betaFeedback

WorksPhysics and Astronomy

Unmyelinated axons alter the biological interpretation of MRI biomarkers of demyelination

Abstract

Quantitative MRI methods are widely used to probe white matter microstructure; yet diffusion- and relaxation-based metrics are often interpreted primarily in terms of myelinated axons, implicitly assuming a negligible contribution from unmyelinated axons. Here, combining diffusion MRI (dMRI), relaxation-based MRI, and electron microscopy (EM) in the cuprizone mouse model of demyelination and recovery, we examine how unmyelinated axons shape MRI-derived microstructural indices, including axonal water fraction and gg-ratio. To enable quantitative cellular-level comparisons, we developed automated tools for instance-level EM segmentation of white matter, allowing morphological analysis of myelinated and unmyelinated axonal populations. We find that, in the mouse model, diffusion-derived estimates of axonal water fraction reflect contributions from both axonal populations at sufficiently short diffusion times; accounting for unmyelinated axons removes an apparent bias and reveals that this metric corresponds to the total intra-axonal space rather than myelinated fibers alone. Similarly, MRI-derived gg-ratio estimates correlate strongly with EM only when both axonal populations are included. We further demonstrate that within-voxel gg-ratio heterogeneity introduces a systematic negative bias of up to 8%8\%, increasing with the fraction of unmyelinated axons and variability in myelin thickness. Together, these findings change the biological interpretation of MRI biomarkers during demyelination by demonstrating that unmyelinated axons are a major determinant of MRI-derived white matter indices and that combining diffusion and relaxation contrasts enables a more accurate and interpretable quantification of demyelination. They further highlight the need for quantitative biomarkers that account for axonal heterogeneity in both health and disease.

Introduction

Quantitative MRI is widely used to characterize white matter microstructure in vivo, with particular relevance to demyelination and remyelination, where changes in myelin content and axonal integrity are central to disease progression and recovery[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]. Measures derived from diffusion MRI (dMRI)[4, 8, 10, 11, 12], transverse relaxation[1, 5, 7], and magnetization transfer[6, 12, 13, 14] are commonly interpreted in terms of myelin loss, axonal density, or gg-ratio, and have been extensively applied and validated in experimental models and human disease. These interpretations, however, implicitly rely on assumptions about the underlying axonal population that may not hold when myelin is disrupted.

Demyelinating insults, such as cuprizone exposure, cause oligodendrocyte death and myelin degeneration[15, 16, 17], altering the myelination status of previously myelinated axons and increasing the proportion of “effectively unmyelinated” axon segments in affected tracts. Whether this expansion reflects selective sparing of intrinsically unmyelinated axons[15, 17] or axons with changed status[18, 19], this shift in the statistical composition of axonal populations fundamentally changes the biophysical environment probed by MRI. Yet, most quantitative MRI models implicitly assume a homogeneous population of uniformly myelinated axons, overlooking this source of variability despite the extensive development of myelin-sensitive biomarkers[4, 5, 6, 8, 11, 20].

A particularly important implication of axonal heterogeneity in demyelinating tissue concerns the interpretation of the gg-ratio. By definition, the gg-ratio relates the inner and outer diameters of a myelinated axon[21, 22] and is therefore formally defined only for myelinated fibers, in both histological and MRI-based formulations[20, 23, 24, 25]. In the presence of demyelination, where unmyelinated axons constitute an increasing fraction of the axonal population[15, 17, 18, 19], voxel- or region-averaged gg-ratio estimates no longer reflect per-axon myelin thickness alone but are influenced by changes in axonal composition. How this affects commonly used gg-ratio measures during demyelination and remyelination is not sufficiently understood.

The cuprizone mouse model provides a highly reproducible system for studying demyelination and recovery[15, 16] and testing the sensitivity and specificity of candidate biomarkers to myelin alterations. Cuprizone exposure causes selective oligodendrocyte loss and progressive demyelination, predominantly in the corpus callosum[26], followed by partial remyelination during continued intoxication and a second demyelination phase that emerges upon prolonged exposure. Withdrawal of cuprizone after either acute or chronic treatment causes remyelination. tion within several weeks[15, 26, 27]. We hypothesize that demyelination-induced changes in the relative abundance of myelinated and unmyelinated axons systematically alter the biological interpretation of quantitative MRI measures of axonal and myelin integrity.

Validating quantitative MRI metrics requires direct comparison with histology[4, 8, 11, 28, 29, 30, 31]. Sub-10-nanometer resolution electron microscopy (EM) enables resolving myelin and single-axon geometry in both myelinated and unmyelinated axons and quantifying their respective volume fractions[32, 33]. However, model validation has largely focused on myelinated axons[22, 34, 35, 36], as automated instance segmentation of unmyelinated axons remains challenging due to their thin single-membrane boundaries and low contrast. Recent methodological advances have substantially expanded this landscape. dMRI models that explicitly describe axonal compartments[9, 10, 37, 38, 39, 40] and orientation dispersion[9, 10, 36, 41], together with improved estimation strategies[39, 42] offer increased sensitivity to axonal composition. In parallel, recent developments in large-scale 3D EM acquisition[43] and automated instance segmentation[43, 44, 45, 46] now enable a precise and comprehensive quantification of both axonal populations.

Here we reanalyze quantitative MRI and transmission EM (TEM) microstructural data acquired in a cuprizone model[8] with a distinct objective from previous work. Rather than validating MRI-derived biomarkers of demyelination[5, 6, 8, 11, 20, 23], we examine how demyelination-induced changes in axonal composition — particularly shifts toward unmyelinated axons[15, 17, 18, 19] — shape the interpretation of voxel-averaged MRI microstructure metrics, including gg-ratio[20, 24, 25].

We show that the dMRI-derived axonal water fraction ff of the Standard Model (SM) of diffusion in white matter[9, 10, 39] measured at short diffusion times, where unmyelinated axons remain effectively impermeable, reflects the total intra-axonal volume rather than myelinated axons alone. Accounting for the combined contribution of myelinated and unmyelinated axons removes the apparent bias between ff and EM-derived axonal volume fractions[22, 32]. On the other hand, the EM-derived myelin volume fraction strongly correlates with MRI measures sensitive to myelin content, including R2R_2 and magnetization transfer measures. Combining these myelin-sensitive MRI measures with the diffusion-derived axonal water fraction ff yields an MRI-derived gg-ratio that agrees with its EM-derived counterpart only when the contribution of unmyelinated axons is explicitly accounted for. Finally, we derive how intra-voxel gg-ratio variability propagates through nonlinear averaging, leading to a systematic negative bias in effective gg-ratio estimates. Together, these findings establish voxel-level heterogeneity, driven largely by unmyelinated axons, as a key determinant of MRI-derived measures of white matter integrity.

Materials and methods

Animal model and imaging

Animal model, MR imaging, and EM preparation are detailed in Ref.[8] and are briefly summarized here.

Animals: An 18-week longitudinal study was conducted on 36 female C57BL/6 mice (8 weeks old at baseline; Taconic Biosciences, Hudson, NY), assigned equally to three diet groups of 12 animals. The control group (Ctl) received standard chow throughout the experiment. The Cpz6 and Cpz12 groups were fed chow supplemented with 0.2% cuprizone for 6 and 12 weeks, respectively, followed by a 6-week recovery period on standard chow. MR imaging was performed in vivo at baseline, week 6, week 12, and week 18. Following imaging, subsets of animals from each group were sacrificed at each time point (6, 12, and 18 weeks) for TEM imaging. Three mice from the Ctl group died during scanning, two at baseline and one at six weeks, so they were excluded.

All mice were treated in strict accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals, and the experimental procedures were performed following the Institutional Animal Care and Use Committee at the New York University School of Medicine.

MR imaging: In vivo MRI data of mice under anesthesia with isoflurane were acquired on a 7 T Bruker. MR imaging was restricted to a single 0.8-mm-thick midsagittal slice, which can be positioned reproducibly across animals and intersects callosal axons predominantly orthogonal to their main orientation. The slice was acquired at 112-μm in-plane resolution.

dMRI data were acquired along with T2T_2-weighted and magnetization transfer (MT) imaging. Diffusion-weighted images were acquired using a 2D GraSE (Gradient and Spin Echo) sequence[47, 48] with bb-values of 1 and 2 ms/μm2^2 and 30 directions each, in addition to six b=0b=0 images (TR=800T_R=800 ms; TE=49T_E=49 ms; Δ=16\Delta=16 ms; δ=2\delta=2 ms). T2T_2-weighted images were obtained using a multi-echo spin-echo sequence (32 echoes; TET_E spacing = 7.1 ms; TR=2T_R=2 s; NA=2N_A=2). Magnetization transfer ratio (MTR) maps were derived from two FLASH acquisitions (TR=47T_R=47 ms; TE=5.5T_E=5.5 ms; flip angle = 8°; NA=40N_A=40) with and without an MT saturation pulse (Gaussian shape; frequency offset = 1.5 kHz; duration = 12 ms; peak power = 4.7 μT).

Electron microscopy: Animals were anesthetized and perfused intracardially with heparinized phosphate-buffered saline, followed by fixation with 2% formaldehyde and 2.5% glutaraldehyde in phosphate buffer. Brains were extracted, and the splenium of the corpus callosum was dissected. Tissue blocks were post-fixed in the same fixative, stained with osmium tetroxide, dehydrated, and embedded in resin. Ultrathin sections were cut using an ultramicrotome and imaged with a Phillips CM12 TEM equipped with a Gatan digital camera, yielding electron micrographs of approximately 6×96 \times9 µm2^2 at 2.36×2.362.36 \times2.36 nm2^2 resolution.

MR image processing

To generate parametric maps, both the T2T_2-weighted images and the dMRI data were processed following the steps in the DESIGNER pipeline (https://nyu-diffusionmri.github.io/DESIGNER-v2/)[49]. Complex-valued images were denoised using MPPCA-Complex[50, 51], corrected for Gibbs ringing[52], and registration[53] was used to correct geometric distortions induced by eddy currents as well as motion.

T2T_2 mapping was performed to probe more specifically the T2T_2 of myelin and given the multi-compartmental nature of T2T_2 relaxation, we estimate R2=1/T2R_2 = 1/T_2 as the slope at TE→0T_E \to0 of the log-signal using a second-order cumulant expansion in TET_E, ln⁡S(TE)≃ln⁡S0−R2TE+c⋅TE2\ln S(T_E) \simeq\ln S_0 - R_2T_E + c \cdot T_E^2, using weighted linear least squares fitting. To reduce the impact of stimulated echoes caused by imperfect refocusing pulses, we used only the even echoes from the T2T_2-weighted images, ranging from 14.168 ms to 226.688 ms in steps of 14.168 ms (excluding the last image due to low SNR).

For MTR mapping, the MTon_{\rm on} image was registered to the MToff_{\rm off} image. MTR was then computed voxel-wise as: MTR = (MToff_{\rm off} − MTon_{\rm on}) / MToff_{\rm off} × 100%, where MToff_{\rm off} is the signal without the MT pulse and MTon_{\rm on} is the signal with the pulse.

To generate the SM maps from dMRI, we first identified outliers[54] by the goodness-of-fit for diffusion kurtosis imaging (DKI) and replaced them with signals predicted from the DKI fit. The SM parameters are the axonal water fraction ff, intra-axonal axial diffusivity DaD_a, extra-axonal parallel and perpendicular diffusivities De∥D_e^{\parallel} and De⊥D_e^{\perp}, and the free water compartment fraction fwf_w. The SM parameters were then estimated following Ref[39], with axial diffusivities DaD_a and De∥D_e^{\parallel} constrained by an upper bound corresponding to the free diffusivity of water 3 µm2^2/ms at body temperature, while half this value was used to constrain the upper value of the radial extra-axonal diffusivity De⊥D_e^{\perp}. Rotational invariants[9] up to ℓ=2\ell= 2 were used for SM parameter estimation, yielding the second-order rotational invariant p2p_2 of the fiber orientation dispersion, e.g., due to crossing fibers in a voxel.

The splenium of the corpus callosum was selected as the region of interest (ROI) as it exhibits robust and reproducible cuprizone-induced demyelination and rapid remyelination[18, 55], and encompasses the anatomical location sampled for 2D TEM. To define the ROI, we first manually drew a bounding box on the red channel (left-right orientation) of the coronal FA map to separate the splenium from the body of the corpus callosum. Within this bounding box, the splenium was delineated automatically using a threshold of 0.8 times the maximum value, and the largest connected component of the resulting binary mask was retained. The corresponding mask was registered to T2T_2 and MTR spaces via the first b=0b = 0 diffusion image and the T2T_2-weighted image with similar TET_E as the dMRI data, with registration verified visually. Mean and standard deviation of all MR-derived metrics were computed within the registered splenium ROI for each animal and time point.

2D EM microstructural segmentation

We developed a DeepACSON-based[56, 57] segmentation pipeline to automatically segment high-resolution TEM images of white matter into myelin, myelinated axons, unmyelinated axons, and extra-axonal space, assigning a unique label to each compartment (https://github.com/ComputationalNeuroanatomy). The pipeline consists of three main steps, as shown in Figure 1: (i) incremental expansion of a training set; (ii) training a UNet[58] for semantic segmentation of white matter microstructure; and (iii) their instance segmentation using a watershed-based method.

Segmentation of white matter microstructure

Figure 1. Segmentation of white matter microstructure. A standard 2D UNet was trained using a small set of manually labeled images until the network effectively memorized the training data. The trained model was then applied to unseen TEM images, and an expert corrected only the segmentation errors. The corrected image-label pairs were added to the training data, and the network was fine-tuned from its previous state. This iterative procedure was repeated, progressively expanding the training set while minimizing manual annotation. The final model was used for performing instance-level segmentation of white matter microstructure.

Training a UNet[58] on a limited dataset poses challenges, yet modern architectures can fully “memorize” even small datasets when trained with standard stochastic gradient descent[59]. We exploited this property as a bootstrapping mechanism to generate a high-quality training set from minimal annotations. We began with a single manually labeled TEM image (512 × 512 pixels), in which neurite boundaries and myelin were annotated. This image was augmented using randomized histogram distortions to ensure invariance to contrast and brightness gradients, as well as by applying affine transformations, random perspective warping, and horizontal/vertical flips.

The initial UNet[58] was trained to fully fit the augmented training set (Dice coefficient ≈ 1). This overfitted network is not intended as a final predictor, but rather to generate initial segmentations for new, unlabeled images. Predictions were manually corrected only where necessary—typically in regions whose structural appearance deviated from the original training example—and the corrected images were added to the training set. This iterative refinement procedure progressively increased the diversity and coverage of the training data while minimizing the manual annotation effort. As the training dataset expanded, the network generalized to increasingly varied tissue appearances, requiring progressively less manual correction.

The model was trained using the cross-entropy loss with Softmax normalization. We set the batch size equal to the number of images in the training set. The training images were standardized to a mean equal to 0.456 and a standard deviation equal to 0.224. For the optimization, we used Adam optimizer[60] and set its initial learning rate α=1×10−4\alpha= 1 \times10^{-4}, the exponential decay rate for the first moment β1=0.9\beta_1 = 0.9, the exponential decay rate for the second-moment β2=0.999\beta_2 = 0.999, and the weight decay =1×10−5= 1 \times10^{-5} (we use the same notation as in ref[60]). We utilized the trained UNet in its final state for the semantic segmentation of TEM images into their microstructural compartments. Finally, we applied a watershed segmentation for the instance segmentation of microstructural compartments.

Statistical analysis

For each histological measure, differences between diet groups were assessed separately at each time point using a linear mixed-effects model, with mouse as a random effect. For comparisons involving more than two groups, pairwise differences were tested using model-based contrasts, and adjusted pp-values were computed using the Holm–Bonferroni correction. Effect sizes were estimated using Cohen’s dd, noted as deffd_{\mathrm{eff}}.

Correlations between EM- and MRI-derived metrics were assessed using weighted, nonparametric Spearman partial correlation analysis, accounting for mouse body weights, as cuprizone-treated mice exhibited reduced weight gain compared to controls. To characterize linear relationships between modalities, least-squares linear regression was performed between corresponding EM and MRI measures, and the resulting slope and intercept were used to assess proportionality and systematic bias. All statistical tests were conducted at the two-sided 5% significance level.

g-Ratio definitions for heterogeneous axonal populations

EM-derived microstructural metrics

The instance-level segmentation of white matter microstructure enables direct access to axonal diameters for individual myelinated and unmyelinated axons, denoted as dmAd_{\mathrm{mA}} and duAd_{\mathrm{uA}}. Our compartmental volumetric analysis includes the myelin volume fraction ϕm\phi_{\mathrm{m}}, the myelinated ϕmA\phi_{\mathrm{mA}} and unmyelinated axon volume fraction ϕuA\phi_{\mathrm{uA}}, and the extra-axonal space volume fraction ϕex\phi_{\mathrm{ex}}, which are

mutually exclusive and ϕm+ϕmA+ϕuA+ϕex=1\phi_{\mathrm{m}} + \phi_{\mathrm{mA}} + \phi_{\mathrm{uA}} + \phi_{\mathrm{ex}} = 1. As these measurements were derived from 2D TEM images, area fractions were used as unbiased estimators of the corresponding volume fractions. We denote the total axonal volume fraction as ϕtA=ϕmA+ϕuA\phi_{\mathrm{tA}} = \phi_{\mathrm{mA}} + \phi_{\mathrm{uA}} (notations are summarized in Table I). When comparing EM-derived axonal volume fractions with the SM-derived axonal water fraction ff, we normalized ϕmA\phi_{\mathrm{mA}}, ϕuA\phi_{\mathrm{uA}}, and ϕtA\phi_{\mathrm{tA}} by 1/(1−ϕm)1/(1 - \phi_{\mathrm{m}}) because ff reflects only “MR visible” water and does not account for the myelin due to the fast T2T_2 relaxation of myelin water. This normalization implicitly assumes comparable transverse relaxation properties across intra- and extra-axonal water pools. The differences in T2T_2 relaxation between intra- and extra-axonal compartments[61, 62] — particularly under demyelinating conditions that we associated with the expansion of the extra-axonal space — are not explicitly accounted for here and may cause slight deviations from linearity in the Cpz12-dMRI correlations.

g-ratio: definitions and generalizations

Originally, the gg-ratio has been defined for individual myelinated axons only as the ratio g=r/Rg = r/R of the inner to outer radii, implicitly assuming a perfectly circular axon cross-section[63]. One can view it as a measure of the thickness lm=R−r=(1/g−1)rl_{\mathrm{m}} = R-r = (1/g-1)r of the myelin sheath, consisting of many wrappings of a lipid bilayer.

For non-circular axon cross-sections, both inner r(θ)r(\theta) and outer R(θ)R(\theta) radii would depend on the polar angle 0≤θ<2π0 \leq\theta< 2\pi, even if the origin is taken to be at the geometric “center of mass”. To characterize myelin, one can either focus on the thickness lm(θ)=R(θ)−r(θ)l_{\mathrm{m}}(\theta) = R(\theta)-r(\theta) or on the ratio g(θ)=r(θ)/R(θ)g(\theta) = r(\theta)/R(\theta); it is rational to pick the metric that depends on θ\theta the least. Here we assume that for the ii-th axon, the gg-ratio gig_i is approximately independent of θ\theta (outer and inner boundaries are geometrically similar), thus the ratio between the inner and outer cross-sectional areas, Ain,iA_{\mathrm{in},i} and Aout,iA_{\mathrm{out},i}, satisfies

Ain,i=∫02πdθ 12ri2(θ)=gi2Aout,iA_{\mathrm{in},i} = \int_{0}^{2\pi} \mathrm{d}\theta\,\frac{1}{2}r_i^2(\theta) = g_i^2 A_{\mathrm{out},i}

independently of the cross-sectional shape.

In EM histology, where individual axons are resolved, we can measure this ratio gig_i at the single-axon level (top panel of Figure 2). We define the mean gg-ratio of myelinated axons within each histological section as ⟨gi∈mA⟩\langle g_{i\in\mathrm{mA}}\rangle; its average across sections for a given animal is ⟨gi⟩\langle g_i\rangle.

Histological and volume-weighted $g$-ratio distributions

Figure 2. Histological and volume-weighted gg-ratio distributions. The top panel shows the histological (unweighted) distribution of the gg-ratio, while the bottom panel shows its volume-weighted MR-relevant analog. Distributions are obtained by pooling gg-ratios from all axons of all histological sections and animals. The peak at gi=1g_i = 1 corresponds to formally including unmyelinated axons into the distributions, relevant for MR measurements (see text). Our main observation is that including the unmyelinated axons shifts the corresponding mean values notably to the right, Eq. (1) to Eq. (5), dashed black vertical lines (red to green), with the shift represented by the black arrows. Furthermore, the nonlinear averaging (2) shifts the resulting effective gg-ratio geffg_{\mathrm{eff}} and gt,effg_{\mathrm{t,eff}} a bit to the left relative to the respective distribution’s mean ⟨gi⟩vol\langle g_i \rangle_{\mathrm{vol}}, represented by the negative variance corrections in Eqs. (4) and (6). This negative correction (blue arrows) is smaller than the major increase due to including the unmyelinated axons (black arrows) in the calculation. The blue arrow correction is larger for the total effective gg-ratio relative to that for myelinated axons alone, as the total distribution is broader.

From the MRI perspective, single-axon gig_i measurements are not accessible. Instead, an effective gg-ratio for myelinated axons geffg_{\mathrm{eff}} is inferred[20] from voxel-averaged area (volume) fractions:

geff=∑i∈mAAin,i∑i∈mAAout,i=ϕmAϕm+ϕmA.(1)g_{\mathrm{eff}} = \sqrt{\frac{\sum_{i\in\mathrm{mA}} A_{\mathrm{in},i}}{\sum_{i\in\mathrm{mA}} A_{\mathrm{out},i}}} = \sqrt{\frac{\phi_{\mathrm{mA}}}{\phi_{\mathrm{m}}+\phi_{\mathrm{mA}}}} . \tag*{(1)}

Eq. (1) nonlinearly pools the underlying intra-voxel distribution of gig_i into an effective

geff=1∑iwigi−2≡⟨gi−2⟩vol−1/2<⟨gi⟩vol,(2)g_{\mathrm{eff}} = \frac{1}{\sqrt{\sum_i w_i g_i^{-2}}} \equiv\langle g_i^{-2}\rangle_{\mathrm{vol}}^{-1/2} < \langle g_i\rangle_{\mathrm{vol}}, \tag*{(2)}

where the weights

wi=Ain,i∑jAin,j(3)w_i = \frac{A_{\mathrm{in},i}}{\sum_j A_{\mathrm{in},j}} \tag*{(3)}

(bottom panel of Figure 2). The largest axons dominate the volume-weighted average ⟨…⟩vol\langle\ldots\rangle_{\mathrm{vol}} due to the weights (3); a similar effect arises for MR-based axon diameters[38, 64]. Furthermore, the nonlinear averaging in Eq. (2) makes the MR-based ratio (1) lower than the volume-weighted mean ⟨gi⟩vol\langle g_i\rangle_{\mathrm{vol}}.

The latter can be seen by expanding in the deviations δgi=gi−⟨gi⟩vol\delta g_i = g_i-\langle g_i\rangle_{\mathrm{vol}} from the mean up to (δgi)2(\delta g_i)^2:

⟨(⟨gi⟩vol+δgi)−2⟩vol−1/2≈⟨gi⟩vol⋅⟨1+3(δgi)2⟨gi⟩vol2⟩vol−1/2\left\langle\left(\langle g_i\rangle_{\mathrm{vol}}+\delta g_i\right)^{-2}\right\rangle_{\mathrm{vol}}^{-1/2} \approx\langle g_i\rangle_{\mathrm{vol}} \cdot\left\langle1+3\frac{(\delta g_i)^2}{\langle g_i\rangle_{\mathrm{vol}}^2}\right\rangle_{\mathrm{vol}}^{-1/2}

since ⟨δgi⟩vol≡0\langle\delta g_i\rangle_{\mathrm{vol}}\equiv0. As a result, the effective gg-ratio

geff≈⟨gi⟩vol⋅(1−32var⁡vol{gi}⟨gi⟩vol2)(4)g_{\mathrm{eff}} \approx\langle g_i\rangle_{\mathrm{vol}}\cdot\left(1-\frac{3}{2}\frac{\operatorname{var}_{\mathrm{vol}}\{g_i\}}{\langle g_i\rangle_{\mathrm{vol}}^2}\right) \tag*{(4)}

is negatively-biased relative to the volume-weighted mean.

In demyelinating conditions, the axonal population is no longer composed solely of myelinated fibers. Unmyelinated or fully demyelinated axons can constitute a substantial fraction of the axonal population. We therefore define the total axonal volume fraction as ϕtA=ϕmA+ϕuA\phi_{\mathrm{tA}}=\phi_{\mathrm{mA}}+\phi_{\mathrm{uA}}, and introduce a total effective gg-ratio

gt,eff=ϕtAϕm+ϕtA,(5)g_{\mathrm{t,eff}} = \sqrt{\frac{\phi_{\mathrm{tA}}}{\phi_{\mathrm{m}}+\phi_{\mathrm{tA}}}}, \tag*{(5)}

which includes unmyelinated axons with gi∈uA=1g_{i\in\mathrm{uA}}=1 by definition. While geffg_{\mathrm{eff}} reflects only the myelinated axon population, gt,eff>geffg_{\mathrm{t,eff}}>g_{\mathrm{eff}} captures changes in the composition of the entire axon population during demyelination and remyelination. Figure 2 shows that the shift of gt,effg_{\mathrm{t,eff}} relative to geffg_{\mathrm{eff}} is very pronounced.

Furthermore, the difference of the corresponding effective gg-ratio from the volume-weighted distribution mean becomes more pronounced for gt,effg_{\mathrm{t,eff}} under demyelinating conditions. Indeed, the presence of demyelinated and unmyelinated axons increases the variance of the intra-voxel gg-ratio distribution through the additional contribution of axons with gi=1g_i=1. Hence, in the analog of Eq. (4),

gt,eff≈⟨gi⟩t,vol⋅(1−32var⁡t,vol{gi}⟨gi⟩t,vol2),(6)g_{\mathrm{t,eff}} \approx\langle g_i\rangle_{\mathrm{t,vol}}\cdot\left(1-\frac{3}{2}\frac{\operatorname{var}_{\mathrm{t,vol}}\{g_i\}}{\langle g_i\rangle_{\mathrm{t,vol}}^2}\right), \tag*{(6)}

the averages are now taken with respect to the broader distribution of gig_i for all axons, which includes a peak at gi=1g_i=1. In other words, the bias in effective gg-ratio due to intra-voxel gg-ratio variability becomes more pronounced in conditions involving myelin damage, where the variability increases due to the gi=1g_i=1 population (compare the two blue arrows in the right panel of Figure 2).

To summarize, including unmyelinated axons into the total gg-ratio distribution makes the MR-relevant gg-ratio (5) notably larger than that for the myelinated axons, Eq. (1). The second-order effect is that these quantities are a bit smaller than the mean values over the corresponding distributions due to the nonlinear character of the averaging (2), with the negative corrections determined by the distribution variance (4) and (6). The third-order effect is that this variance-based correction is more pronounced for the total distribution than for myelinated axons alone, as the total distribution is broader.

Results

Segmentation of white matter microstructure. Figure 3 shows the segmentation of TEM images of the splenium of the corpus callosum, acquired at an ultra-high resolution, 2.36 nm, enabling visualization of individual cellular and axonal membranes. We performed semantic segmentation of the tissue into myelin, myelinated axons, unmyelinated axons, and extra-axonal space, using the trained UNet[58] as shown in the middle row of Figure 3. Subsequently, each myelinated and unmyelinated axon has been assigned a unique label to achieve instance-level segmentation, enabling downstream morphometric and statistical analyses.

White matter microstructure segmentation in TEM images

Figure 3. White matter microstructure segmentation in TEM images. (Top) TEM images are acquired at ultra-high resolution with a voxel size of 2.36×2.36 nm22.36 \times2.36\ \mathrm{nm}^{2}: three representative TEM images at 12 weeks from the Ctl, Cpz6, and Cpz12 groups. (Middle) We applied a deep learning-based pipeline for the semantic segmentation of myelin, myelinated and unmyelinated axons, and the extra-axonal spaces; colors indicate microstructural classes. (Bottom) Instance segmentation and the zoomed-in view of myelinated and unmyelinated axons; colors are randomly assigned to individual axons.

Microstructural analysis of cuprizone exposure. Figure 4 summarizes the quantitative analysis of TEM images and reveals microstructural changes consistent with the general expected trends in cuprizone-induced demyelination and recovery.

Bar plots of myelin- and axonal-related parameters

Figure 4. Morphological analysis of microstructural changes in cuprizone-induced demyelination. Myelin- and axonal-related parameters derived from TEM images are shown for each group and time point. Following cuprizone intoxication, demyelination is expected to reduce the myelin volume fraction ϕm\phi_m and the volume fraction of myelinated axons ϕmA\phi_{mA}, while increasing the volume fractions of unmyelinated axons ϕuA\phi_{uA} and extra-axonal space ϕex\phi_{ex}. The total axonal volume fraction ϕtA\phi_{tA} is expected to remain approximately unchanged. Significant differences are asterisked: * : p<0.05p < 0.05; ** : p<0.01p < 0.01; *** : p<0.001p < 0.001.

At 6 weeks of cuprizone exposure, we observed moderate reductions in ϕm\phi_m and ϕmA\phi_{mA} (deff=−0.71d_{\mathrm{eff}} = -0.71 and −0.82-0.82, respectively), accompanied by increases in ϕuA\phi_{uA} and ϕex\phi_{\mathrm{ex}} (deff=0.39d_{\mathrm{eff}} = 0.39 and 1.241.24), with ϕtA\phi_{tA} remaining largely unchanged (deff=−0.14d_{\mathrm{eff}} = -0.14).

MetricDefinition
ϕm\phi_mMyelin volume fraction
ϕmA\phi_{mA}Volume fraction of myelinated axons
ϕuA\phi_{uA}Volume fraction of unmyelinated axons
ϕex\phi_{ex}Volume fraction of the extra-axonal space
ϕtA\phi_{tA}Total axonal volume fraction: ϕmA+ϕuA\phi_{mA} + \phi_{uA}
dmAd_{mA}Myelinated axon diameter
duAd_{uA}Unmyelinated axon diameter
geffg_{\mathrm{eff}}Eq. (1) Effective gg-ratio of myelinated axons
gt,effg_{t,\mathrm{eff}}Eq. (5) Effective gg-ratio of total axons
gt,eff∗g^{*}_{t,\mathrm{eff}}Eq. (6) Bias-corrected effective gg-ratio
ϕm,MR\phi_{m,\mathrm{MR}}MRI-derived myelin water fraction
ϕA,MR\phi_{A,\mathrm{MR}}MRI-derived axonal water fraction
gMRg_{\mathrm{MR}}MRI-derived gg-ratio
ffStandard Model axonal water fraction

Table 1. EM- and MR-derived microstructural metrics.

By 12 weeks, demyelination was markedly more pronounced. Both Cpz6 and Cpz12 groups exhibited significant decreases in ϕm\phi_m (deff=−8.73d_{\mathrm{eff}} = -8.73 and −14.20-14.20) and ϕmA\phi_{mA} (deff=−2.80d_{\mathrm{eff}} = -2.80 and −7.60-7.60). In contrast, we observed an increase in ϕuA\phi_{uA} (deff=3.50d_{\mathrm{eff}} = 3.50 and 6.096.09) and ϕex\phi_{ex} (deff=3.53d_{\mathrm{eff}} = 3.53 and 4.674.67). ϕtA\phi_{tA} remained unchanged in the Cpz6 group (deff=−0.87d_{\mathrm{eff}} = -0.87) and increased significantly only in the Cpz12 group (deff=1.37d_{\mathrm{eff}} = 1.37).

At 18 weeks, significant differences persisted between Cpz12 and control mice, with continued reductions in ϕm\phi_m and ϕmA\phi_{mA} (deff=−1.70d_{\mathrm{eff}} = -1.70 and −1.59-1.59), concomitant increases in ϕuA\phi_{uA} and ϕex\phi_{ex} (deff=1.15d_{\mathrm{eff}} = 1.15 and 2.802.80), and no change in ϕtA\phi_{tA} (deff=0.42d_{\mathrm{eff}} = 0.42). Notably, the opposite directions of change in ϕmA\phi_{mA} and ϕuA\phi_{uA} yield a total axon volume fraction ϕtA\phi_{tA} approximately constant, suggesting overall axonal preservation despite extensive myelin loss.

To evaluate the recovery, we examined within-group

changes over time. Although Cpz6 mice remained significantly different from controls at 12 weeks and Cpz12 mice at 18 weeks, increases in ϕm\phi_m and ϕmA\phi_{mA} from week 6 to week 12 in Cpz6 mice (deffϕm=0.4d^{\phi_m}_{\mathrm{eff}} = 0.4, deffϕmA=0.64d^{\phi_{mA}}_{\mathrm{eff}} = 0.64), and from week 12 to week 18 in Cpz12 mice (deffϕm=0.93d^{\phi_m}_{\mathrm{eff}} = 0.93, deffϕmA=1.54d^{\phi_{mA}}_{\mathrm{eff}} = 1.54), indicate partial remyelination following cuprizone withdrawal.

The diameter of myelinated and unmyelinated axons. To assess whether the observed changes in axonal volume fractions were accompanied by alterations in axonal diameter, we compared the diameter distributions of myelinated and unmyelinated axons across groups and time points. Across all groups and time points, myelinated axons exhibited a mean diameter of 0.55±0.19 μm0.55 \pm0.19\ \mu\mathrm{m}, compared with 0.23±0.12 μm0.23 \pm0.12\ \mu\mathrm{m} for unmyelinated axons, indicating that myelinated axons are, on average, approximately twice as large. As shown in Figure 5, no significant differences in axonal diameter were observed for either population across treatment groups and time points, except for a transient difference in the diameter of myelinated axons between Ctl and Cpz6 mice at week 6. This effect was not sustained at later time points.

Axon diameter distributions for myelinated and unmyelinated axons

Figure 5. Axon diameter distributions are largely preserved across treatments. Distributions of axonal diameters for myelinated dmAd_{mA} and unmyelinated duAd_{uA} axons show no differences between control and cuprizone-treated groups at most time points. A significant difference between Ctl and Cpz6 is observed at week 6; however, this effect is not sustained at later time points and is not accompanied by consistent shifts in diameter distributions. This relative consistency of axon diameter indicates that the observed changes in axonal volume fractions primarily reflect alterations in axonal population density and composition rather than pervasive diameter remodeling of individual axons. Significant differences are asterisked: * : p<0.05p < 0.05; ** : p<0.01p < 0.01; *** : p<0.001p < 0.001.

Unmyelinated axons contribute to MRI-derived axonal water fraction. Here, we evaluate how intra-axonal water fraction ff correlates with the EM-measures of axonal volume fractions. For comparison with ff, EM measures were expressed relative to the MR-visible compartment by normalizing with (1−ϕm)(1-\phi_m). We ob-

served strong and statistically significant correlations (ρ≥∣0.71∣\rho\geq\lvert0.71\rvert, p<0.001p<0.001) between ff and all normalized EM-derived metrics (Figure 6). The negative correlation between ff and ϕuA/(1−ϕm)\phi_{uA}/(1-\phi_m) reflects their opposing changes during demyelination.

Correlation plots between intra-axonal water fraction and normalized EM-derived volume fractions

Figure 6. Correlation between the MR-derived intra-axonal water fraction ff and EM-derived microstructural volume fractions. To enable direct comparison with MR-derived ff, we normalize EM-derived volume fraction by 1/(1−ϕm)1/(1-\phi_m) as myelin does not contribute to the dMRI signal. Dashed lines indicate linear fits (y=c1+c2xy=c_1+c_2x), where xx and yy are defined by the axis labels. The intra-axonal water fraction ff correlates strongly (Spearman correlation coefficient ρ\rho) with the normalized volume fractions of myelinated axons ϕmA\phi_{mA}, unmyelinated axons ϕuA\phi_{uA}, and the total axonal population ϕtA\phi_{tA}. However, the linear fit between ff and ϕtA\phi_{tA} exhibits a negligible intercept, in contrast to the plots showing ff vs ϕmA\phi_{mA} and ff vs ϕuA\phi_{uA}, indicating that ff reflects contributions from both myelinated and unmyelinated axons rather than selectively capturing myelinated fibers alone.

Interestingly, the linear relationship between ff and ϕtA/(1−ϕm)\phi_{tA}/(1-\phi_m) exhibits a negligible intercept (c1=−0.04c_1=-0.04), in contrast to the substantial non-zero intercept observed when ff is regressed against the ϕmA/(1−ϕm)\phi_{mA}/(1-\phi_m) (c1=0.20c_1=0.20). This difference indicates that ff reflects contributions from both myelinated and unmyelinated axonal populations rather than selectively capturing myelinated fibers alone. The slope deviates from unity, indicating a scaling effect between modalities.

Effect of cuprizone exposure on gg-ratio. Figure 7 compares gg-ratio definitions, derived from histology as described in Methods, highlighting their sensitivity to demyelination and recovery. The mean ⟨gi⟩\langle g_i\rangle and the effective geffg_{\mathrm{eff}} in Eq. (1) reflect myelination within the myelinated axonal population. Despite their different constructions, they fall within a similar numerical range and remain largely insensitive to demyelination and recovery, showing no significant changes across control and cuprizone groups at all time points.

g-ratio analysis in cuprizone-induced demyelination

Figure 7. gg-ratio analysis in cuprizone-induced demyelination. Following cuprizone intoxication, the mean microscopic gg-ratio ⟨gi⟩\langle g_i\rangle and the effective geffg_{\mathrm{eff}} of myelinated axons Eq. (1) showed no significant differences among groups. In contrast, the effective gt,effg_{t,\mathrm{eff}} of the total axonal population Eq. (5) and its bias-corrected gt,eff∗g^{*}_{t,\mathrm{eff}} analog Eq. (4) (corrected for intra-voxel gg-ratio variability) captured demyelination and increased significantly in cuprizone-treated groups at 12 and 18 weeks. Significant differences are asterisked: ∗*: p<0.05p<0.05; ∗∗**: p<0.01p<0.01; ∗∗∗***: p<0.001p<0.001.

In contrast, the total effective gt,effg_{\mathrm{t,eff}} defined in Eq. (5) and its bias-corrected form gt,eff∗g_{\mathrm{t,eff}}^{*} in Eq. (6), which corrects the negative bias from nonlinear averaging over heterogeneous gg-ratios, explicitly incorporate the area-weighted contribution of unmyelinated axons, for which gi∈uA=1g_{i\in uA}=1. As a result, they attain a higher numerical range compared to ⟨gi⟩\langle g_i\rangle and geffg_{\mathrm{eff}}. Now, both gt,effg_{\mathrm{t,eff}} and gt,eff∗g_{\mathrm{t,eff}}^{*} capture demyelination-related changes, showing significant increases in cuprizone-treated groups relative to controls at 12 weeks and remained higher in the Cpz12 group at 18 weeks.

Unmyelinated axons contribute to experimental MR gg-ratio. To estimate an experimental MR-derived gg-ratio, denoted as gMRg_{\mathrm{MR}}, we utilized R2R_2 and MTR that are, by design, most sensitive to myelin content. As shown in Figure 8, both metrics were evaluated against ϕm\phi_m, where R2R_2 exhibited the strongest correlation (ρ=0.81\rho= 0.81, p<0.001p < 0.001), whereas MTR followed the same trend but with lower sensitivity (ρ=0.45\rho= 0.45, p=0.05p = 0.05).

Correlation between MRI measures of myelination and histology

Figure 8. Correlation between MRI measures of myelination and histology. R2R_2 shows a strong and significant correlation with the EM-derived myelin volume fraction ϕm\phi_m, while MTR follows the same trend but with reduced sensitivity, resulting in only marginal statistical significance. Dashed lines indicate linear fits (y=c1+c2xy = c_1 + c_2x), where xx and yy are defined by the axis labels.

Based on the strong linear relationship between R2R_2 and ϕm\phi_m as shown in Figure 8, we estimated the MRI-derived myelin volume fraction as ϕm,MR=(R2−15.27)/25.12\phi_{m,\mathrm{MR}} = (R_2 - 15.27)/25.12, obtained by inverting the fitted linear model. An MRI-derived axonal volume fraction was then estimated as ϕA,MR=f(1−ϕm,MR)\phi_{A,\mathrm{MR}} = f(1 - \phi_{m,\mathrm{MR}}), where the factor (1−ϕm,MR)(1 - \phi_{m,\mathrm{MR}}) accounts for the non–MR-visible myelin compartment. The experimental MRI-derived gMRg_{\mathrm{MR}} was then calculated

gMR=ϕA,MRϕm,MR+ϕA,MR.g_{\mathrm{MR}} = \sqrt{\frac{\phi_{A,\mathrm{MR}}}{\phi_{m,\mathrm{MR}} + \phi_{A,\mathrm{MR}}}.}

This construction yields an effective MRI-derived gg-ratio analogous to Eqs. (1) and (5), whose correspondence to either the myelinated or total axonal population is evaluated below.

Comparison of gMRg_{\mathrm{MR}} with the gg-ratio definitions (see Figure 9) introduced in Methods showed poor correlation with both ⟨gi⟩\langle g_i\rangle and geffg_{\mathrm{eff}}. In contrast, when unmyelinated axons were explicitly included, strong and significant correlations emerged between gMRg_{\mathrm{MR}} and both the total effective gt,effg_{\mathrm{t,eff}} and gt,eff∗g_{\mathrm{t,eff}}^{*} (ρ=0.77\rho= 0.77, p<0.001p < 0.001). This finding indicates that the MRI-derived gg-ratio reflects the combined contribution of myelinated and unmyelinated axons rather than selectively reporting the myelin fraction of myelinated fibers alone.

Correlation plots between histology- and MR-derived g-ratio

Figure 9. Correlation between histology- and MR-derived gg-ratio. The MR estimation of gg-ratio correlates strongly with the histology measure of gg-ratio when the contribution of unmyelinated axons is included, as in gtAg_{\mathrm{tA}} and the dispersion corrected gtA∗g_{\mathrm{tA}}^{*}. Dashed lines indicate linear fits (y=c1+c2xy = c_1 + c_2x), where xx and yy are defined by the axis labels.

Discussion

We studied white-matter microstructure in a cuprizone-induced model of demyelination and recovery, emphasizing the role of unmyelinated axons in shaping MR-derived microstructural metrics. The instance-level segmentation of myelinated and unmyelinated axons in high-resolution 2D TEM images, a task that is intrinsically difficult due to single-layer unmyelinated axons, allowed us to resolve axon-specific geometries and derive the full set of volume fractions, including that of myelin, myelinated and unmyelinated axons, and the extra-axonal space, together with gg-ratio measures accounting for their intra-voxel variability. Integration of these EM-based measurements with quantitative MRI revealed a nuanced picture of demyelination characterized by a pronounced loss of myelin and myelinated axons, accompanied by a relative expansion of unmyelinated axons and the extra-axonal compartment. Consistent with this structural reorganization, the SM-derived axonal water fraction reflected contributions from both myelinated and unmyelinated axons, reinforcing its interpretation as a measure of total intra-axonal volume fraction rather than selectively probing myelinated fibers. This contribution of unmyelinated axons also substantially affected the MR-derived gMRg_{\mathrm{MR}}-ratio.

The cuprizone-induced shift in axonal composition had a systematic impact on both histology- and MRI-derived measures of white matter integrity. In particular, this reorganization explains why the MRI-derived axonal water fraction correlated more strongly with the total axonal volume fraction than with the volume fraction of myelinated axons. dMRI is therefore sensitive to intra-axonal water from both myelinated and unmyelinated axons and should not be interpreted as a measure specific to myelinated axons alone.

Importantly, this interpretation applies to the relatively short diffusion time used in the present study ($\Delta =16= 16 ms). At diffusion times shorter than the exchange time, water associated with unmyelinated axons contributes directly to the MR-measured axonal water fraction. As diffusion time exceeds the exchange time, intercompartmental exchange between unmyelinated axons and extra-axonal space progressively transfers this contribution to the extra-axonal compartment[65, 66]. Therefore, the contribution of unmyelinated axons to the measured axonal water fraction decreases with increasing diffusion time. This suggests that clinical measurements, which typically operate at longer diffusion times, may be less sensitive to the distinct contribution of unmyelinated axons than shorter-time preclinical experiments. Distinguishing between myelinated and unmyelinated axons is critical in pathological contexts in which myelin loss occurs without overt axonal degeneration, such as the early demyelinating or remyelinating stages of multiple sclerosis[17, 67], where apparent reductions in myelin content may coexist with preserved total axonal density.

The EM-derived metrics exhibited the expected trends of cuprizone intoxication and recovery. Following the chronic cuprizone exposure, the volume fraction of myelin and myelinated axons decreased, while the volume fraction of unmyelinated axons increased. During the recovery phase, these effects were partially reversed, consistent with prior reports of demyelination and remyelination dynamics in the cuprizone model[15, 18, 55, 68, 69] and the increase of the population of intrinsically unmyelinated axons as an adaption mechanism[19] or, at minimum, a selective sparing of unmyelinated axons from cuprizone toxicity[15]. The near-constancy of the total axonal volume fraction across treatment conditions suggests that cuprizone primarily affects the myelin sheath rather than causing gross axonal loss.

The analysis of axonal diameters further supports this interpretation (for a 3D EM study that quantifies myelinated and unmyelinated axons along axonal trajectories, see Ref.[32]). The absence of systematic changes in axonal diameters indicates that the observed alterations in axonal volume fractions are not driven by swelling, shrinkage, or caliber remodeling of individual axons. Instead, they primarily reflect changes in axonal population density and composition within the tissue. Note that we do not distinguish between axons that were intrinsically unmyelinated and those that became unmyelinated following demyelination.

The myelin-sensitive MR metrics R2R_2 and MTR were sensitive to cuprizone-induced demyelination and recovery at all time points, in line with previous studies[4, 70, 71, 72, 73], with R2R_2 showing the strongest correlation. However, MTR showed only a modest correlation with ϕm\phi_m in our data. Although previous studies reported significant correlations between MTR and either myelin basic protein content[71] or myelin volume fraction[4], MTR is not specific to myelin. In particular, significant correlations between MTR and the volume fraction of non-myelinated cells have also been reported[4], yet it can be sensitive to inflammation and edema[74, 75]. These limitations in myelin specificity should therefore also be considered when interpreting gg-ratio, which combines myelin- and axonal water fraction.

Despite pronounced reductions in myelin volume fraction, the mean gg-ratio ⟨gi⟩\langle g_i \rangle and the effective geffg_{\mathrm{eff}} of myelinated axons in Eq. (1) remained largely unchanged across treatment groups, underscoring the maintained ratio between the inner and outer diameters in surviving myelinated axons. In contrast, the total effective gt,effg_{\mathrm{t,eff}} in Eq. (5), which explicitly incorporates the area-weighted contribution of unmyelinated axons (gi∈uA=1g_{i\in\mathrm{uA}} = 1), increased significantly during demyelination and corresponds best to the experimental MR-derived gMRg_{\mathrm{MR}} as measured from quantitative dMRI and T2T_2-MRI. This increase reflects the growing fraction of unmyelinated axons and links the total effective gg-ratio to the effective redistribution of myelination within an MRI voxel, rather than to changes in per-axon myelin thickness.

The effective geffg_{\mathrm{eff}} combines estimates of myelin and axonal volume fractions to yield a voxel-wise index of myelination[20, 25]. Yet, this model implicitly assumes a uniform gg-ratio within each voxel, an approximation that neglects intra-voxel heterogeneity in myelin thickness and the presence of unmyelinated axons. Such heterogeneity biases effective gg-ratio estimates toward lower values[25]. As shown in Figure 10, this bias becomes substantial in tissue with a high fraction of unmyelinated or demyelinated axons, reaching up to 8% and potentially introducing apparent group differences or obscuring true biological effects.

Bar plots showing within-voxel variability in the effective g-ratio estimates for myelinated axons and the total axonal population

Figure 10. Within-voxel variability in the gg-ratio results in a systematic negative bias in the MR-derived gg-ratio. Relative error in effective gg-ratio estimates arising from discarding within-voxel gg-ratio variability, shown separately for myelinated axons gMRg_{\mathrm{MR}} and the total axonal population gt,MRg_{\mathrm{t,MR}}. The bias is larger for gt,MRg_{\mathrm{t,MR}}, reflecting the increased variance introduced by including unmyelinated axons with gi=1g_i = 1. In some comparisons, this bias reaches statistical significance. Significant differences are asterisked: *: p<0.05p < 0.05; **: p<0.01p < 0.01; ***: p<0.001p < 0.001.

We derived an analytical correction that links the variance of the gg-ratio distribution to the macroscopic bias observed in both geffg_{\mathrm{eff}} and gt,effg_{\mathrm{t,eff}}. This framework establishes a quantitative connection between microscopic gg-ratio variability and its measurable effective counterpart. In practice, however, applying this correction led to overcompensation: the bias-corrected measure gt,eff∗g_{\mathrm{t,eff}}^{*} increased the mismatch with gMRg_{\mathrm{MR}} (bias 0.02 →\rightarrow 0.08). This overshoot suggests that the magnitude of var⁡vol{gi}\operatorname{var}_{\mathrm{vol}}\{g_i\} required to reconcile MRI- and EM-derived gg-ratio estimates is larger than what is observed in the data, implying that gg-ratio heterogeneity alone cannot fully explain the observed discrepancy between modalities.

In conclusion, the diffusion-derived axonal water fraction reflects the total axonal volume fraction at short diffusion times, encompassing both myelinated and unmyelinated axons, as confirmed by its strong correspondence with EM-derived microstructural measures. The same principle applies to the gg-ratio: accurate interpretation requires inclusion of unmyelinated axons, whose contribution substantially influences voxel-averaged estimates. Moreover, intra-voxel heterogeneity in gg-ratio distributions systematically underestimates the effective gg-ratio, with the magnitude of this reduction increasing alongside the fraction of unmyelinated axons and the gg-ratio variability. Together, these findings demonstrate that unmyelinated axons are integral to the microstructural contrast detected by MRI and must be considered explicitly when quantifying axonal and myelin integrity in both healthy and demyelinated white matter.

Data availability

The high-resolution EM data are publicly available through the Open Science Framework at https://osf.io/adrcs/overview. DeepACSON is publicly available at https://github.com/ComputationalNeuroanatomy, and the DESIGNER pipeline is available at https://nyu-diffusionmri.github.io/DESIGNER-v2.

Acknowledgments

We thank Prof. James L. Salzer and Prof. Jiangyang Zhang for their discussions and feedback on this work. We thank Hong-Hsi Lee for his feedback. We also thank Ileana O. Jelescu and Magdalena Zurek for the data used in this study.

Funding

The research was supported by the NIH under awards R01 NS088040 and R21 NS081230, and by the Irma T. Hirschl fund. This research was supported by the Research Council of Finland (grant #360360) awarded to A.A. The research was performed at the Center of Advanced Imaging Innovation and Research (CAI2R, www.cai2r.net), a Biomedical Technology Resource Center supported by NIBIB with award P41 EB017183.

Author contributions

A.A., D.S.N., and E.F. conceived the project and designed the study. A.A. conducted EM image analysis and analyzed the data. R.C.-L. conducted MRI processing and analysis. E.C. contributed to EM analysis. S.G.K. contributed to data acquisition. A.A. and D.S.N. derived the analytical expressions. A.A. and E.F. wrote the manuscript. D.S.N. and E.F. supervised the project. All authors commented on and approved the final manuscript.

Competing interests

The authors declare no competing financial interests.

References

References

  1. [1]A. Mackay, K. Whittall, J. Adler, D. Li, D. Paty, and D. Graeb, In vivo visualization of myelin water in brain by magnetic resonance, Magnetic Resonance in Medicine 31, 673 (1994).
  2. [2]E. Fieremans, D. S. Novikov, J. H. Jensen, and J. A. Helpern, Monte Carlo study of a two-compartment exchange model of diffusion, NMR in Biomedicine 23, 711 (2010).DOI
  3. [3]E. Fieremans, J. H. Jensen, and J. A. Helpern, White matter characterization with diffusional kurtosis imaging, NeuroImage 58, 177 (2011).DOI
  4. [4]J. D. Thiessen, Y. Zhang, H. Zhang, L. Wang, R. Buist, M. R. Del Bigio, J. Kong, X.-M. Li, and M. Martin, Quantitative MRI and ultrastructural examination of the cuprizone mouse model of demyelination, NMR in biomedicine 26, 1562 (2013).DOI
  5. [5]M. Filippi, P. Preziosa, and M. A. Rocca, Magnetic resonance outcome measures in multiple sclerosis trials: Time to rethink?, Current Opinion in Neurology 27, 290 (2014).
  6. [6]N. Weiskopf, S. Mohammadi, A. Lutti, and M. F. Callaghan, Advances in MRI-based computational neuroanatomy: from morphometry to in-vivo histology, Current Opinion in Neurology 28, 313 (2015).DOI
  7. [7]A. L. MacKay and C. Laule, Magnetic Resonance of Myelin Water: An in vivo Marker for Myelin, Brain Plasticity 2, 71 (2016).
  8. [8]I. O. Jelescu, M. Zurek, K. V. Winters, J. Veraart, A. Rajaratnam, N. S. Kim, J. S. Babb, T. M. Shepherd, D. S. Novikov, S. G. Kim, and E. Fieremans, In vivo quantification of demyelination and recovery using compartment-specific diffusion MRI metrics validated by electron microscopy, NeuroImage 132, 104 (2016).DOI
  9. [9]D. S. Novikov, J. Veraart, I. O. Jelescu, and E. Fieremans, Rotationally-invariant mapping of scalar and orientational metrics of neuronal microstructure with diffusion MRI, NeuroImage 174, 518 (2018).DOI
  10. [10]D. S. Novikov, E. Fieremans, S. N. Jespersen, and V. G. Kiselev, Quantifying brain microstructure with diffusion MRI: Theory and parameter estimation, NMR in Biomedicine 32, 10.1002/nbm.3998 (2019).
  11. [11]S. Papazoglou, M. Ashtaraye, J. M. Oeschger, M. F. Callaghan, M. D. Does, and S. Mohammadi, Insights and improvements in correspondence between axonal volume fraction measured with diffusion-weighted MRI and electron microscopy, NMR in Biomedicine 37, e5070 (2024).DOI
  12. [12]Y. V. Sui, H. Bertisch, D. C. Goff, A. Samsonov, and M. Lazar, Quantitative magnetization transfer and g-ratio imaging of white matter myelin in early psychotic spectrum disorders, Molecular Psychiatry 30, 2739 (2025).DOI
  13. [13]K. Schmierer, F. Scaravilli, D. R. Altmann, G. J. Barker, and D. H. Miller, Magnetization transfer ratio and myelin in postmortem multiple sclerosis brain, Annals of Neurology 56, 407 (2004).
  14. [14]C. Laule, I. M. Vavasour, S. H. Kolind, D. K. B. Li, T. L. Traboulsee, G. R. W. Moore, and A. L. MacKay, Magnetic resonance imaging of myelin, Neurotherapeutics: The Journal of the American Society for Experimental NeuroTherapeutics 4, 460 (2007).
  15. [15]G. K. Matsushima and P. Morell, The neurotoxicant, cuprizone, as a model to study demyelination and remyelination in the central nervous system, Brain Pathology (Zurich, Switzerland) 11, 107 (2001).
  16. [16]J. Praet, C. Guglielmetti, Z. Berneman, A. Van der Linden, and P. Ponsaerts, Cellular and molecular neuropathology of the cuprizone mouse model: Clinical relevance for multiple sclerosis, Neuroscience & Biobehavioral Reviews 47, 485 (2014).
  17. [17]E. Schäffner, M. Bosch-Queralt, J. M. Edgar, M. Lehning, J. Strauß, N. Fleischer, T. Kungl, P. Wieghofer, S. A. Berghoff, T. Reinert, M. Krueger, M. Morawski, W. Möbius, A. Barrantes-Freer, J. Stieler, T. Sun, G. Saher, M. H. Schwab, C. Wrede, M. Frosch, M. Prinz, D. S. Reich, A. Flügel, C. Stadelmann, R. Fledrich, K.-A. Nave, and R. M. Stassart, Myelin insulation as a risk factor for axonal degeneration in autoimmune demyelinating disease, Nature Neuroscience 26, 1218 (2023).DOI
  18. [18]M. F. Stidworthy, S. Genoud, U. Suter, N. Mantei, and R. J. M. Franklin, Quantifying the early stages of remyelination following cuprizone-induced demyelination, Brain Pathology (Zurich, Switzerland) 13, 329 (2003).
  19. [19]F. Pfeiffer, G. Frommer-Kaestle, and P. Fallier-Becker, Structural adaption of axons during de- and remyelination in the Cuprizone mouse model, Brain Pathology 29, 675 (2019).DOI
  20. [20]N. Stikov, J. S. Campbell, T. Stroh, M. Lavallée, S. Frey, J. Novek, S. Nuara, M. K. Ho, B. J. Bedell, R. F. Dougherty, I. R. Leppert, M. Boudreau, S. Narayanan, T. Duval, J. Cohen-Adad, P. A. Picard, A. Gasecka, D. Côté, and G. B. Pike, In vivo histology of the myelin g-ratio with magnetic resonance imaging, NeuroImage 118, 397 (2015).DOI
  21. [21]T. Chomiak and B. Hu, What is the optimal value of the g-ratio for myelinated fibers in the rat CNS? A theoretical approach, PLoS ONE 4, 10.1371/journal.pone.0007754 (2009).DOI
  22. [22]A. Behanova, A. Abdollahzadeh, I. Belevich, E. Jokitalo, A. Sierra, and J. Tohka, gACSON software for automated segmentation and morphology analysis of myelinated axons in 3D electron microscopy, Computer Methods and Programs in Biomedicine 220, 106802 (2022).DOI
  23. [23]D. S. Novikov and E. Fieremans, Relating Extracellular Diffusivity to Cell Size Distribution and Packing Density as Applied to White Matter, Proceedings of the International Society of Magnetic Resonance in Medicine 20, p.1829 (2012).
  24. [24]K. L. West, N. D. Kelm, R. P. Carson, and M. D. Does, A Revised Model for Estimating g-ratio from MRI, NeuroImage 125, 1155 (2016).DOI
  25. [25]J. S. W. Campbell, I. R. Leppert, S. Narayanan, M. Boudreau, T. Duval, J. Cohen-Adad, G. B. Pike, and N. Stikov, Promise and pitfalls of g-ratio estimation with MRI, NeuroImage 182, 80 (2018).DOI
  26. [26]T. Skripuletz, V. Gudi, D. Hackstette, and M. Stangel, De- and remyelination in the CNS white and grey matter induced by cuprizone: The old, the new, and the unexpected, Histology and Histopathology 26, 1585 (2011).DOI
  27. [27]M. Zinggbl, P. Assinck, A. Sizov, A. V. Caprariello, and J. R. Plemel, Oligodendrocyte death and myelin loss in the cuprizone model: an updated overview of the intrinsic and extrinsic causes of cuprizone demyelination, Molecular Neurodegeneration 17, 34 (2022).
  28. [28]R. A. Salo, I. Belevich, E. Manninen, E. Jokitalo, O. Grö̈hn, and A. Sierra, Quantification of anisotropy and orientation in 3D electron microscopy and diffusion tensor imaging in injured rat brain, NeuroImage 172, 404 (2018).DOI
  29. [29]I. S. M. Molina, R. A. Salo, A. Abdollahzadeh, J. Tohka, O. Grö̈hn, and A. Sierra, In vivo diffusion tensor imaging in acute and subacute phases of mild traumatic brain injury in rats, eNeuro 7, 1 (2020).
  30. [30]M. Fratini, A. Abdollahzadeh, M. DiNuzzo, R. A. Salo, L. Maugeri, A. Cedola, F. Giove, O. Gröhn, J. Tohka, and A. Sierra, Multiscale Imaging Approach for Studying the Central Nervous System: Methodology and Perspective, Frontiers in Neuroscience 14, 1 (2020).DOI
  31. [31]A. Abdollahzadeh, R. Coronado-Leija, H.-H. Lee, A. Sierra, E. Fieremans, and D. S. Novikov, Scattering approach to diffusion quantifies axonal damage in brain injury, Nature Communications 16, 9808 (2025).
  32. [32]A. Abdollahzadeh, R. Coronado-Leija, S. Mehrin, H.-H. Lee, A. Sierra, E. Fieremans, and D. S. Novikov, Characterization of myelinated and unmyelinated axons in white matter from diffusion MRI perspective, in Proceedings of the International Society for Magnetic Resonance in Medicine (2023).DOI
  33. [33]A. Abdollahzadeh, R. Coronado-Leija, E. Chasen, D. S. Novikov, and E. Fieremans, Evaluating sensitivity of quantitative mri to myelin and axonal integrity using histological validation, in ISMRM Workshop on 40 Years of Diffusion: Past, Present and Future Perspectives (2025).
  34. [34]A. Abdollahzadeh, I. Belevich, E. Jokitalo, J. Tohka, and A. Sierra, Automated 3D Axonal Morphometry of White Matter, Scientific Reports 9, 6084 (2019).
  35. [35]H.-H. Lee, K. Yaros, J. Veraart, J. L. Pathan, F.-X. Liang, S. G. Kim, D. S. Novikov, and E. Fieremans, Along-axon diameter variation and axonal orientation dispersion revealed with 3D electron microscopy: implications for quantifying brain white matter microstructure with histology and diffusion MRI, Brain Structure and Function 224, 1469 (2019).DOI
  36. [36]R. Coronado-Leija, A. Abdollahzadeh, H.-H. Lee, S. Coelho, B. Ades-Aron, Y. Liao, R. A. Salo, J. Tohka, A. Sierra, D. S. Novikov, and E. Fieremans, Volume electron microscopy in injured rat brain validates white matter microstructure metrics from diffusion MRI, Imaging Neuroscience 2, 1 (2024).
  37. [37]E. Kaden, N. D. Kelm, R. P. Carson, M. D. Does, and D. C. Alexander, Multi-compartment microscopic diffusion imaging, NeuroImage 139, 346 (2016).
  38. [38]J. Veraart, D. Nunes, U. Rudrapatna, E. Fieremans, D. K. Jones, D. S. Novikov, and N. Shemesh, Noninvasive quantification of axon radii using diffusion MRI, eLife 9, 10.7554/eLife.49855 (2020).DOI
  39. [39]S. Coelho, S. H. Baete, G. Lemberskiy, B. Ades-Aron, G. Barba, J. Veraart, D. S. Novikov, and E. Fieremans, Reproducibility of the Standard Model of diffusion in white matter on clinical MRI systems, NeuroImage 257, 119290 (2022).
  40. [40]M. Veldmann, L. J. Edwards, K. J. Pine, P. Ehses, M. Ferreira, N. Weiskopf, and T. Stoecker, Improving MR axon radius estimation in human white matter using spiral acquisition and field monitoring, Magnetic Resonance in Medicine 92, 1898 (2024).DOI
  41. [41]M. Tariq, T. Schneider, D. C. Alexander, C. A. Gandini Wheeler-Kingshott, and H. Zhang, Bingham–NODDI: Mapping anisotropic orientation dispersion of neurites using diffusion MRI, NeuroImage 133, 207 (2016).DOI
  42. [42]M. Reisert, E. Kellner, B. Dhital, J. Hennig, and V. G. Kiselev, Disentangling micro from mesostructure by diffusion MRI: A Bayesian approach, NeuroImage 147, 964 (2017).DOI
  43. [43]P. Schlegel, Y. Yin, A. S. Bates, S. Dorkenwald, K. Eichler, P. Brooks, D. S. Han, M. Gkantia, M. dos Santos, E. J. Munnelly, G. Badalamente, L. Serratosa Capdevila, V. A. Sane, A. M. C. Frangier, L. Kiassat, M. W. Pleijzier, T. Stürner, I. F. M. Tamimi, C. R. Dunne, I. Salgarella, A. Javier, S. Fang, E. Perlman, T. Kazimierz, S. R. Jagannathan, A. Matsliah, A. R. Sterling, S.-c. Yu, C. E. McKellar, M. Costa, H. S. Seung, M. Murthy, V. Hartenstein, D. D. Bock, and G. S. X. E. Jefferis, Whole-brain annotation and multi-connectome cell typing of Drosophila, Nature 634, 139 (2024).
  44. [44]A. Sheridan, T. M. Nguyen, D. Deb, W.-C. A. Lee, S. Saalfeld, S. C. Turaga, U. Manor, and J. Funke, Local shape descriptors for neuron segmentation, Nature Methods 20, 295 (2023).
  45. [45]A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, P. Dollár, and R. Girshick, Segment anything, arXiv preprint arXiv:2304.02643 (2023).arxiv.org/abs/2304.02643
  46. [46]A. Archit, L. Freckmann, S. Nair, N. Khalid, P. Hilt, V. Rajasekhar, M. Freitag, C. Teuber, G. Buckley, S. von Haaren, S. Gupta, A. Dengel, S. Ahmed, and C. Pape, Segment Anything for Microscopy, Nature Methods 1 (2025).
  47. [47]K. Oshio and D. A. Feinberg, GRASE (Gradient-and Spin-Echo) imaging: A novel fast MRI technique, Magnetic Resonance in Medicine 20, 344 (1991).
  48. [48]M. Aggarwal, S. Mori, T. Shimogori, S. Blackshaw, and J. Zhang, Three-dimensional diffusion tensor microimaging for anatomical characterization of the mouse brain, Magnetic Resonance in Medicine 64, 249 (2010).
  49. [49]J. Chen, B. Ades-Aron, H.-H. Lee, S. Mehrin, M. Pang, D. S. Novikov, J. Veraart, and E. Fieremans, Optimization and validation of the DESIGNER preprocessing pipeline for clinical diffusion MRI in white matter aging, Imaging Neuroscience 2, 1 (2024).DOI
  50. [50]J. Veraart, D. S. Novikov, D. Christiaens, B. Ades-Aron, J. Sijbers, and E. Fieremans, Denoising of diffusion MRI using random matrix theory, NeuroImage 142, 394 (2016).
  51. [51]B. Ades-Aron, S. Coelho, G. Lemberskiy, J. Veraart, S. H. Baete, T. M. Shepherd, D. S. Novikov, and E. Fieremans, Denoising improves cross-scanner and cross-protocol test–retest reproducibility of diffusion tensor and kurtosis imaging, Human Brain Mapping 46, e70142 (2025).DOI
  52. [52]E. Kellner, B. Dhital, V. G. Kiselev, and M. Reisert, Gibbs-ringing artifact removal based on local subvoxel-shifts, Magnetic Resonance in Medicine 76, 1574 (2015).
  53. [53]B. Avants, C. Epstein, M. Grossman, and J. Gee, Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain, Medical Image Analysis 12, 26 (2008).
  54. [54]Q. Collier, J. Veraart, B. Jeurissen, A. J. den Dekker, and J. Sijbers, Iterative reweighted linear least squares for accurate, fast, and robust estimation of diffusion magnetic resonance parameters, Magnetic Resonance in Medicine 73, 2174 (2015).DOI
  55. [55]A. J. Steelman, J. P. Thompson, and J. Li, Demyelination and remyelination in anatomically distinct regions of the corpus callosum following cuprizone intoxication, Neuroscience Research 72, 32 (2012).DOI
  56. [56]A. Abdollahzadeh, I. Belevich, E. Jokitalo, A. Sierra, and J. Tohka, DeepACSON automated segmentation of white matter in 3D electron microscopy, Communications Bi- ology 4, 179 (2021).
  57. [57]A. Abdollahzadeh, A. Sierra, and J. Tohka, Cylindrical Shape Decomposition for 3D Segmentation of Tubular Objects, IEEE Access 9, 23979 (2021).DOI
  58. [58]O. Ronneberger, P. Fischer, and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, in MICCAI 2015, edited by N. Navab, J. Hornegger, W. Wells, and A. Frangi (Springer International Publishing, Cham, 2015) pp. 234–241.
  59. [59]C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, Understanding deep learning (still) requires rethinking generalization, Communications of the ACM 64, 107 (2021).
  60. [60]D. P. Kingma and J. L. Prodanova, ADAM: A Method for Stochastic Optimization, AIP Conference Proceedings 1631, 58 (2014).arxiv.org/abs/1412.6980
  61. [61]J. Vervaart, D. S. Novikov, and E. Fieremans, TE dependent Diffusion Imaging (TEdDI) distinguishes between compartmental T2 relaxation times, NeuroImage 182, 360 (2018).DOI
  62. [62]C. M. W. Tax, E. Kleban, M. Chamberland, M. Baraković, U. Rudrapatna, and D. K. Jones, Measuring compartmental T2-orientational dependence in human brain white matter using a titable fr coil and diffusion-T2 correlation mri, NeuroImage 236, 117967 (2021).DOI
  63. [63]W. A. H. Rushton, A theory of the effects of fiber size in medullated nerve, The Journal of Physiology , 101 (1951).
  64. [64]L. M. Burcaw, E. Fieremans, and D. S. Novikov, Mesoscopic structure of neuronal tracts from time-dependent diffusion, NeuroImage 114, 18 (2015).DOI
  65. [65]R. Coronado-Leija, D. S. Novikov, and E. Fieremans, Quantifying unmyelinated axons from time-dependent radial kurtosis in brain white matter, in Proceedings of the International Society for Magnetic Resonance in Medicine (2025).DOI
  66. [66]R. Coronado-Leija, Y. Bekku, J. Zhang, J. Salzer, D. S. Novikov, and E. Fieremans, In vivo quantification of unmyelinated axons in brain white matter using diffusion mri, in Proceedings of the International Society for Magnetic Resonance in Medicine (2026).
  67. [67]H. Lassmann, Multiple sclerosis pathology, Cold Spring Harbor Perspectives in Medicine 8, a028936 (2018).
  68. [68]M. Lindner, S. Heine, K. Haastert, N. Garde, J. Fokuhl, F. Linsmeier, C. Grothe, W. Baumgärtner, and M. Stangel, Sequential myelin protein expression during remyelination reveals fast and efficient repair after central nervous system demyelination, Neuropathology and Applied Neurobiology 34, 105 (2008).
  69. [69]M. Xie, J. E. Tobin, M. D. Budde, C.-I. Chen, K. Trinkaus, A. H. Cross, D. P. McDaniel, S.-K. Song, and R. C. Armstrong, Rostrocaudal analysis of corpus callosum demyelination and axon damage across disease stages refines diffusion tensor imaging correlations with pathological features, Journal of Neuropathology and Experimental Neurology 69, 704 (2010).
  70. [70]S.-K. Song, J. Yoshino, T. Q. Le, S.-J. Lin, S.-W. Sun, A. H. Cross, and R. C. Armstrong, Demyelination increases radial diffusivity in corpus callosum of mouse brain, NeuroImage 26, 132 (2005).
  71. [71]W. Zaaraoui, M. Deloire, M. Merle, C. Girard, G. Raffard, M. Biran, M. Inglese, K. G. Petry, O. Gonen, B. Brochet, J.-M. Franconi, and V. Dousset, Monitoring demyelination and remyelination by magnetization transfer imaging in the mouse brain at 9.4 T, Magnetic Resonance Materials in Physics, Biology and Medicine 21, 357 (2008).DOI
  72. [72]Q.-Z. Wu, Q. Yang, H. S. Cate, D. Kemper, M. Binder, H.-X. Wang, K. Fang, M. J. Quick, M. Marriott, T. J. Kilpatrick, and G. F. Egan, MRI identification of the rostral-caudal pattern of pathology within the corpus callosum in the cuprizone mouse model, Journal of Magnetic Resonance Imaging 27, 446 (2008).
  73. [73]M. F. Falangola, D. N. Guilfoyle, A. Tabesh, E. S. Hui, X. Nie, J. H. Jensen, S. V. Gerum, C. Hu, J. LaFrancois, H. R. Collins, and J. A. Halpern, Histological correlation of diffusional kurtosis and white matter modeling metrics in cuprizone-induced corpus callosum demyelination, NMR in Biomedicine 27, 948 (2014).DOI
  74. [74]P. J. Gareau, B. K. Rutt, S. J. Karlik, and J. R. Mitchell, Magnetization transfer and multicomponent T2 relaxation measurements with histopathologic correlation in an experimental model of MS, Journal of Magnetic Resonance Imaging 11, 586 (2000).
  75. [75]I. M. Vavasour, C. Laule, D. K. Li, A. L. Traboulsee, and A. L. MacKay, Is the magnetization transfer ratio a marker for myelin in multiple sclerosis?, Journal of Magnetic Resonance Imaging 33, 710 (2011).

Paper details

vlnc:W_hjzhfhxh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents