarse‐graining solution (13) predicted the simulated extra‐cellular tortuosity (1) well when the overall intra‐cellular volume fraction was low (fc+fs≤0.3), whereas the theory deviated from simulations for high intra‐cellular volume fractions (fc+fs>0.6). In contrast, the simultaneous coarse‐graining solution (14) predicted the simulated extra‐cellular tortuosity (1) well not just for low intra‐cellular volume fractions but also the high ones. This observation was independent of the intrinsic diffusivity D0=1−3μm2/ms and sphere diameters ds=10 and 2 μm tested in the simulations. In the realistic extra‐cellular space derived from mouse SBEM data, all solutions performed well. Comparison of (A) mixture solution (9) and (10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=3μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1 μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. For wide pulsed‐gradient and oscillating‐gradient waveforms (Figure 4, Figures S7–S11), the simultaneous coarse‐graining solution (14) predicted the extra‐cellular MD(Δ,δ) (for pulsed‐gradient) and MD(ω) (for oscillating‐gradient) quite well across low to high intra‐cellular volume fractions. For both wide pulsed‐gradient and oscillating‐gradient waveforms, we observed slightly higher deviations between the proposed solution (14) and simulations with lower intrinsic diffusivity and larger sphere diameter (Figure S7). In addition, the solution performed well in the realistic extra‐cellular space derived from mouse SBEM data. The simulated diffusivity time‐dependence across the three diffusion protocols with varying (Δ, δ) was less than 2%; the simulated diffusivity frequency‐dependence across the two frequencies ω was less than 1.5%. Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=3μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1 μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent MD(Δ,δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54, 55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we are showing effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse has been inverted to account for the refocusing effect. In Vivo MRI To demonstrate the applicability of the proposed tortuosity relation (14) in vivo, we introduced a constrained SANDI with reduced data requirements. The standard SANDI was fitted to an eight‐shell data as a reference, while the constrained model was applied to a two‐shell subset of the same data to estimate intra‐soma and intra‐neurite fractions. Both approaches yielded comparable values of tissue parameters and regional contrast across both white and gray matter (Figure 5). In vivo application of the tortuosity relation in soma and neurite density imaging (SANDI). (Top) The standard SANDI was fitted to diffusion data acquired with eight b‐shells (0.05–6 ms/μm2) in seven subjects on Connectome 1.0 scanner (300 mT/m maximum gradient strength), serving as the reference. (Bottom) A constrained SANDI was fitted to a subset of the same data using only two b‐shells (0.8 and 3.45 ms/μm2). The constraints included fixing intra‐neurite diffusivity to 1.85 μm2/ms and expressing the extra‐cellular diffusivity in terms of the intra‐soma and intra‐neurite fractions (corresponding to fs and fc, respectively) via the tortuosity relation (14). Discussion In this study, we proposed EMT solutions for the extra‐cellular tortuosity relation in a gray matter‐mimicking medium consisting of densely packed spheres and randomly oriented cylinders, including the mixture solution (9) and (10), the sequential coarse‐graining solution (12) and (13), and the simultaneous coarse‐graining solution (14), and evaluated these solutions using Monte Carlo simulations of diffusion. The simulations were performed in both idealized substrates and a realistic extra‐cellular space derived from SBEM data, although the latter did not contain somas and therefore did not fully represent the mixed soma‐neurite architecture of cortical gray matter. As a proof‐of‐concept application, the resulting tortuosity relations were incorporated into biophysical modeling of dMRI and microstructural MRI in brain gray matter on a Connectome scanner [16, 65]. While the in vivo demonstration was performed in a small cohort of healthy participants and did not constitute biological validation, it explained how the proposed tortuosity relation may reduce the number of independent fitting parameters. Furthermore, the relation may facilitate protocol optimization to shorten scan times on clinical MRI scanners for clinical translation [36, 66] and on preclinical scanners for high‐resolution microstructure imaging [52, 61, 67, 68]. The current framework is based on EMT, which assumes weak spatial correlations and homogeneous mixing of grains in a medium. In tissues with densely packed cells, spatial correlations and the connectivity of the extra‐cellular space may play an important role, since transport becomes increasingly governed by constrained pathways. Such correlations are not explicitly considered in classical EMT and may lead to deviations, especially at high intra‐cellular fractions. However, we observe that the proposed simultaneous coarse‐graining solution (14) remains accurate across densely packed geometries. This suggests that, though empirical, the proposed theory may effectively capture the combined influence of volume exclusion and connectivity constraints. We also note a generally better agreement of the mean‐field‐like approach of the present 3d setup with MC simulations than in the previously considered 2d problem [34], which required modeling transport via narrow gorges in‐between disks, as well as employing models with the periodic geometry for the packed smaller disks, to reach an agreement with the simulations. Better performance of mean‐field approaches is generally expected in higher spatial dimensions, as the role of fluctuations decreases. In addition, correlation functions of tissue microstructure have been shown to characterize diffusivity time‐dependence [69] and diffusion‐diffraction signal patterns [70] in heterogeneous media and may provide complementary information for relating tissue microstructure to extra‐cellular transport beyond the long‐time tortuosity limit. Future work will explore the applicability and generalization of these relations to more complex geometries, such as realistic cellular structures observed on microscopic imaging of human brain tissue specimens prepared to preserve the extra‐cellular space [35], as well as the interconnected extracellular network spanned by tunnels around synapses, and axons and sheets around astrocytes [71]. Following limitations or simplifying assumptions should be acknowledged. Our validation mainly focused on diffusion simulations of the extra‐cellular diffusivity in the long‐time limit, which remains relevant to many practical scenarios. In this limit, the measured diffusivity becomes independent of the specific gradient waveform [6], and clinical scanners typically operate with long effective diffusion times of ∼30 ms or longer. In contrast, shorter diffusion times are more accessible on preclinical systems, high‐performance systems [52, 67, 72], and/or employing oscillating‐gradient waveforms [73], where residual time‐ or frequency‐dependence, although small, may introduce biases. Extending our current framework to numerically model these effects across scanner platforms and waveform protocols will be investigated in future work. Furthermore, our theoretical derivations ignored the effect of structural disorder caused by neurite branching, beading, undulation (micro‐orientational dispersion), and glial processes on the diffusivity tortuosity in the extra‐cellular space. These microstructural features may affect not only the intra‐neurite diffusivity and its time‐dependence [33, 48, 51, 74, 75, 76, 77, 78, 79] but also the coarse‐grained extra‐cellular diffusivity at long times [80, 81]. The proposed extra‐cellular tortuosity relation may change in pathological conditions if the tissue microstructure is altered sufficiently strongly, such that our EMT approach fails. The present validation is based on idealized substrates and a single realistic micro‐geometry, and extending the analysis to a broader range of realistic substrates is an important direction for future work. Finally, we assumed impermeable cell membranes and neglected the transcytolemmal water exchange [82, 83]. In biological tissue, membrane permeability is heterogeneous and cell‐type dependent [84], which may lead to deviations from the proposed tortuosity relation. The impermeable assumption is expected to be appropriate when the characteristic exchange time is long compared to the diffusion time, such that water molecules predominantly remain within their compartments during the encoding period. In contrast, for faster exchange regimes or for encoding schemes with enhanced exchange sensitivity [85, 86, 87, 88, 89], deviations from the proposed relations are expected. Incorporating exchange within the present framework, for example, via frequency‐domain solution [90] of the Kärger exchange model [85, 86, 87, 88, 91], may account for these effects and their interplay with structural disorder, and will be examined in future studies. Conclusions We have proposed and validated analytical solutions: A mixture, Equations (9) and (10); sequential coarse‐graining (12) and (13); and simultaneous coarse‐graining (14), for extra‐cellular tortuosity of a medium composed of randomly packed spheres and randomly oriented cylinders, mimicking microgeometry of the gray matter, and validated the theory in Monte Carlo simulations of diffusion with a wide range of tissue parameters. Surprisingly, the least justifiable simultaneous coarse‐graining solution (14) performs the best. The proposed analytical approach helps simplify biophysical models of dMRI in gray matter, facilitating diffusion protocol optimization on clinical scanners and the broader translation of microstructure imaging to the neurological and psychiatric disorders. Funding This work was supported by National Institute of Biomedical Imaging and Bioengineering (Grant Nos. P41EB015896, P41EB017183, P41EB030006, U01EB026996, UH3EB034875), the National Research Foundation of Korea (Grant No. RS‐2024‐00411768), the Academy of Finland (Grant No. 360360), Office of the Director and National Institute of Dental and Craniofacial Research (Grant No. DP5OD031854), and the National Institute of Neurological Disorders and Stroke (Grant Nos. R01NS088040, R01NS118187, R21NS081230, U24NS137077), National Institute on Aging (Grant No. R21AG085795). Supporting information Section S1: Estimation of Cylinder Number in Overlapping Configurations. Section S2: Oscillating‐gradient Waveform. Figure S1: Oscillating‐gradient waveform used in diffusion simulations. The trapezoidal‐cosine waveform can be parametrized by four quantities: the inter‐waveform interval Δ, waveform duration δ, number of oscillations N, and rise time tr. Alternatively, it can be equivalently described by the total waveform time T, mixing time tm, N, and tr. The waveform amplitude is scaled by the Larmor gradient strength g. Additional parameters, such as the first plateau time p1 and the second plateau time p, are then determined accordingly. Here, the actual gradient waveform applied on the scanner is shown, without accounting for the refocusing effect of the 180° pulse. Figure S2: Comparison of (A) mixture solution (9)–(10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=1μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. Figure S3: Comparison of (A) mixture solution (9)–(10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=2μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. Figure S4: Comparison of (A) mixture solution (9)–(10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=1μm2/ms in a medium composed of spheres of ds=2μm in diameter and cylinders of 1μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. Figure S5: Comparison of (A) mixture solution (9)–(10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=2μm2/ms in a medium composed of spheres of ds=2μm in diameter and cylinders of 1μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. figure S6: Comparison of (A) mixture solution (9)–(10), (B) sequential (13), and (C) simultaneous coarse‐graining solutions (14) for extra‐cellular tortuosity with simulation results of intrinsic diffusivity D0=3μm2/ms in a medium composed of spheres of ds=3μm in diameter and cylinders of 1μm in diameter. In left panels, simulation results in extra‐cellular space of mouse SBEM are shown in pentagrams. In right panels, simulation results are shown as data points, and theoretical predictions are shown as curves. Figure S7: Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=1μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent mean diffusivity MD(Δ, δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54, 55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent mean diffusivity MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we showed effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse was inverted to account for the refocusing effect. Figure S8: Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=2μm2/ms in a medium composed of spheres of ds=10μm in diameter and cylinders of 1μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent mean diffusivity MD(Δ, δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54, 55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent mean diffusivity MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we showed effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse was inverted to account for the refocusing effect. Figure S9: Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=1μm2/ms in a medium composed of spheres of ds=2μm in diameter and cylinders of 1μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent mean diffusivity MD(Δ, δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54, 55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent mean diffusivity MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we showed effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse was inverted to account for the refocusing effect. Figure S10: Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=2μm2/ms in a medium composed of spheres of ds=2μm in diameter and cylinders of 1μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent mean diffusivity MD(Δ, δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54, 55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent mean diffusivity MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we showed effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse was inverted to account for the refocusing effect. Figure S11: Comparison of simultaneous coarse‐graining solution (14) for extra‐cellular tortuosity with simulation results of (A) wide pulsed‐gradient and (B) oscillating‐gradient sequences. We performed diffusion simulations with intrinsic diffusivity D0=3μm2/ms in a medium composed of spheres of ds=2μm in diameter and cylinders of 1μm in diameter. For pulsed‐gradient waveform, we simulated the time‐dependent mean diffusivity MD(Δ, δ) for diffusion protocols (inter‐pulse interval Δ, pulse width δ) on three different scanners, such as Connectome 2.0 (500 mT/m) [52], Connectome 1.0 (300 mT/m) [53], and clinical scanners (80 mT/m) [54,55]. For oscillating‐gradient waveform, we simulated the frequency‐dependent mean diffusivity MD(ω) for the IMPULSED protocol on clinical scanners (80 mT/m) [58]. Here, we showed effective gradient waveforms of oscillating‐gradient sequences, where the gradient polarity after the 180° pulse was inverted to account for the refocusing effect.